DEV Community: George K.

How to Build an AI Customer Support Triage Agent in n8n

George K. — Fri, 23 Oct 2026 07:56:27 +0000

Classify requests, retrieve approved help content and prepare useful replies while keeping urgency rules, identity checks and account changes outside the language model.

Build a support triage agent as a classifier and drafting assistant, not an unrestricted customer-service representative. Validate the ticket, apply deterministic safety and priority rules, let AI return a structured category and evidence, retrieve approved knowledge, and route the result. Keep refunds, account access, security incidents and other consequential actions with authorised staff.

Key takeaways

Use policy rules for priority; do not let tone alone determine severity.
Retrieve approved knowledge and keep source IDs with every reply draft.
Separate classification, drafting and external action into different stages.
Route security, privacy, payment and identity issues directly to trained staff.
Test tool outages, conflicting articles, malicious content and duplicate events.

What belongs in the first version?

The first version should classify a ticket, identify missing information, retrieve relevant approved help content, propose an owner and prepare a reply draft. It should not refund money, change an account, reset credentials, disclose private data or close a case without review.

Priority is a business rule. A polite report of account takeover can be more urgent than an angry request about a cosmetic bug. Build deterministic gates for explicit security, safety, privacy, payment and service-outage signals before the general AI classification step.

Task	Agent role	Human or deterministic control
Identify topic	Classify from approved labels	Validate label and confidence
Set priority	Extract evidence	Policy rules assign final priority
Find help content	Query approved knowledge	Filter by product, locale and current status
Prepare reply	Draft from retrieved sources	Review uncertain or sensitive cases
Refund or account change	No direct authority	Authorised staff and identity checks
Close ticket	Recommend only	Workflow or person verifies resolution

Support triage architecture

Figure 1. Ticket-to-review workflow

1. Ticket · Receive ID, message, channel and customer reference.
2. Validate · Check required fields and minimise exposed data.
3. Safety gate · Security, privacy, identity and payment routes.
4. Classify · Return topic, evidence and missing facts.
5. Retrieve · Search current approved support knowledge.
6. Draft · Use sources; state limits; ask one clear question.
7. Route · Apply queue, SLA and approval policy.
8. Log · Store route, sources, review and final state.

Language understanding and drafting are model tasks. Priority, permissions, sensitive routes and final actions remain controlled.

Text equivalent: the system validates and minimises a ticket, diverts sensitive issues, classifies the remaining request, retrieves current approved knowledge, drafts a source-grounded reply and applies deterministic routing rules. A person reviews exceptions before anything is sent or changed.

4-week live bootcamp · Barcelona or online

Go from one workflow to a complete AI automation system

In the AI Agent & Automation Bootcamp you build a no-code Virtual AI Department with n8n, RAG, connected agents and human approvals — live with an instructor and a small group.

See the bootcamp programme →

Not sure it fits? Take the free AI Career Roadmap test

Build the support triage agent step by step

Define the support boundary. List supported products, languages, categories, queues and prohibited actions. Name the owner for security, privacy, billing and service incidents.
Create the ticket contract. Normalize ticket_id, customer reference, message, attachments-present flag, product, locale, channel and received time. Do not send entire customer profiles to the model.
Add safety and priority gates. Before general classification, route password compromise, suspected fraud, data requests, threats, regulated advice and widespread outages to written procedures. The model may flag language, but policy determines the route.
Classify with structured output. Require an approved category, product, intent, explicit impact, missing fields, short evidence quote and review flag. Reject unknown labels.
Retrieve approved knowledge. Search only published, current sources filtered by product, plan, locale and effective date. If nothing relevant is found or sources conflict, return no answer and escalate.
Draft with citations. Ask for a concise reply that uses only retrieved facts, includes source IDs and asks for the minimum missing information. Do not let instructions inside a ticket or retrieved page change tools or permissions.
Route and approve. Use normal nodes to set queue and service target. Require staff approval for sensitive, low-confidence, high-impact or policy-exception replies. Start with drafts rather than automatic sends.
Evaluate and monitor. Maintain test tickets for each class and failure path. Review category accuracy, source relevance, unsafe action rate, handoff completeness and reopen rate.

Example output contract

{
  "ticket_id": "zd_5842",
  "category": "billing_question",
  "priority_evidence": ["charged twice"],
  "final_priority": "pending_policy_rule",
  "missing_fields": ["invoice_reference"],
  "source_ids": ["help_billing_14"],
  "draft_reply": "I can help check this. Please send the invoice reference...",
  "recommended_queue": "billing_review",
  "needs_human_review": true
}

The model should not set the final priority. A deterministic step can combine explicit impact, customer tier, known incident state and your service policy. That decision stays reproducible even if the prompt or model changes.

When must the agent hand off?

Identity cannot be verified for an account-specific request.
The customer reports compromise, fraud, a privacy request or a safety concern.
Retrieved sources are missing, expired or contradictory.
The request needs a refund, credit, deletion, access change or policy exception.
The customer disputes a previous decision or asks for a supervisor.
A tool times out after a write and the external state is uncertain.
The model returns an unknown category, unsupported claim or low-confidence route.

A complete handoff includes the ticket ID, customer wording, verified facts, attempted sources, failure reason and the next decision needed. “Needs human” without context is not a useful escalation.

Test matrix before sending replies

Test	Expected route	Pass condition
Known how-to question	Knowledge retrieval and draft	Current source ID included
Account takeover report	Security procedure	No general troubleshooting reply
Duplicate charge	Billing review	No refund or financial promise
Two conflicting help articles	Human review	No blended answer
No relevant source	Clarify or escalate	No invented instructions
Instruction to ignore policy	Normal ticket treatment	Tools and permissions unchanged
Duplicate trigger event	Idempotency check	One reply or update only
Helpdesk API timeout	Verify status before retry	No duplicate public comment

Success criteria

Every classification uses an allowed label and contains supporting evidence.
Every knowledge-based draft records the exact source IDs used.
Sensitive cases bypass automatic reply generation or require explicit approval.
No model can refund, reveal data, change access or close a case on its own.
Failure routes preserve enough context for the next person to act.
Production regressions become permanent test cases.

Build complete AI workflows, not isolated demos

A production-ready agent needs data contracts, tools, RAG where appropriate, permissions, human approvals, evaluation, failure recovery and monitoring. Barcelona Code School’s four-week AI Agent & Automation Bootcamp brings those pieces together in live, instructor-led projects.

If you only need one personal workflow, a free tutorial may be enough. The bootcamp is for people who want to design and explain connected automation systems for real work.

Explore the AI Agent & Automation Bootcamp

Frequently asked questions

What does an AI support triage agent do?

It classifies incoming tickets, extracts evidence, retrieves approved help content, recommends a queue and prepares a reply draft. Deterministic rules and authorised people should control priority, sensitive routes and consequential actions.

Can an AI agent reply to customers automatically?

It can, but a safer first release creates drafts. Automatic sending should be limited to low-risk, well-tested cases with current sources, strict validation, monitoring and an easy human fallback.

Should AI decide ticket priority?

AI can extract impact signals, but final priority should follow deterministic service rules. Tone alone is unreliable; security, safety, privacy, known incidents and customer commitments need explicit policy logic.

How does RAG help customer support?

RAG retrieves relevant approved articles at request time so the draft can use current product information and keep source IDs. It does not guarantee correctness, so retrieval relevance and answer grounding still need tests.

Which support tickets always need a person?

Route suspected compromise, fraud, privacy requests, identity problems, refunds, account changes, policy exceptions, conflicting sources, low-confidence results and disputed decisions to authorised staff.

Sources

Continue learning

Start with the platform-neutral AI agent build guide

Add human approval and failure recovery

See the AI Agent & Automation Bootcamp

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

How to Enter Tech or Stay in Demand in the AI Era

George K. — Thu, 22 Oct 2026 09:33:22 +0000

New roles such as AI Automation Specialist, AI Agent Developer and AI Transformation Specialist are creating a new route into tech for beginners who can turn business problems into working systems.

Published 1 October 2026 · Event held 30 September 2026 · Barcelona Code School

The clearest message from our panel was that AI automation offers an unusually accessible route into tech. The roles are new, few candidates have years of direct experience, and companies across industries want people who can turn AI ambitions into working automations. A beginner can learn this work without first becoming a programmer, then compete with a portfolio of powerful, connected agent systems rather than an empty job title.

On 30 September, Barcelona Code School Product Manager Alena Dzinman hosted a live Q&A with school founder George Kovalev and Career Advisor Diane Serra. Together, they discussed what AI is changing in hiring, why emerging AI roles can lower the barrier to entering tech, and how experienced professionals can remain valuable as tools and roles evolve.

Alena Dzinman · Product Manager at Barcelona Code School · Host

Diane Serra · Career Advisor at Barcelona Code School · BCS alumna and design and development team leader

George Kovalev · Founder of Barcelona Code School

Five takeaways from the Q&A

Companies want AI transformation, but many still need people who can identify what to automate and build the solution.
New roles include AI Automation Specialist, AI Agent Developer, AI Implementation Specialist and AI Transformation Specialist.
The field has less entrenched seniority competition because almost nobody has many years of direct AI automation experience.
No-code removes the programming prerequisite, not the depth of the work. Beginners can still build complex, connected systems of agents.
Your previous career is not baggage. Combined with new technical skills, it can become your advantage.

AI is changing work, not removing the need for judgement

Diane described a first wave of urgency: leaders wanted visible AI adoption, and some companies assumed that using AI meant they could remove design or other specialist roles. Then the limitations became clearer. Teams still needed people who could judge quality, understand context and take responsibility for the result.

She also sees prospective clients asking agencies for “AI” without being able to define the problem they want to solve. That gap creates an opening for professionals who can move the conversation from a fashionable tool to a useful outcome.

“You still need human discernment.”

- Diane Serra

The opportunity, then, is not simply to know how to open an AI tool. It is to help a company decide where AI is appropriate, what the solution should cost, which risks need controls and how success will be measured.

New AI roles create an opening for beginners

AI-related titles remain inconsistent. Similar work may appear under AI Automation Specialist, AI Agent Developer, AI Adoption Manager, AI Transformation Manager, AI Automation Engineer or AI Implementation Specialist. Diane's advice was to watch the language companies use and keep your CV and LinkedIn positioning current.

“The one I'm seeing most is the AI Automation Engineer.”

- Diane Serra

George explained why this changing market can favour a newcomer. Companies are trying to become AI-first or AI-native and automate work across departments, but the professional field is too new to have a large pool of people with long, directly relevant careers. There is competition, including from candidates around the world, but it is less entrenched than in established technology roles with rigid experience ladders.

“And here's the interesting part: in AI agents and automation, almost nobody has years of experience, because the field itself is new.”

- George Kovalev

That does not make the work automatic or guarantee a job. It removes one of the usual barriers: beginners are not always competing against professionals with ten years in the exact same role. Employers need proof that you can do the work, and a portfolio of functioning AI agents and automations can provide it.

“It's almost like the word "junior" barely applies here.”

- Diane Serra

Can junior developers and designers still get hired?

Yes, but the route may be more strategic than sending the same application to every famous technology company. Diane pointed to midsize companies, startups, small businesses and local government as places where entry-level talent can solve real problems. Not every organisation needs or can afford an all-senior team.

Freelancing can be another route. A new developer or designer can begin with small-business clients, build experience and a portfolio, and grow beyond the junior label through delivered work. Company size also changes what “junior” means: in a small team, responsibility can expand quickly; in a large organisation, progression may involve many formal levels.

“So don't get discouraged if you're just starting out.”

- Diane Serra

What employers want now

For an early-career candidate, employers need proof of practical ability and enough clarity to see where that person fits. For an experienced engineer, Diane looks for something more: the ability to understand what the business needs, propose a solution and continue improving it.

This is where technical execution meets initiative. A strong candidate does not merely say, “I know AI.” They can say, “Here is the process I studied, here is the problem I found, here is the system I built, and here is how I know it helped.”

“From my day-to-day experience, the best developers I have right now are the ones who know AI - meaning they can tell me what the business needs and show me what that looks like, then keep building on the solution they proposed.”

- Diane Serra

A beginner-friendly route into a new tech field

Enter tech through the AI Agent & Automation Bootcamp

You do not need to know how to code before you start. The course is no-code, but the systems you build are not simple: you create a connected Virtual AI Department with n8n workflows, business data, AI agents, knowledge bases, human approvals, testing and monitoring.

In four full-time weeks, you learn how to understand a company's processes, choose what to automate and turn the requirements into a working system. You finish with practical projects that can demonstrate your ability to employers or clients in this fast-emerging field.

Explore the AI Agent & Automation Bootcamp

Your CV, LinkedIn and portfolio should tell one story

A portfolio cannot do all the work if the rest of your professional story points somewhere else. Diane recommended treating your CV, LinkedIn profile and portfolio as three parts of the same narrative: what you know, what you have done and what you want to do next.

That story must also feel true to you. If you can explain it clearly in an interview, a hiring manager can understand how you might contribute to a team. If AI is genuinely part of your skill set, include it in your positioning, but support the label with projects and specific examples.

“The clearer and more coherent that story is across all three, the more prepared you'll be walking into an interview, because you'll be able to speak to it confidently.”

- Diane Serra

Your previous career can be your advantage

Diane entered tech from design, teaching and commercial arts. Her experience since completing a bootcamp has shown her that the industry has room for many backgrounds. Knowledge from healthcare, finance, education, logistics, hospitality, design or another field can help you recognise problems that a purely technical candidate might miss.

The useful combination is domain knowledge plus the ability to build. Take what you already understand about people, processes or an industry; add technical skill; then use both to create better solutions. A career change does not have to erase what came before it.

“Take what you already know, add what you learn in a bootcamp, and bring both to your career.”

- Diane Serra

In-house role, consulting, or both?

George described two valid paths. A larger company may hire someone permanently to build and improve automations, while another business may bring in a consultant for a defined implementation. Some professionals may combine an in-house role with carefully scoped consulting work.

The first step is the same in each case: understand the company's processes, needs, bottlenecks and constraints before choosing a tool or proposing an implementation. Small, recurring tasks may benefit from in-house ownership; a complex one-off project may justify external expertise.

“Yes - the first step is always understanding their processes, their needs, their bottlenecks, and the problems they're facing.”

- George Kovalev

What should you do next?

Choose a direction. Decide whether you want to enter development, design, AI automation, product or another part of tech. You can refine it later, but you need a starting point.
Map your transferable experience. List the industries, users, processes and problems you already understand.
Study real vacancies. Search several related titles and compare the responsibilities and evidence employers request.
Build one complete project. Solve a real, bounded problem and document the decisions, tools, tests and result.
Align your public story. Make your CV, LinkedIn and portfolio support the same target role without pretending to have experience you do not have.
Apply beyond the obvious companies. Include smaller businesses, startups, public organisations, agencies and freelance clients where your combined skills may be especially useful.

The field is open, but proof matters

The panel did not promise an effortless route or a guaranteed job. Competition, persistence and the quality of your work still matter. The opportunity is that AI automation has a lower entry barrier than many established technology tracks: the roles are new, demand is spreading across industries, and no-code tools let beginners build substantial systems without first spending years learning software development.

You do not need to predict the final job title. You need to understand a business problem, build something useful and explain the value clearly. The AI Agent & Automation Bootcamp is designed to help a beginner produce exactly that kind of evidence, while experienced professionals can use the same skills to stay valuable as the tools change.

Continue exploring your route into tech

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

n8n Tutorial: Build Your First AI Workflow

George K. — Wed, 21 Oct 2026 07:57:25 +0000

Build a small, testable workflow that turns a fictional customer enquiry into structured triage data and a reply draft, without sending anything outside n8n.

Build your first n8n AI workflow with seven nodes: a manual trigger, mock input, validation, one bounded AI step, output validation, a review-ready result and an error route. Keep the workflow in manual test mode, use fictional data and stop before any external action. The first success criterion is predictable structured output, not automatic email or CRM updates.

Key takeaways

Start with manual execution and fictional data, not a live webhook or inbox.
Give the AI one job: classify an enquiry and draft a reply in a fixed schema.
Use deterministic nodes to validate input and model output.
End with a review record; add external actions only after testing.
An error path and an execution log are part of the first workflow.

What will you build?

The workflow receives a fictional customer message, checks that the required fields exist, asks an AI model to return a category and reply draft, validates the result and produces a record for human review. It does not send an email or update a live CRM.

This is an AI workflow, not yet an autonomous AI agent. The route is chosen in advance, and the model performs one narrow language task. That boundary keeps the first project easier to understand and test. For the general agent architecture, use the BCS guide to building an AI agent.

What input and output does the workflow use?

Field	Type	Required	Rule
`request_id`	string	Yes	Unique fictional test identifier
`customer_message`	string	Yes	Non-empty; no real personal data
`consent_to_process`	boolean	Yes	Must be true before a model call
`category`	enum	Output	Only `sales`, `support` or `other`
`reply_draft`	string	Output	Draft only; no claims outside supplied facts
`needs_human`	boolean	Output	True for uncertainty, missing facts or sensitive requests

A test record can use a message such as: “I run a small design studio and want to understand which course teaches workflow automation.” Do not paste a real customer conversation into an unfinished workflow.

{
  "request_id": "test-001",
  "customer_message": "I run a small design studio and want to learn workflow automation.",
  "consent_to_process": true
}

What does the finished workflow look like?

Figure 1. Seven-node first AI workflow

1. Manual Trigger · Run only when the builder starts a test.
2. Mock input · Add request ID, message and consent.
3. Input check · Reject missing text or false consent.
4. AI step · Return category, draft and human flag.
5. Output check · Allow only the defined schema and labels.
6. Review output · Store or display the draft; do not send.
7. Error route · Record failure and alert the workflow owner.
Execution log · Keep input, route, output and status.

The model performs one bounded language task. The surrounding workflow controls data, validation, review and failure handling.

Text equivalent: manual execution creates a fictional record. An input check blocks incomplete or unauthorised data. The AI step returns three defined fields. A second check rejects unknown labels or missing fields. Valid results become a review-ready draft. Errors are recorded and routed to a person instead of being hidden.

4-week live bootcamp · Barcelona or online

Go from one workflow to a complete AI automation system

In the AI Agent & Automation Bootcamp you build a no-code Virtual AI Department with n8n, RAG, connected agents and human approvals — live with an instructor and a small group.

See the bootcamp programme →

Not sure it fits? Take the free AI Career Roadmap test

How do you build the workflow step by step?

Create a manual test trigger.

Open a new workflow and start with the manual trigger available in your n8n version. Manual execution prevents the unfinished workflow from accepting live events. Name the workflow clearly, such as TEST - Enquiry triage draft.

Add fictional input data.

Add a field-setting node and create the three input fields shown above. Use a new request_id for each test. Execute this node and inspect its output before connecting anything else.

Validate required fields.

Add conditional logic that continues only when customer_message is not empty and consent_to_process is true. Route the false branch to a safe result such as status: rejected_input. The model should never repair missing consent.

Add one bounded AI step.

Connect a supported model credential using the narrowest access available. Ask the model to classify the message as sales, support or other, draft a short reply using only supplied facts and set needs_human. Require structured output. Node names and options can change, so confirm them in the documentation for your installed version.

   Classify the customer message as sales, support, or other.
   Draft a reply using only facts in the input.
   Set needs_human to true if facts are missing or the request is sensitive.
   Return only: category, reply_draft, needs_human.

Validate the model output.

Add a deterministic check after the model. Confirm that all three fields exist, the category is one of the allowed labels and the draft is within your chosen length. Route every unexpected result to status: needs_review. Do not ask the same model to approve its own output.

Create a review-ready output.

Map the validated fields into a final record with the original request ID, model result and status: draft_ready. n8n supports mapping data from previous nodes through expressions or the input pane. Inspect the actual node output before choosing a field path.

Add errors, tests and activation rules.

Create an error workflow or alert route that includes the workflow name, execution ID and failure message without exposing secrets. Keep the workflow inactive until the test matrix passes. If you later add email or CRM actions, require human approval and a duplicate-safe operation ID first.

Where do permissions and human approval belong?

The tutorial needs a model credential but no live customer-system credential. Keep the scope small and never place a secret inside prompt text, mock data or a screenshot. Anyone who can run or edit the workflow may be able to use connected credentials depending on the n8n role and project configuration.

A future send step belongs after output validation and human review. n8n documents simple send-and-wait approval through supported communication nodes, and suggests the Wait node for more complex approval processes. Approval should show the exact recipient, subject and message that will be sent.

The first workflow finishes at “draft ready.” External action is a separate capability with its own permission, approval and test cases.

What should happen when a node fails?

Record the execution ID, failing node, error type and request ID. A malformed input should stop without calling the model. An invalid model result should enter review. A temporary API error may be retryable, but only within a set limit.

n8n execution history can show failed, running, successful and waiting runs, and failed executions can be retried. Once a workflow has external side effects, never retry blindly: first check whether the previous run already created the record or sent the message.

How do you test the workflow?

Case	Input	Expected result
Valid sales enquiry	Complete fictional record	Allowed category and `draft_ready`
Empty message	Blank text	`rejected_input`; no model call
No consent	Consent false	`rejected_input`; no model call
Ambiguous request	Insufficient facts	`needs_human: true`
Unexpected category	Model returns another label	`needs_review`; no external action
Prompt injection text	Message asks to ignore workflow rules	Treated as customer data; schema unchanged
Model timeout	Simulated provider error	Error route records execution and alerts owner
Duplicate request ID	Repeat `test-001`	Detected before any future external write

Criteria for success

Only complete, consented fictional records reach the model.
The AI returns exactly the three defined fields.
Unknown labels and missing fields enter human review.
The workflow creates a draft but sends nothing externally.
Every failure is visible with an execution and request ID.
No credential or personal data appears in the workflow example.
All eight test cases reach their expected result.

Go from one workflow to a complete automation system

This tutorial is enough for a first controlled build. Barcelona Code School’s AI Agent & Automation Bootcamp is for people who want deeper practice with n8n, business data, tool-using agents, RAG, approvals, QA, error handling, monitoring and multi-agent orchestration in one connected portfolio project.

The bootcamp is a four-week, full-time career programme; you do not need it to test one small workflow. Use the live page for current dates, format, tuition and prerequisites.

Explore the AI Agent & Automation Bootcamp

Frequently asked questions

What should a first n8n AI workflow do?

A good first workflow should complete one bounded task with a clear input and output. For example, it can turn a fictional enquiry into a category and reply draft. It should not send messages, spend money or change important records while you are still testing it.

Do I need an AI Agent node for this n8n tutorial?

No. A direct model or extraction step is enough when the workflow needs one bounded judgement. Use an AI Agent node only when the system must choose among tools or decide several steps within defined limits.

How do I pass data between n8n nodes?

n8n lets you map values from previous nodes into later node parameters with expressions or by selecting data from the input pane. Check the actual input and output of each node rather than assuming a field name.

Should the first workflow send the AI-written email?

No. End the first version with a review-ready draft. After the workflow passes its tests, add a separate approval step before any email node sends to a real recipient.

How do I know the n8n workflow is ready to activate?

Activate it only when valid inputs produce the expected output, invalid data stops safely, unexpected model output enters review, credentials have minimum access, errors alert an owner and retries cannot duplicate an external action.

Sources

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

AI Agent Guardrails: Human Approval and Failure Recovery

George K. — Tue, 20 Oct 2026 08:08:20 +0000

A practical control system for deciding what an agent may do, what a person must approve and how the workflow returns to a safe state when something fails.

Reliable AI agents use several controls, not one safety prompt.

Give the agent least-privilege access, validate inputs and tool calls, require human approval before consequential side effects, verify the real result and define recovery states for every failure. If the system cannot establish what happened, it should stop, preserve the evidence and escalate instead of guessing or repeating the action.

Key takeaways

Permissions decide what the system can do; instructions only tell the model what it should do.
Guardrails should cover inputs, outputs and individual tool calls.
Approval belongs immediately before a consequential side effect, with enough context for a real decision.
Retries need limits and idempotency; an unknown outcome is not a failed action that can safely be repeated.
Every production workflow needs a human queue, action log and tested recovery path.

What are AI agent guardrails?

AI agent guardrails are checks and limits placed around the model and the tools it can use. An input guardrail can reject missing or suspicious data. A tool guardrail can inspect a proposed action before execution and validate the returned result. An output guardrail can check the final answer for required evidence, format or policy.

Guardrails are one layer. They do not replace authentication, authorisation or database constraints. If an agent has a tool capable of deleting every customer record, “do not delete important data” is not an adequate permission model. The application should expose only the narrow action and records the workflow needs.

A prompt is guidance for the model. A permission is an enforced boundary. Use both, but never confuse them.

Which controls belong in the system?

A practical design uses five layers. First, identity and permissions establish who initiated the task and which resources are available. Second, input validation checks structure, provenance and required fields. Third, policy guardrails evaluate the proposed decision and tool arguments. Fourth, a human gate pauses high-impact actions. Fifth, verification and recovery determine what really happened and what to do next.

Action	Default control	Reason	Safe alternative
Read an approved product document	Automatic, logged	Reversible, low impact	Deny access outside the approved collection
Create an internal draft	Automatic with output validation	No external side effect	Mark as draft and cite source fields
Update a low-risk CRM field	Policy check; approval depends on field	May affect operations	Write to a review queue or staging field
Send, publish or make a commitment	Human approval before execution	External and reputational consequence	Prepare a preview only
Spend money, change access or delete data	Human approval plus strong authorisation	Financial, security or irreversible impact	Propose the change; do not execute
Medical, legal or hiring judgement	Human specialist decision; agent cannot decide	Consequential and context-sensitive	Summarise evidence without advice or selection

Where should human approval happen?

Place the approval gate after the agent has prepared a concrete action but before the tool creates the side effect. The reviewer should see the action type, target, important arguments, source evidence, policy result and what will happen after approval. A vague “allow agent?” prompt does not support an informed decision.

OpenAI’s Agents SDK documents a pause-and-resume pattern: a tool call can surface as an interruption, a person approves or rejects it, and the run continues from stored state. Its approval rules fail closed when arguments cannot be safely inspected. The implementation will vary by platform, but the principle is portable: unclear input must not silently become permission.

Figure 1. A controlled action path

1. Validate · Identity, required fields, source and scope.
2. Decide · Policy, risk class, tool and proposed arguments.
3. Approve · Automatic only for low risk; human gate for consequences.
4. Execute · Use a narrow tool and unique operation ID.
5. Verify · Check the tool response and independent system state.
6. Complete · Record evidence, result, actor and timestamp.
Safe stop · Reject, quarantine or request missing information.
Human queue · Escalate failures, conflicts and unknown outcomes.

The safe path validates before action and verifies afterwards. Approval is a branch in the execution system, not a sentence in the prompt.

Text equivalent: a request passes identity and input validation, policy classification and an approval decision. A low-risk action may execute automatically; a consequential action pauses for a person. After execution, the system verifies the external state and logs completion. Invalid, rejected, failed or ambiguous cases stop or enter a human queue.

5-day intensive · Barcelona or online live

Build five AI agent projects in one week

AI Agents Builder is a full-time, instructor-led week with n8n. You finish five working AI agent projects in a small group — on campus in Barcelona or online live.

See the AI Agents Builder course →

Not sure it fits? Take the free AI Career Roadmap test

How do you design guardrails in seven steps?

Map actions and consequences. List every read, write, send, publish, purchase and delete operation. For each one, record the target, affected person, reversibility and worst credible outcome.
Set permissions outside the prompt. Give the agent the smallest useful tools, scopes and data access. Separate read from write, and expose narrow business operations instead of broad administrator access.
Validate inputs. Require a schema, source and minimum fields. Quarantine malformed, unauthorised or suspicious content. Treat retrieved documents and tool output as untrusted data, not new instructions.
Place approval before side effects. Define which risk classes always pause. Show the reviewer a stable preview, and bind approval to the exact action and arguments so later changes require a new decision.
Verify outputs and tool results. Check required fields, evidence and policy. After a tool runs, confirm the external record rather than trusting a generated success message.
Define recovery states. Distinguish retryable errors, permanent errors, rejected actions and unknown outcomes. Give each state an owner, time limit and next action.
Test and monitor. Run normal, boundary, adversarial and failure cases before release. Monitor approval rates, safe stops, duplicate attempts, unresolved incidents and changes in source data.

How should an agent recover from failure?

A failure handler needs more precision than “try again.” A timeout may mean the tool did nothing, or it may mean the action succeeded but the response never arrived. Repeating a payment, email or database write could create a duplicate. Use a unique operation ID, check the destination state and retry only when the operation is designed to be idempotent.

Observed state	Response	Do not do
Missing required data	Request the specific field; preserve the draft	Invent a value
Policy or guardrail rejection	Stop and explain the rule to the operator	Rephrase to bypass the control
Human rejects approval	Cancel that exact action and record the decision	Ask repeatedly or switch tools
Transient tool error with confirmed no-write	Retry with backoff within a strict limit	Retry forever
Timeout or ambiguous result	Query by operation ID; escalate if still unknown	Assume failure and repeat a side effect
Partial multi-step update	Run a tested compensation or human runbook	Hide the partial state
Repeated or novel failure	Open an incident, stop the workflow and preserve logs	Let the model improvise a production fix

What should you test before release?

Use a fixed test set with expected states, not only examples that should succeed. The system passes when it produces the expected result or the expected safe stop without an unapproved side effect.

Test	Expected result	Evidence to capture
Valid low-risk request	Completes automatically	Input, tool call, verified result and log
Missing or malformed fields	Stops before tool execution	Validation reason and zero writes
Instruction hidden in retrieved content	Treated as data; policy remains unchanged	Source, blocked instruction and chosen action
Consequential action	Pauses before execution	Approval preview and exact arguments
Rejected approval	Action cancelled	Decision, reviewer and zero side effects
Tool timeout after possible write	No blind retry; state checked	Operation ID and destination lookup
Duplicate request	One outcome only	Idempotency record
Maximum attempts reached	Safe stop and human escalation	Attempt count, last error and queue item

Criteria for success

No action can exceed the initiating user’s permission.
Every consequential side effect is approved against exact arguments.
Invalid and adversarial inputs stop before a protected tool runs.
Every external action has a verifiable result and audit record.
Retries cannot create duplicate side effects.
Unknown outcomes and repeated failures reach a named human owner.
Visible FAQ, runbook and monitoring rules match the implemented behaviour.

Learn to make one AI workflow reliable

Barcelona Code School’s AI Agents Builder course teaches professionals to turn a real work process into a controlled system. Students define inputs and outputs, connect approved tools, handle bad inputs, add human approval and test what can go wrong.

The course is for practical workplace automation, not a promise of risk-free autonomy. Check the live page for the current format, dates and tuition.

Explore the AI Agents Builder course

Frequently asked questions

What are guardrails for an AI agent?

AI agent guardrails are checks and limits around inputs, outputs and tool calls. They can validate data, block prohibited actions, require evidence, constrain formats and stop a run. They work best with real permissions, approval gates, logging and recovery paths.

Which AI agent actions need human approval?

Require approval for actions with financial, legal, privacy, reputational or operational consequences: sending important messages, publishing, spending money, deleting data, changing access, making commitments or acting on ambiguous evidence. The exact boundary depends on the business.

Is human approval enough to make an AI agent safe?

No. A reviewer can miss problems, and an approval screen may omit important context. Use least-privilege permissions, input and output validation, application-level authorisation, logs, post-action verification and safe recovery as well as human approval.

What should happen when an AI agent fails?

The agent should record the failure, avoid repeating unsafe side effects, classify the outcome and either retry within a strict limit, request missing information, compensate through a tested action or escalate to a person. Ambiguous outcomes should not be treated as success.

How do you test AI agent guardrails?

Test valid work, missing fields, malformed data, prompt injection, excessive permissions, rejected approvals, tool timeouts, duplicate requests, partial writes and ambiguous responses. Verify that each case reaches its expected result, safe stop or human queue without an unapproved side effect.

Sources

Continue learning

Build your first AI agent

Choose between an agent, chatbot and workflow

Compare BCS AI learning paths

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

How to Build an AI Appointment Booking Agent for a Hair Salon

George K. — Sun, 18 Oct 2026 06:26:45 +0000

One agent, one business workflow

Let AI understand “a trim after work with anyone available.” Let rules and live calendar data decide what can actually be booked.

Published 28 September 2026 · Barcelona Code School

Build a hair salon booking agent by separating language from scheduling.

The AI agent interprets the customer's request, identifies missing details and selects approved read-only tools. A deterministic workflow retrieves the service duration, eligible stylists and real calendar availability. Only after the customer confirms a specific option should the workflow recheck the slot, create the appointment, record the action and send a confirmation.

Key takeaways

The language model may interpret “colour and a trim,” but the salon's service catalogue must supply the official duration, price and eligible staff.
Availability must come from the booking system or calendar, never from the model.
Check the slot twice: when proposing it and immediately before creating the appointment.
Require explicit confirmation of service, stylist, date, time and price before the booking tool can write.
Escalate exceptions, repeated failures, complaints and ambiguous services to salon staff.

What should the salon booking agent do?

The agent has one job: move a booking request from natural language to a confirmed appointment without guessing business facts. A customer might ask, “Can I get highlights next Friday afternoon with Maria?” The agent extracts the service, preferred stylist and time range, then checks structured salon data.

The workflow should support a small, explicit action set: answer booking-policy questions, find services, find eligible staff, search real availability, propose slots, create a confirmed booking and request staff help. Rescheduling and cancellation can be added later after the basic booking path is reliable.

Business boundary

The agent does not decide prices, invent duration, override working hours, squeeze a client into a busy calendar or promise that a complex service is suitable. Consultation-required services, policy exceptions and unclear requests go to a person.

What data must exist before you build the agent?

A booking agent cannot repair missing business rules with better prompting. Create a structured service catalogue first. Give every service and staff member a stable ID so changes to display names do not break the workflow.

Data set	Minimum fields	Why the agent needs it	Owner
Services	`service_id`, name, duration, buffer, price, consultation flag	Prevents invented duration and price	Salon manager
Staff	`staff_id`, services, calendar ID, working hours, active status	Limits each service to qualified, available staff	Salon manager
Policies	lead time, cancellation window, late-arrival rule, deposit rule, version	Keeps answers consistent and auditable	Business owner
Calendar	busy intervals, holidays, breaks, time zone	Supplies real availability	Booking system
Booking	service, staff, start/end, customer contact, consent, status, idempotency key	Creates one traceable appointment	Workflow

What architecture prevents invented slots and double bookings?

Figure 1. Hair salon booking state machine

1. Understand

Extract service, preferred staff, date range and contact channel. Ask for missing fields.

2. Search

Read service rules and eligible calendars. Return two or three genuinely free slots.

3. Confirm

Repeat the exact service, staff, time, duration and price. Wait for an explicit yes.

4. Commit

Recheck availability, create once, log the result and send confirmation. Escalate on conflict.

The model controls interpretation and questions. Deterministic tools control catalogue facts, time calculations, availability and calendar writes.

The text equivalent of the diagram is: interpret the request; validate the service; retrieve eligible staff and live busy intervals; calculate candidate slots; obtain customer confirmation; recheck the selected slot; create one booking; record the result. Any unresolved ambiguity, calendar conflict or tool failure leaves the booking unconfirmed and routes the case to staff.

How do you build the salon booking agent step by step?

Define the booking boundary. Write the actions the agent may perform and the actions it must refuse or escalate. Begin with new bookings only. Keep refunds, deposits, disputes and unusual service combinations outside the first version.
Create the service catalogue. Store official service IDs, names, aliases, durations, buffers, prices, consultation flags and eligible staff. Do not put these changing facts only inside the prompt.
Define booking states. Use explicit statuses such as collecting_details, searching, awaiting_confirmation, committing, confirmed, conflict and human_help. A message is not a confirmed booking unless the write tool returned a booking ID.
Connect read-only tools first. Give the agent tools such as get_services, get_eligible_staff and check_availability. Keep create_booking unavailable until validation and confirmation pass.
Write narrow agent instructions. Tell the model to use catalogue data for facts, ask when the service is ambiguous, show no more than three suitable slots and never describe a booking as confirmed before receiving a booking ID.
Calculate candidate slots deterministically. Query busy ranges for eligible staff, subtract them from working hours, then apply service duration, buffer, holidays and lead-time rules. The model selects how to present valid options; it does not perform the calendar arithmetic.
Require explicit customer confirmation. Show service, stylist or “any available stylist,” local date, local time, duration and current catalogue price. Store the customer's affirmative response with the chosen slot.
Recheck and create atomically. Query availability again immediately before the write. Create the event only if the slot is still free. Use an idempotency key derived from the booking attempt so a network retry cannot create a duplicate.
Log, notify and test. Store the request, selected catalogue IDs, tool results, confirmation, booking ID and final status. Test normal bookings, ambiguity, conflicts and failures before allowing customers to use the workflow.

4-week live bootcamp · Barcelona or online

Go from one workflow to a complete AI automation system

In the AI Agent & Automation Bootcamp you build a no-code Virtual AI Department with n8n, RAG, connected agents and human approvals — live with an instructor and a small group.

See the bootcamp programme →

Not sure it fits? Take the free AI Career Roadmap test

What should the agent instructions contain?

The following is a design example, not a complete production prompt. Tool permissions and business rules must also be enforced in the workflow:

ROLE
You help customers find and request hair salon appointments.

SOURCE OF TRUTH
Use get_services for duration, price and consultation rules.
Use check_availability for slots. Never invent either.

REQUIRED BEFORE SEARCH
service_id, preferred date or range, customer timezone.

REQUIRED BEFORE WRITE
selected service, staff or any-staff choice, exact start and end,
customer name, confirmed contact method, explicit confirmation.

STOP AND ESCALATE
Ambiguous service after two questions; consultation-required service;
policy exception; complaint; calendar error; two failed tool attempts.

CONFIRMATION RULE
Never say “booked” until create_booking returns booking_status=confirmed
and a booking_id.

How should the n8n workflow be divided?

n8n can coordinate the conversation trigger, structured data, AI Agent node, calendar operations, validation branches, notifications and logs. Keep the model-facing tools small and precisely named. The AI agent should call read-only tools during discovery; a normal workflow branch should perform the final write after confirmation.

Component	Responsibility	Must remain deterministic?
Message or form trigger	Receive request and attach channel/customer identifiers	Yes
Input validator	Check type, length, required identifiers and allowed action	Yes
AI Agent	Interpret request, ask questions and select approved read tools	No, but bounded
Service/staff lookup	Return structured catalogue records	Yes
Availability calculator	Combine working hours, duration, buffers and busy intervals	Yes
Confirmation gate	Verify selected slot and explicit affirmative response	Yes
Calendar write	Recheck and create one appointment	Yes
Error and human route	Stop retries, open staff task and preserve context	Yes

How do you prevent double-booking?

Checking availability once is not enough. Another customer or receptionist may take the slot while the agent waits for confirmation. The write path must check the selected staff calendar again immediately before creation.

Google Calendar's Freebusy endpoint returns busy periods for specified calendars and time ranges. Event creation uses an event with start and end values. Google also documents that a caller-provided event ID can help keep a local system in sync and prevent duplicate event creation after an uncertain operation. If the salon uses a dedicated booking platform, apply the equivalent availability and idempotency controls supported by that platform.

Which actions need human help?

A new customer cannot identify the service after two clarification questions.
Colour correction, extensions or another service requires consultation.
The customer requests an exception to price, timing, deposit or cancellation policy.
The calendar or booking tool returns an error twice.
The requested stylist is unavailable and the customer rejects alternatives.
The customer makes a complaint or asks for compensation.
The workflow detects conflicting customer or booking records.

Human escalation should include a compact summary, collected fields, attempted tools, the last error and the exact decision required. Staff should not have to restart the conversation.

How do you test the agent before launch?

Test	Expected behaviour	Pass condition
Clear service, flexible stylist	Offers valid slots for eligible staff	All proposed slots are free and long enough
Ambiguous “colour” request	Asks which catalogue service applies	No duration, price or slot invented
Preferred stylist unavailable	Offers other times or eligible staff	Customer preference remains visible
Customer changes date	Discards previous candidates and searches again	No stale slot can reach write gate
Slot taken during confirmation	Recheck fails and new options are offered	No overlapping booking created
Duplicate message or retry	Returns existing result for same attempt key	Only one booking ID exists
Calendar unavailable	Stops after retry threshold and escalates	No claim that booking succeeded
Consultation-required service	Routes to staff or consultation flow	Standard appointment not created
Policy exception	Quotes written policy and requests staff decision	Agent does not override policy
Prompt injection in customer text	Treats text as customer data, not instructions	Tool scope and system rules unchanged
Missing contact method	Requests one before write	Booking cannot enter committing state
Successful booking	Returns confirmed details and booking ID	Calendar, log and customer message agree

Criteria for success

No suggested slot conflicts with the source-of-truth calendar at suggestion time.
No booking is created without explicit confirmation and a second availability check.
Repeated delivery of the same confirmed request creates no duplicate booking.
Every confirmed message maps to a stored booking ID and audit event.
Unsupported services and tool failures reach staff with useful context.
Test results are recorded before the workflow is opened to real customers.

Learn to build the whole workflow, not only the chat

Barcelona Code School's current AI Agent & Automation Bootcamp is a four-week, full-time, no-code programme built around n8n. The live curriculum covers workflows, tool-using agents, structured business data, human approvals, RAG, QA, reliability and connected business systems.

This programme fits people who want to design and build complete business automations. If you only want a scripted FAQ bot, a shorter chatbot tutorial may be enough.

See the current AI Agent & Automation Bootcamp

The live course page is the source for current format, dates, tools and tuition.

Frequently asked questions

Can an AI agent book hair salon appointments automatically?

Yes, if the agent reads real availability through an approved tool and a deterministic workflow validates the booking before writing it. The agent should never invent a slot or create an appointment before the customer confirms the service, staff, date and time.

How does a salon booking agent avoid double-booking?

The workflow checks availability when suggesting slots and checks it again immediately before the write. The booking operation should use an idempotency key so a retry cannot create the same appointment twice. If the slot changed, the workflow offers new options instead of forcing the booking.

What information does a hair salon booking agent need?

Before searching, it needs a service and preferred date or range. Before booking, it needs the selected staff member or any-staff preference, confirmed start and end, customer name, one contact method, timezone, policy version and explicit customer confirmation.

Should the language model calculate appointment duration?

No. Duration, preparation time and cleanup buffer should come from the salon's structured service catalogue. The model may identify the likely service from the customer's words, but it must ask when the choice is ambiguous.

Can this workflow use n8n and Google Calendar?

Yes. n8n provides AI Agent and Google Calendar nodes, while Google Calendar provides free-busy lookup and event creation. A salon using another booking system should keep that system as the source of truth and connect only through its supported API or integration.

When should the booking agent hand the conversation to salon staff?

Escalate when the requested service is unclear, the customer asks for an exception, calendar access fails, the same action fails repeatedly, a complaint appears, a complex colour correction needs consultation or the requested change falls outside the salon's written policy.

Sources

OpenAI: A practical guide to building agents — agent components, tool use, guardrails and human intervention.
n8n documentation: AI Agent node.
n8n documentation: Google Calendar node.
Google Calendar API: Freebusy query — calendar busy intervals.
Google Calendar API: Create events — event fields and duplicate-prevention option.
Barcelona Code School: How to Build an AI Agent — canonical general build guide.
Barcelona Code School: AI Agent & Automation Bootcamp — current course information.

Continue learning

Read the canonical guide to building an AI agent

Compare Barcelona Code School's AI courses

Explore the AI Agent & Automation Bootcamp

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

Single Agent vs Multi-Agent System: How to Choose

George K. — Fri, 16 Oct 2026 12:44:36 +0000

Architecture decision

More agents do not automatically produce a better system. Compare the architectures by ownership, context, permissions, cost and failure paths.

Published 28 September 2026 · Barcelona Code School

Start with a single AI agent unless the task has a proven reason to split. Add specialist agents when work contains independent branches, distinct tool permissions, separate evaluation criteria or parallel research that one agent cannot handle reliably. A multi-agent system may improve focus and scale, but it also adds routing errors, context loss, latency, cost and coordination failures.

Key takeaways

One agent with clear tools is the default architecture for one bounded workflow.
A manager agent keeps ownership and calls specialists for narrow subtasks.
A handoff transfers active control to a specialist and must define the context that moves.
Parallel agents fit independent branches; dependent steps still need ordered orchestration.
Compare multi-agent results with a single-agent baseline before accepting extra complexity.

What is the difference between a single-agent and multi-agent system?

A single-agent system gives one model-driven agent the instructions, context and tools needed to complete a task. The same agent plans, calls tools, checks results and produces the final output. It may use many deterministic tools without becoming a multi-agent system.

A multi-agent system contains two or more model-driven agents with separate roles. They may be coordinated by a manager, transfer control through handoffs or run independent branches in parallel. OpenAI’s Agents SDK documents manager-as-tools and handoffs as two common patterns. These are architecture choices, not maturity levels.

Figure 1. Three practical agent architectures

Single agent: Request → One agent → Tools → Result. One owner keeps the full context and decides which deterministic tool to use.
Manager + specialists: Manager → Research agent → Data agent → Writer agent. The manager calls specialists as tools, combines their outputs and owns the final result.
Triage + handoff: Triage → Billing → Booking → Support. The router chooses one specialist, which becomes responsible for the next part of the conversation.

The key question is who owns the next decision and final result. More boxes are justified only when they create a useful boundary.

Text equivalent: the single-agent pattern has one agent between the request and its tools. In the manager pattern, one agent owns the result while specialist agents perform bounded subtasks. In the handoff pattern, a triage agent routes the request and the chosen specialist takes control.

How do the architectures compare?

Criterion	Single agent	Manager with specialists	Handoffs
Final owner	One agent	Manager	Active specialist
Context	Usually shared in one run	Manager selects specialist input	Transferred and optionally filtered
Best fit	One bounded workflow	Several narrow subtasks feeding one result	Routing to distinct service domains
Control	Simplest trace	Central policy point	Distributed after transfer
Parallel work	Limited by one loop	Possible for independent calls	Not the main purpose
Typical risk	Overloaded prompt or tool choice	Bad delegation or conflicting outputs	Wrong route or lost context
Cost and latency	Usually lower	More model calls	More turns and transfer overhead
Testing	Task and tool loop	Agents, routing and synthesis	Agents, route and context transfer

When is one agent enough?

Use one agent when one role can understand the goal, access the necessary context and select from a manageable tool set. Lead intake, appointment preparation, support drafting and internal reporting often begin this way. A tool that calculates, searches or writes a record does not need its own agent if it follows deterministic logic.

A single agent is easier to trace because one instruction set owns the decision loop. It also avoids repeated context packaging and conflicting specialist outputs. Anthropic’s architecture guidance recommends starting with the simplest system and adding complexity only when it produces a measured improvement.

Do not create a specialist agent for every tool. Create a specialist only when it needs its own judgement, context, permissions or evaluation.

When does a multi-agent system help?

Separate agents can help when a task divides into independent areas that require different instructions or tool access. A research manager might ask parallel agents to explore distinct markets, then compare their evidence. A customer-service router might hand a billing case to a billing specialist with limited account tools.

Parallel research is a strong example because branches can explore different directions without waiting for each other. Anthropic reported that its own multi-agent research system outperformed its single-agent baseline by 90.2% on an internal research evaluation. That result belongs to that system and test set; it does not establish a general advantage for multi-agent architecture.

Good reasons to split

independent work can run in parallel;
specialists need different data or tool permissions;
each specialty has a distinct evaluation set;
one prompt has become difficult to understand and maintain;
a routing boundary mirrors a real operational boundary;
one manager must combine several evidence-producing subtasks.

4-week live bootcamp · Barcelona or online

Go from one workflow to a complete AI automation system

In the AI Agent & Automation Bootcamp you build a no-code Virtual AI Department with n8n, RAG, connected agents and human approvals — live with an instructor and a small group.

See the bootcamp programme →

Not sure it fits? Take the free AI Career Roadmap test

Should you use a manager or handoffs?

Use a manager when one component must retain the user context, apply shared policy and produce the final answer. Specialists behave like tools: they receive a bounded task and return a result. This is useful for reports, proposals and investigations where several contributions become one deliverable.

Use handoffs when routing is the workflow. The specialist takes over, such as billing, booking or technical support. OpenAI’s handoff documentation allows typed routing metadata and context filters. The design still needs an application-level permission check before any side effect; a model-generated handoff reason is not authorisation.

How do you choose in seven steps?

Define one measurable result. State what finishes the run and who owns the business outcome.
Build a single-agent baseline. Give one agent only the context and tools needed for the task.
Record the real failure. Identify whether failures come from context overload, tool confusion, missing expertise, latency or weak source data.
Test a simpler fix. Improve tool descriptions, structured outputs, retrieval and deterministic routing before adding another agent.
Name the boundary. Give a new agent separate instructions, tools, permissions and evaluation only when those boundaries matter.
Choose manager, handoff or code orchestration. Keep one owner for synthesis, transfer control for service routing, or use code when the sequence must be predictable.
Compare against the baseline. Measure task success, routing, evidence, latency, cost and recovery on the same test set.

What can fail in a multi-agent system?

Failure	Control	Safe outcome
Wrong specialist selected	Structured route labels and route evals	Return to triage or human queue
Required context omitted	Typed handoff payload and minimum fields	Specialist requests missing context
Too much sensitive context transferred	Context filter and least privilege	Transfer only task-required fields
Specialists disagree	Source requirements and manager comparison	Show conflict; do not invent consensus
Agents call each other repeatedly	Maximum turns and visited-route state	Stop and escalate with logs
Parallel branch times out	Branch timeout and partial-result policy	Mark missing branch explicitly
Approval is bypassed after handoff	Application-level action gate	No side effect without approval

How should you test the decision?

Test each agent in isolation, then every routing or handoff boundary, then the whole system. Use the same normal, ambiguous and adversarial cases for the single-agent and multi-agent versions. Record every agent call, tool call, transfer, source and approval result so a reviewer can reconstruct the run.

Criteria for success

The chosen architecture completes the defined task more reliably than the baseline.
Every route is correct or fails into a safe, visible queue.
Transferred context is sufficient but contains no unnecessary sensitive data.
Specialist outputs use a defined schema and cite their evidence.
Loops stop at the configured limit.
Consequential actions remain behind the same approval gate after every handoff.
The improvement justifies additional latency, cost and operational work.

Learn orchestration by building the whole workflow

Barcelona Code School’s AI Agent & Automation Bootcamp covers tool-using agents, shared business data, human approvals, RAG, QA, reliability and multi-agent orchestration with n8n. It fits people who want to design connected business systems and test them end to end.

If one prompt and two tools solve your task reliably, you do not need a multi-agent system or a full bootcamp. Use the live course page for current format, dates and tuition.

Explore the AI Agent & Automation Bootcamp

Frequently asked questions

Is a multi-agent system better than a single agent?

Not by default. A multi-agent system can help when work splits into independent specialties, permissions or parallel branches. It also adds handoffs, context transfer, latency, cost and new failure modes. Start with one agent and add specialists only when evaluation shows a clear improvement.

When should I use a single AI agent?

Use one agent when one instruction set, context window and tool set can complete the task reliably. A single agent is usually easier to test, trace and operate, especially for a first version or a workflow with one clear owner.

What is a manager agent?

A manager agent keeps control of the task and calls specialist agents as tools. It combines their outputs and owns the final response. This pattern suits work where one component must enforce shared policy, maintain context and assemble the result.

What is an AI agent handoff?

A handoff transfers active control from one agent to a specialist. It suits routing workflows where the selected specialist should own the next part of the conversation. The system must define what context and permissions move with the handoff.

How do you test a multi-agent system?

Test each agent, every routing decision, the transferred context and the complete workflow. Include wrong routes, missing context, conflicting outputs, timeouts, repeated handoffs, permission failures and approval rejection. Compare the result with a single-agent baseline.

Sources

Continue learning

Build one reliable AI agent first

Compare BCS AI learning paths

Explore multi-agent orchestration training

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

What Is n8n and How Is It Used for AI Automation?

George K. — Thu, 15 Oct 2026 07:57:21 +0000

n8n connects triggers, data, AI models and business applications in one visible workflow. The useful question is not whether every step can use AI, but where AI judgement improves a controlled process.

Published 28 September 2026 · Barcelona Code School

n8n is a workflow automation platform that connects apps, APIs, databases and AI models through visual nodes.

For AI automation, n8n can receive an event, prepare data, ask a model to classify, extract or draft, apply deterministic rules, request human approval and update a business system. n8n orchestrates the process; the AI model performs only the bounded judgement steps.

Key takeaways

n8n is an orchestration layer, not an AI model.
A workflow begins with a trigger and moves data through connected nodes.
Fixed rules belong in deterministic nodes; AI belongs where language or variable content needs interpretation.
Credentials, approvals, execution logs and error paths are part of the workflow design.
Cloud is simpler to operate; self-hosting adds control and infrastructure responsibility.

What is n8n?

n8n describes itself as a fair-code licensed workflow automation tool that combines AI capabilities with business process automation. It can connect applications that expose APIs, transform their data and pass results between systems with little or no code. Workflows are assembled from nodes, and each node performs a defined job.

A trigger starts the workflow. It might be a form submission, a webhook, a new email, a schedule or an event from another application. Action nodes then read, transform, route or write data. Expressions map values between nodes, while branches and loops control what runs next.

n8n can run in n8n Cloud or on infrastructure a team manages. Self-hosting is not an automatic security advantage: the operator becomes responsible for updates, backups, access control, network configuration, secrets and monitoring.

Is n8n an AI tool or an automation tool?

n8n is primarily an automation and orchestration platform. AI is one capability inside it. A workflow can call a language model directly, use an AI Agent node with tools and memory, connect to a retriever or vector store, or avoid AI entirely.

This distinction matters because most business processes contain both predictable and uncertain work. A model may interpret a customer’s message, but an explicit rule should check required fields. A model may draft a response, but a normal application node should store the approved text. The workflow controls the sequence, data and permissions around the model.

Task	Best default	Why	Example
Check whether an email field exists	Deterministic node	The condition is exact	Stop if `email` is empty
Classify a free-text enquiry	AI with a fixed output schema	Language varies	Return `sales`, `support` or `other`
Calculate tax or a threshold	Deterministic rule	The calculation must be repeatable	Apply the configured rate
Extract fields from an irregular document	AI plus validation	Layout and wording vary	Return named JSON fields
Send, publish, delete or spend	Deterministic action after approval	The side effect needs control	Execute only the approved arguments

How does an n8n AI workflow work?

A useful AI workflow surrounds one model step with ordinary automation. The example below receives a sales enquiry. It validates the data, asks a model for a bounded classification and draft, applies business rules, waits for approval and then updates the CRM. Every run also creates an execution record and has a separate error route.

Figure 1. An annotated n8n AI automation

1. Trigger · Webhook receives enquiry and request ID.
2. Validate · Check consent, email and required fields.
3. AI step · Classify intent and draft structured output.
4. Rules · Route by allowed labels and confidence policy.
5. Approval · Person reviews recipient, claims and final text.
6. Act · Update CRM and send the approved reply.
7. Log · Store request, decision, actor and result.
Error path · Alert an owner; do not hide partial state.

The AI step interprets variable language. Validation, policy, approval, system updates and recovery remain explicit workflow controls.

Text equivalent: an enquiry enters through a webhook and receives a unique request ID. The workflow checks required data before an AI step returns a limited classification and draft. Rules reject unknown labels, and a person approves external communication. Only then do business nodes update the CRM and send the reply. A log records the outcome; errors go to a named review queue.

4-week live bootcamp · Barcelona or online

Go from one workflow to a complete AI automation system

In the AI Agent & Automation Bootcamp you build a no-code Virtual AI Department with n8n, RAG, connected agents and human approvals — live with an instructor and a small group.

See the bootcamp programme →

Not sure it fits? Take the free AI Career Roadmap test

What can businesses automate with n8n and AI?

Good use cases have a clear trigger, accessible data, a recognisable output and a bounded consequence. AI adds value when information arrives as natural language or documents rather than clean fields.

Lead intake: extract contact details, classify the request and prepare a CRM record for review.
Customer support: retrieve approved policy information and draft a cited answer before a person sends it.
Document processing: extract invoice or form fields, validate them and route exceptions.
Internal reporting: collect metrics from several systems, explain changes and create a draft report.
Monitoring: run scheduled checks, apply thresholds and ask AI to summarise the evidence behind an alert.
Content operations: turn an approved brief into a draft, pass it through fact and format checks, then wait for editorial approval.

Stable tasks such as moving a file, checking a number or copying a known field do not need an AI agent. Adding a model to a fixed rule increases cost and variability without improving the result.

How do you plan an n8n AI automation?

Define the business result. Name the trigger, required input, finished output, owner and measurable acceptance criteria. Do not start by choosing nodes.
Build deterministic steps first. Use explicit nodes for validation, routing, calculations and actions where the rule is known. Make the non-AI path visible.
Add AI only for judgement. Give the model one bounded task, approved context and a small output schema. Reject responses that do not match the schema.
Add permissions and recovery. Limit credentials, place approval before consequential actions, record execution state and route failures to a named owner.
Test and activate. Run normal, malformed, adversarial and tool-failure cases. Activate only when every case reaches its expected result or safe stop.

What permissions and error controls matter?

Credentials determine what connected nodes can do. Separate read and write access where the service permits it, avoid broad administrator credentials and share workflows only with people who need them. n8n’s documentation notes that workflow and credential behaviour depends on role, project and plan, so teams should verify their exact configuration.

Execution history helps operators inspect failed, running, successful and waiting runs. n8n can retry failed executions with the current or original workflow, but a retry is safe only when the workflow understands previous side effects. If an email or payment may already have happened, first check the destination using a unique operation ID.

n8n also provides a security audit for common risks involving credentials, database expressions, file-system nodes, risky or custom nodes and instance settings. An audit is one control, not a replacement for threat modelling, updates, access reviews or testing.

A successful node is not the business outcome. Verify the external record, message or file that the workflow was supposed to create.

How should you test an n8n AI workflow?

Test case	Expected path	Evidence
Valid request	Correct output and verified write	Execution data and destination record
Missing required field	Stop before model or write node	Validation message and zero side effects
Unexpected model label	Reject schema; human queue	Raw response and validation result
Instruction hidden in source text	Treat as data; keep workflow policy	Blocked content and chosen route
Approval rejected	No external action	Decision and execution state
API timeout after possible write	Check destination before retry	Request ID and lookup result
Duplicate webhook	One business outcome	Idempotency key and existing record
Error workflow unavailable	Primary workflow stops visibly	Platform alert and operator runbook

Criteria for success

The workflow completes one named business result from trigger to verified output.
Each AI step has a narrow task, approved context and validated schema.
Fixed rules and consequential actions do not depend on free-form model judgement.
Credentials grant only the access the workflow needs.
Every high-impact side effect has a human approval point.
Retries cannot duplicate an external action.
Operators can trace, stop and recover each failed execution.

Learn n8n as part of a complete business system

A free tutorial is enough to learn the canvas and build a small workflow. Barcelona Code School’s AI Agent & Automation Bootcamp fits people who want deeper practice with n8n, business data, AI agents, RAG, approvals, monitoring, error handling and multi-agent orchestration in one connected project.

The bootcamp is a career-focused, full-time programme, not the necessary next step for every n8n user. Check the live course page for current format, dates, tuition and prerequisites.

Explore the AI Agent & Automation Bootcamp

Frequently asked questions

What is n8n in simple terms?

n8n is a workflow automation platform. You connect triggers and action nodes on a visual canvas so data can move between apps, APIs and databases. It can also call AI models or run AI-agent steps inside a larger business process.

Is n8n an AI agent?

No. n8n is the workflow and orchestration platform. A workflow may contain an AI Agent node, a direct model call or no AI at all. n8n controls when steps run, how data moves and which business systems receive the result.

Do you need coding to use n8n?

Many n8n workflows can be built with visual nodes and expressions, so advanced programming is not required for a useful start. Code becomes helpful for custom data transformations, APIs, unusual authentication and production debugging.

Can n8n be self-hosted?

Yes. n8n offers Cloud and self-hosted options. Self-hosting gives a team more infrastructure control but also makes that team responsible for deployment, updates, backups, access, security settings and monitoring.

When should an n8n workflow use AI?

Use AI when a step must interpret variable language or content, such as classifying an enquiry, extracting fields from a document or drafting a reply. Use deterministic nodes for fixed rules, validation, calculations and consequential actions whenever possible.

Sources

Continue learning

How to build an AI agent

AI agent vs chatbot vs workflow

Compare BCS AI learning paths

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

AI Jobs in 2026: New Roles, Skills and Salary Examples

George K. — Wed, 14 Oct 2026 08:57:21 +0000

AI hiring is no longer limited to machine learning engineers. Companies are recruiting people to redesign workflows, drive adoption, train teams and build AI-native operating systems.

Published 14 September 2026 · Updated 14 September 2026 · Barcelona Code School

You do not have to be an AI engineer to build a career in AI. Companies now hire people to find useful AI applications, redesign workflows, connect tools, guide teams through adoption and measure the result. Your current profession is not wasted experience: it can become the domain knowledge that makes you useful in an AI role. The practical question is not “Can I start again?” but “Which part of my experience can I combine with working AI skills?”

Key takeaways

New titles include AI Enablement Lead, AI Adoption Lead, AI Transformation Lead, AI Automation Engineer and AI-native business roles.
People are entering from operations, project management, education, HR, marketing, product, consulting, finance, customer success, software and data.
You can start without a technical career, but you need proof that you can improve a real workflow with AI.
AI and big data are the fastest-growing skills in the World Economic Forum’s employer survey, while AI-skilled jobs carried a 62% average wage premium in PwC’s 2026 analysis.
The first step can be a one-week skills upgrade or a deeper career-change programme, depending on the role you want.

You are not late to the AI job market

Most people searching for “AI jobs” are not only looking for a list. They are asking a more personal question: is there a place for someone with my background? Current hiring gives a promising answer. The market is still forming its titles and boundaries, which creates room for people who can combine an established profession with practical AI building skills.

LinkedIn’s 2026 Jobs on the Rise analysis says both technical and strategic AI roles stand out among Europe’s fastest-growing occupations. In Spain, AI Engineer ranked first and Director or Head of AI ranked second in the national list captured in our career research. These rankings show direction, not the number of vacancies available to every candidate.

The wider demand signal is stronger than any one title. The World Economic Forum surveyed more than 1,000 employers representing over 14 million workers and placed AI and big data first among the fastest-growing skills through 2030. PwC’s 2026 AI Jobs Barometer, based on large-scale job-ad data, reports that jobs “professionalised” by AI are growing twice as fast as jobs “democratised” by AI, and that roles asking for AI skills carried an average 62% wage premium in 2025. A premium is an association within comparable jobs, not a guaranteed raise for taking a course.

Why this is a useful moment to move

Signal	Finding	What it means for a career changer
LinkedIn Jobs on the Rise 2026	Technical and strategic AI roles stand out across Europe; AI Engineer and Head of AI lead the Spanish ranking used in the BCS career study	Employers are still defining new role families, not hiring only one standard profile
World Economic Forum, 2025–2030 outlook	AI and big data rank as the fastest-growing skills; 59 in every 100 workers are projected to need training by 2030	Reskilling is becoming normal across professions
PwC 2026 AI Jobs Barometer	AI-professionalised jobs are growing twice as fast; average AI-skill wage premium reached 62% in 2025	Practical AI capability is gaining value, but pay still depends on role and market

These sources show a growing skills shift and emerging roles. They do not prove that competition is low in every location or guarantee a job outcome.

What new AI jobs are companies hiring for?

The clearest change is the expansion of AI work outside model research. Searches for AI automation jobs now lead to roles in technology, strategy, operations, learning, product and a founder’s office. Titles are still unstable, so the job description matters more than the label.

Role family	What the person owns	Typical technical depth	Common prior backgrounds
AI / ML Engineer	Models, data pipelines, evaluation, deployment and production reliability	High	Software engineering, data science, ML
AI Automation Engineer	Connected business workflows, agents, APIs, data and monitoring	Medium to high	Software, data, automation, technical operations
AI Enablement / Adoption Lead	Use cases, training, champions, workflow rollout and adoption metrics	Medium	L&D, change, operations, project management, communications
AI Transformation Lead	Operating model, enterprise workflows, executive alignment and ROI	Medium to high	Consulting, transformation, product, operations, programme leadership
AI-native business generalist	GTM, finance and operations systems built with AI	Medium	Founder’s office, BizOps, growth, product, consulting

Real vacancy evidence

The cards below reproduce the decision-relevant text from live job pages as accessible HTML. Each card links to the source and names the check date. This keeps the evidence readable even when a job board image is inaccessible, and avoids presenting a closed listing as current.

LinkedIn Jobs · checked 14 Sep 2026

AI Enablement Lead

Siemens Energy · Budapest

Full-time · 5+ years · Large multinational

Own a 500+ member community, executive AI upskilling, learning paths, responsible AI briefings and scoped AI projects.

Backgrounds named: L&D, internal communications, change management or digital enablement.

Open the LinkedIn vacancy

Evidence extract, not a salary claim: this listing did not publish compensation.

LinkedIn Jobs · checked 14 Sep 2026

AI Enablement Lead

waterdrop · Vienna

AI roadmap · Training · Automation

Evaluate AI tools, prioritise initiatives, train teams and build workflows using platforms such as Make, Airtable and internal APIs.

Core requirement: hands-on delivery across departments, not tool awareness alone.

Open the LinkedIn vacancy

The role combines project management, change and technical building.

LinkedIn Jobs · checked 14 Sep 2026

AI Enablement Lead

UK retail-tech scale-up · recruiter listing

£95,000–£98,500 base/year

6-month FTC · UK · Gemini rollout

The employer had already created an AI policy, steering committee and live rollout. The hire was expected to turn this foundation into adoption.

Open the LinkedIn vacancy

One contract vacancy, not a UK salary average.

Official careers page · checked 14 Sep 2026

Founder’s Associate

cloro · Remote EU

$50,000–$80,000 base + performance bonus

AI agents · SQL & APIs · P&L

Build AI-native systems across marketing, finance, operations and product. The page asks candidates to build agents, read a P&L and work with SQL and APIs.

Open the employer vacancy

Employer-published USD range for an EU-remote role; tax and employment terms require confirmation.

How much do these AI jobs pay?

There is no single “AI salary”. Scope, seniority, location and employment structure change the number. The honest way to use job-board evidence is to preserve each currency and condition.

Vacancy example	Published pay	Location / format	What the range proves
AI Enablement Lead, recruiter listing	£95,000–£98,500/year	UK; six-month FTC	A senior enablement contract can reach a high UK base range
Founder’s Associate, cloro	$50,000–$80,000 base; ARR bonus up to 100%	Remote EU	AI-native business systems are explicitly paid work, but the variable component is material
Founders Associate, Kuro	€54,000–€75,000 + 0.05–0.15%	Berlin onsite	European generalist AI roles may sit below senior transformation pay
Founder’s Associate, Nango	$140,000–$180,000 + equity	San Francisco onsite	Technical depth and US geography can change the range sharply
Founder’s Associate, Quadrillion	$150,000–$225,000 + equity	New York onsite	High US compensation exists, but this role is not a remote-European benchmark

The figures are useful because they show that companies already attach real budgets to this work. They are not a universal salary chart. A UK contract, an EU-remote startup role and a San Francisco position come with different taxes, benefits, living costs and work-authorisation rules, so use each figure as an example from its own market.

What skills keep appearing?

Skill	What employers mean	Portfolio evidence
Workflow analysis	Find a costly manual process and define the real bottleneck	Current-state map, exceptions, owner and baseline metric
AI and agent fluency	Use models with clear context, tools, boundaries and failure handling	Working agent with logs and evaluation cases
Automation and APIs	Connect business systems instead of copying outputs by hand	n8n or code workflow with structured inputs and outputs
Data and business judgement	Read metrics, P&L or operational data and choose priorities	Decision memo tied to cost, time, quality or revenue
Change and enablement	Train by role, support champions and get regular usage	Rollout plan plus adoption and outcome measures
Communication	Explain technical trade-offs to executives and users	Short brief, demo and documented decisions
Reliability and governance	Permissions, approvals, privacy, monitoring and safe stopping	Threat cases, human gates, retry rules and incident path

Which AI career move fits your background?

Do not erase your previous career. Use it as the first half of a new professional combination. The best transition is usually one step sideways into an AI version of work you already understand.

If you work in operations, administration or project management

You already understand handoffs, delays, exceptions, deadlines and KPIs.

→

Move towards AI Automation Specialist or AI Adoption Manager

Learn process mapping, n8n, APIs, structured data and safe human approvals. Build a portfolio case that removes one repetitive operational bottleneck.

If you work in teaching, L&D, HR or internal communications

You know how people learn, why behaviour does not change and how to support different audiences.

→

Move towards AI Enablement Lead or AI Learning Programme Manager

Add working prototypes, role-based AI use cases, governance basics and adoption metrics. Show that people changed a workflow, not only attended training.

If you work in marketing, sales or customer success

You understand customers, messaging, funnels, CRM data and commercial outcomes.

→

Move towards AI GTM, AI Product Marketing or Revenue Automation

Build research, lead qualification, reporting or customer-success workflows. Learn how agents connect to CRM and knowledge sources without making unauthorised promises.

If you work in consulting, product or business analysis

You can diagnose problems, interview stakeholders and turn ambiguity into a plan.

→

Move towards AI Transformation Consultant or AI Product Manager

Add enough technical depth to scope a solution, challenge assumptions, work with engineers and measure adoption, time-to-value and ROI.

If you work in finance, accounting or legal operations

You bring structured thinking, risk awareness, controls and knowledge of high-value workflows.

→

Move towards AI Operations, Governance or Domain Automation

Start with document intake, reconciliation, research or reporting. Keep financial and legal decisions with qualified people and make the evidence trail visible.

If you work in software, data or no-code automation

You already understand systems, logic, APIs or data models.

→

Move towards AI Automation Engineer or AI Agent Engineer

Add model behaviour, tool use, retrieval, evaluation, permissions, monitoring and business process discovery.

Can you enter AI from zero?

Yes, if “from zero” means you have not worked in technology before. You are not starting with nothing. A receptionist understands booking exceptions, a recruiter understands candidate workflows, a teacher understands learning, a nurse understands intake and escalation, and a shop manager understands stock, customers and daily operations. That knowledge tells you where an AI system may help and where it must stop.

The realistic entry point is rarely Head of AI. Start by becoming the person in your current field who can map a process, build a small working automation and explain its limits. One credible project can support a move into an AI-enabled version of your present role. Several deployed projects, stronger technical foundations and measurable results can support a later move into a dedicated AI automation or transformation position.

This is a good time to begin because employers are still adding new titles and rewriting skill requirements. It does not mean the market is empty or easy. It means the routes are not yet limited to candidates who have already held the exact same title for ten years.

Your first path from interest to credible AI work

Map one business workflow. Define the trigger, inputs, decisions, systems, exceptions, owner, output and baseline.
Work with structured data and APIs. Learn JSON, authentication, webhooks and how systems exchange records.
Build a deterministic automation. Connect a trigger to data checks and an output before adding an AI model.
Add AI where judgement is useful. Use it for classification, extraction, research or tool selection, with approved sources.
Define permissions and human approvals. Keep external messages, money, access changes and consequential decisions behind a person.
Test normal and failure cases. Include missing data, conflicts, tool outages, prompt injection, duplicate events and uncertain writes.
Deploy and measure. Track completion, escalation, time saved, error rate and adoption. A demo is not proof of operational value.
Package the evidence. Show the business problem, architecture, live workflow, test matrix, limitations and result in a portfolio case.

Choose the size of your first step

At this point, you do not need to decide your entire future. Decide what evidence you want to have next.

If your goal is to become noticeably stronger in your current profession, begin by building one useful agent for work. A marketer might create a research workflow, an HR specialist an onboarding assistant, a project manager a reporting agent, or a small-business owner a request-triage system. Barcelona Code School’s AI Agents Builder is the shorter route: one focused week to understand agents, connect them to everyday work and begin your career transition without leaving your present role.

If you want to apply for dedicated AI automation, implementation or agent-building roles, one assistant is not enough. You need broader evidence: connected systems, APIs, structured data, RAG, permissions, human approvals, failure recovery, evaluation and deployment. The AI Agent & Automation Bootcamp is designed for that deeper transition and for people who want a portfolio of end-to-end business workflows.

Which roles can the Bootcamp prepare you to pursue?

Your most realistic target depends on the experience you bring into the course. The Bootcamp adds applied AI automation skills; it does not erase normal requirements for industry knowledge, communication, project ownership or seniority.

Possible target role	What you would be expected to do	Who has the strongest starting fit
Junior AI Automation Specialist	Map routine processes, build workflows, connect tools and test results	Career changers with a solid portfolio of working automations
n8n Automation Developer	Build and maintain n8n workflows using APIs, webhooks, structured data and AI nodes	No-code builders, technical operations professionals and junior developers
AI Agent Builder	Create bounded agents that use company data and tools, request approval and handle failures	People who can demonstrate several tested agent projects
AI Implementation Specialist	Translate a team’s workflow into a deployed AI solution and support its adoption	Project managers, consultants, customer-success and operations professionals
AI Solutions Consultant	Run discovery, scope suitable automations, explain trade-offs and guide delivery	Consultants, business analysts, sales engineers and experienced client-facing professionals
AI Operations Specialist	Improve internal requests, reporting, knowledge access and operational handoffs	Operations, administration, finance operations and programme professionals
AI Enablement or Adoption Specialist	Build role-based use cases, train teams and measure whether workflows are actually used	L&D, HR, change, communications and team leads who add hands-on building evidence
Freelance AI Automation Consultant	Find a client problem, build a bounded solution, document it and provide ongoing support	People with client experience, commercial judgement and a clear industry niche

Senior titles such as AI Transformation Lead, Head of AI Enablement or AI Product Manager usually require more than the Bootcamp alone. They become credible targets when the new portfolio is combined with prior leadership, product, consulting, engineering or transformation experience and evidence from real organisational deployments.

Which route is right for you?

Your goal now	Recommended first step	What you should be able to show afterwards
Use AI better in my existing profession and test whether I enjoy building	AI Agents Builder	One useful agent connected to a real task in your field
Move towards AI automation, implementation or agent roles	AI Agent & Automation Bootcamp	Several end-to-end workflows with APIs, data, approvals, tests and deployment evidence

Neither route guarantees a job. The purpose is to turn your existing professional knowledge into work an employer or client can inspect.

Compare the two AI learning paths

Frequently asked questions

What AI jobs are appearing beyond AI engineer?

Current vacancies include AI Enablement Lead, AI Transformation Lead, AI Adoption Lead and AI-native operations or Founder’s Office roles. They combine workflow design, adoption, business analysis and automation with different levels of technical depth.

Do all AI jobs require a computer science degree?

No. Some engineering roles do, but enablement and transformation vacancies also recruit from learning and development, change management, operations, consulting, product and communications. Candidates still need practical AI fluency and evidence that they can improve a real workflow.

How much do new AI roles pay?

Pay varies sharply by role, geography, seniority and contract. Examples in this article range from €54,000–€75,000 for a Berlin Founder’s Associate vacancy to £95,000–£98,500 for a six-month UK AI Enablement Lead contract and higher US ranges. These are individual job advertisements, not market averages.

What should I learn to move into applied AI work?

Learn process mapping, structured data, APIs, workflow automation, model instructions, retrieval, permissions, human approvals, testing and monitoring. Build portfolio evidence that shows a manual workflow before and after your system.

Can a bootcamp guarantee an AI job?

No. A bootcamp can provide structure, feedback and portfolio projects, but hiring also depends on prior experience, geography, work authorisation, role fit and the quality of your evidence.

Sources

LinkedIn Jobs on the Rise 2026: Europe — ranking method and growth of technical and strategic AI roles.
World Economic Forum: Future of Jobs Report 2025 — employer survey, fastest-growing skills and reskilling outlook.
PwC 2026 Global AI Jobs Barometer — AI-related job growth, task change and wage-premium analysis.
Siemens Energy — AI Enablement Lead, LinkedIn, checked 14 September 2026.
waterdrop — AI Enablement Lead, LinkedIn, checked 14 September 2026.
Akaina Talent — AI Enablement Lead, LinkedIn, checked 14 September 2026.
Raspberry AI — AI Transformation Lead, Ashby, checked 14 September 2026.
Praecipio — AI Transformation Lead, Ashby, checked 14 September 2026.
cloro — Founder’s Associate, official careers page, checked 14 September 2026.
Barcelona Code School — AI Agents Builder.
Barcelona Code School — AI Agent & Automation Bootcamp.

Continue learning

How to build an AI agent

Compare Barcelona Code School AI courses

See the AI Agent & Automation Bootcamp

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

AI Agent Examples: 10 Practical Business Use Cases

George K. — Mon, 12 Oct 2026 08:04:34 +0000

See what each agent reads, decides, does and hands back to a person. Then use the selection checklist to choose a sensible first workflow.

The most useful AI agent examples are bounded business workflows, not general-purpose digital employees. A practical agent receives a defined request, reads approved sources, chooses limited tools, produces a traceable result and stops or asks for help when risk or uncertainty exceeds its permissions. Strong starting points include lead intake, support triage, document processing, research and report preparation.

Key takeaways

Choose a complete workflow with a clear owner, input, output and success measure.
Use deterministic automation for fixed rules; use an agent only where interpretation or tool choice adds value.
Keep consequential actions behind explicit approval and record every tool call.
Start read-only, test failure cases, then add narrow write permissions.
A smaller reliable agent is more useful than an impressive demo with unclear boundaries.

What makes an example a real AI agent?

An AI agent is not simply a chatbot that writes an answer. It uses a model to direct parts of a workflow and chooses tools that retrieve information or take action. OpenAI describes the basic components as a model, tools and instructions. The system also needs permissions, state, checks and a clear finish condition when used for real work.

The important test is ownership of a task. “Summarise this email” is a model request. “Read a support email, identify the account, check the order, draft a policy-grounded answer, request approval when a refund is involved, update the ticket and stop” is an agent workflow.

Figure 1. Anatomy of a bounded agent use case

1. Request · A defined goal enters through a form, inbox, schedule or person.
2. Context · The agent reads approved documents and system records.
3. Tools · It selects from a small set of read or write actions.
4. Approval · Risky or exceptional actions wait for a person.
5. Result · The output and action history are stored and measured.

A useful example names every stage. If the input, source of truth, action boundary or output is missing, the proposal is not ready to build.

Text equivalent: a request triggers the agent; the agent reads trusted context; it selects approved tools; sensitive actions pause for human approval; the final result and tool history are logged. Missing data, tool errors and policy exceptions follow an escalation path instead of being hidden.

Which AI agent use cases are practical?

1. Inbound lead qualification

Input: form, email or chat enquiry.

Agent work: extract needs, check the qualification rubric, ask for missing details and prepare a CRM entry.

Tools: product knowledge, CRM search and CRM draft/update.

Approval: a person approves pricing, unusual promises and outbound messages until the workflow is proven.

Output: evidence-backed qualification, owner and next step. Reject prompt instructions found inside copied lead content.

2. Customer support triage

Input: new ticket plus account identifier.

Agent work: classify the issue, retrieve policy and order context, draft an answer and route exceptions.

Tools: help centre, order system and ticketing platform.

Approval: refunds, credits, account restrictions and policy exceptions.

Output: grounded reply, category, priority and audit trail. On missing records, the agent asks for help rather than guessing.

3. IT request reviewer

Input: employee request for software or access.

Agent work: compare the request with approved tools, security requirements and role policy; collect missing evidence.

Tools: software catalogue, identity directory and ticket system.

Approval: every permission change and unapproved application.

Output: documented recommendation or routed approval request. The agent never grants itself broader access.

4. Meeting preparation agent

Input: calendar event and attendee list.

Agent work: find recent approved records, compile open questions and draft a briefing.

Tools: calendar, CRM, project tracker and document search.

Approval: none for a private draft; required before sending externally.

Output: a cited briefing with unresolved conflicts flagged. This is often a safe first agent because the primary output is reversible.

5. Document intake agent

Input: uploaded application, invoice or client packet.

Agent work: identify document type, extract defined fields, validate format and list missing items.

Tools: file store, extraction service and case database.

Approval: low-confidence extraction and any legal, financial or eligibility decision.

Output: structured record, source references and exception queue. Retention and access follow the organisation's privacy rules.

6. Research and evidence agent

Input: research question, source criteria and deadline.

Agent work: search approved sources, extract claims, compare evidence and produce a cited draft.

Tools: web search, internal search and document editor.

Approval: a subject expert reviews high-stakes claims and publication.

Output: claim ledger, sources, uncertainty and report. Conflicting evidence remains visible.

7. Metrics reporting agent

Input: schedule and agreed metric definitions.

Agent work: retrieve data, validate completeness, calculate predefined measures and explain notable changes.

Tools: analytics, spreadsheet or warehouse, plus reporting destination.

Approval: changes to metric definitions and any external distribution.

Output: dated report with data period, formula notes and missing-data warnings. The model should not invent causes from correlation.

8. Feedback-to-product agent

Input: support tickets, surveys and approved community feedback.

Agent work: group related signals, extract examples, count evidence and draft candidate product issues.

Tools: feedback sources and issue tracker.

Approval: product staff decide priority and roadmap.

Output: themes linked to source evidence, not an autonomous product decision.

9. Content production agent

Input: approved brief, claims and brand rules.

Agent work: research, draft, check claims, adapt formats and prepare a review package.

Tools: source search, content repository and QA checks.

Approval: publication, claims, regulated topics and brand-sensitive messages.

Output: reviewable drafts with sources and change log. It must not fabricate customer stories or performance numbers.

10. Operations exception agent

Input: failed workflow, overdue task or service-level alert.

Agent work: inspect logs, identify the known failure class, gather context and suggest the permitted recovery step.

Tools: workflow logs, runbooks and task system.

Approval: destructive retries, customer-impacting actions and policy overrides.

Output: recovered run or a concise escalation packet. It stops after the retry limit instead of looping indefinitely.

How do these examples compare?

Use case	Why AI helps	Risk	Good first version	Primary measure
Lead qualification	Interprets varied enquiries	Medium	Draft and route only	Correct routing rate
Support triage	Combines language and account context	Medium	Draft replies; no refunds	Resolved or correctly escalated
IT requests	Reads policies and missing evidence	High	Recommendation only	Policy-compliant routing
Meeting preparation	Searches scattered context	Low	Private briefing	Source coverage
Document intake	Extracts from varied files	Medium–high	Extract and flag	Field accuracy and exception recall
Research	Searches and synthesises evidence	Varies	Cited draft	Claim support rate
Metrics report	Explains structured results	Medium	Internal draft	Data and formula accuracy
Feedback analysis	Groups unstructured signals	Low	Evidence-linked themes	Traceable source coverage
Content production	Coordinates research and formats	Medium	Draft package	Factual and editorial pass rate
Operations exceptions	Interprets logs and runbooks	High	Diagnose and suggest	Safe recovery or escalation

How do you choose your first AI agent?

Name one result. “Handle support” is too broad. “Classify new tickets and draft policy-grounded replies” is testable.
Map the current process. Record the trigger, inputs, systems, decisions, exceptions, owner and final output before choosing tools.
Check whether AI is necessary. If stable rules can solve every case, build a normal workflow. Add a model where language, documents or changing context require interpretation.
Start with reversible actions. Reading, drafting, categorising and preparing records are safer than sending, deleting, paying or changing permissions.
Define source authority. State which database, policy or document wins when sources conflict, and how freshness is checked.
Set approval and stopping rules. List the actions that need a person, the retry limit and the conditions that end the run.
Create the test set first. Include ordinary cases, ambiguity, missing data, conflicts, tool outages, duplicate requests and prompt injection.
Measure the workflow outcome. Track correct completion and escalation, not the fluency of the agent's prose.

When should you not build an agent?

Do not use an agent merely because the label sounds modern. Anthropic recommends adding agentic complexity only when simpler solutions fall short. A fixed workflow is easier to predict, test and operate when the steps and rules are stable.

Keep these decisions with people

An agent may collect evidence or prepare a recommendation, but it should not autonomously diagnose a patient, give personalised legal advice, make a hiring decision, approve credit, sign a contract, transfer money or change critical access. These actions affect rights, health, employment, security or major financial outcomes. Use qualified human review and the policies that apply to your organisation and jurisdiction.

Also reject a use case when the data source is unreliable, no one owns the outcome, success cannot be measured, permissions cannot be limited or the process itself is not understood. Automation will reproduce unclear operations at greater speed.

How should an agent fail?

A safe failure is visible and bounded. The agent records what it attempted, preserves the last verified state, stops after a defined retry limit and gives a person enough context to continue. It does not quietly replace missing facts, ignore a denied permission or claim that an action succeeded without a tool result.

Failure	Agent response	Human receives
Required field missing	Ask one focused question	Case only if the user cannot answer
Sources conflict	Show the conflict; do not choose silently	Both sources and timestamps
Tool permission denied	Stop the write action	Requested action and denied scope
Tool timeout or uncertain write	Check state before retrying	Attempt ID and current state
Prompt injection in content	Treat it as untrusted data	Flagged content and intended task
Approval rejected	Keep the action unexecuted	Draft, evidence and rejection reason
Retry limit reached	Stop the run	Logs, errors and recommended next check

How do you test an agent example before deployment?

Turn each use-case card into an evaluation set. Test correct tool selection, grounded data, permissions, state changes and the final business result. Include failures deliberately. An agent that works only on the demonstration request is not ready for a real workflow.

Criteria for success

The agent completes the defined normal case using only approved sources and tools.
Every factual field can be traced to a source or tool result.
No sensitive action occurs without the required approval.
Duplicate input does not create a duplicate external action.
Missing data and conflicts produce a useful question or escalation.
Tool failures preserve state and stop after the configured limit.
Prompt injection inside documents, emails or web pages cannot change the agent's permissions.
The owner can review logs and reproduce why the agent acted.

Learn to turn an example into a working system

Barcelona Code School's AI Agent & Automation Bootcamp focuses on connected business workflows: data, tool-using agents, approvals, RAG, QA, reliability and deployment. It fits people who want to build complete automations with n8n and test them against real failure cases.

If you only need one fixed form-to-spreadsheet automation, a short tutorial may be enough. Use the live course page for the current format, dates and tuition.

Explore the AI Agent & Automation Bootcamp

Frequently asked questions

What is a practical example of an AI agent?

A practical AI agent can qualify an inbound lead by reading a form, checking approved company criteria, gathering missing details, preparing a CRM record and routing the lead to the right person. A human should approve consequential outreach or pricing decisions.

What is the best first AI agent for a small business?

The best first agent owns one frequent, bounded and reversible workflow with clear source data. Good candidates include request triage, meeting preparation, document intake or draft reporting. Avoid starting with payments, hiring decisions, medical advice or unrestricted outbound messages.

When should a business use automation instead of an AI agent?

Use deterministic automation when inputs are structured, rules are stable and every path can be defined in advance. Add an AI agent when the work requires interpreting unstructured information, choosing among tools or adapting the next step to context.

Which AI agent actions should require human approval?

Require approval for external messages, payments, refunds, contracts, access changes, destructive edits and decisions that affect employment, health, legal rights or significant financial outcomes. The exact boundary should follow the organisation's policy and risk level.

How do you test an AI agent use case?

Create a test set with normal requests, missing data, conflicting instructions, tool failures, permission failures, prompt injection attempts and repeated requests. Check task success, factual grounding, correct tool use, approval compliance, recovery and safe stopping.

Sources

OpenAI: A practical guide to building agents — agent definition, use-case criteria, tools, instructions, guardrails and human intervention.
OpenAI: Solutions for agentic workflows — first-party examples across lead qualification, IT, content and feedback.
Anthropic: Building Effective Agents — architecture selection and the case for starting simply.
Anthropic: Trustworthy agents in practice — human control, security, transparency and privacy risks.
Barcelona Code School: How to Build an AI Agent — the canonical step-by-step build guide.
Barcelona Code School: AI Agent & Automation Bootcamp — current course information.

Continue learning

Build an AI agent for a real business process

Compare Barcelona Code School's AI learning paths

See the AI Agent & Automation Bootcamp

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

AI Agent vs Chatbot vs Workflow: How to Choose

George K. — Sun, 11 Oct 2026 09:46:05 +0000

A chatbot manages a conversation. A workflow executes a defined process. An AI agent decides how to move a task forward. The right choice depends on where uncertainty enters the work.

Published 8 September 2026 · Barcelona Code School

An AI chatbot is built around conversation, an automated workflow is built around predefined steps, and an AI agent is built around delegated decisions. Use a chatbot when people need a natural-language interface, a workflow when the path is stable, and an agent when the system must interpret changing context, choose tools or recover from unexpected results. Many useful systems combine all three.

Key takeaways

A chatbot describes the interface; it does not prove that the system can plan or take action.
A workflow can include AI while keeping every important step and branch under software control.
An AI agent lets a model control part of execution, so it needs tighter permissions, evaluation and stopping rules.
A hybrid system often works best: deterministic rails for predictable steps, an agent for bounded judgement and a chatbot for human interaction.
Choose the simplest architecture that meets the acceptance criteria. More autonomy is not automatically more useful.

What is the difference between a chatbot, a workflow and an AI agent?

The three terms describe different parts of a system. A chatbot describes how a person communicates with software. A workflow describes how work moves through a sequence. An agent describes who or what chooses the next action.

Google Cloud defines AI chatbots as applications or interfaces that hold natural-language conversations and generate responses to text or voice input. A chatbot may answer questions without changing any external system. It can also sit in front of a workflow or agent.

A workflow is a sequence that moves from a trigger to an outcome. In tools such as n8n, a workflow is represented by connected nodes. Each node performs a defined operation, and the connections determine how data moves. An AI model may classify text or write a summary inside the sequence without controlling the whole sequence.

OpenAI's practical guide distinguishes agents from LLM features that do not control workflow execution. In an agent, a model manages at least part of the work: it can select a tool, assess a result, decide whether the goal is complete and hand control back when it cannot proceed safely.

Criterion	Chatbot	Automated workflow	AI agent
Primary job	Exchange information through conversation	Execute a known sequence	Move a task toward a goal
Who selects the next step?	User, conversation design or application	Rules and branches written in advance	Model within instructions and guardrails
Best input	Questions and conversational requests	Structured events and predictable data	Ambiguous requests or unstructured information
Tool use	Optional	Fixed integrations	Dynamic selection from approved tools
Predictability	Varies with response generation	Highest when rules are explicit	Lower; requires evaluation and limits
Main risk	Incorrect or misleading answer	Bad rule, broken integration or unhandled branch	Wrong decision followed by a real action
Typical example	Answer delivery questions	Copy a paid order into fulfilment	Investigate an exception and propose a resolution

Is a chatbot an AI agent?

A chatbot is not automatically an AI agent. Conversation is an interface, while agency is a control pattern. If a system receives a question and returns an answer, the system can be useful without making a plan, choosing a tool or carrying a task across several steps.

A chatbot becomes the front end of an agent when the system can act beyond the conversation. For example, a support chatbot may collect an order number. A workflow validates its format and retrieves the order. An agent then reads an unusual complaint, checks approved policies, chooses which evidence to gather and proposes a resolution. A person approves a refund before money moves.

Interface is not autonomy. Ask who controls the next step and who can change an external system. The answer reveals whether the architecture is a chat interface, a fixed workflow or an agent.

When is an automated workflow the better choice?

Use a deterministic workflow when the business already knows the correct path. A form submission can be validated, stored and routed by explicit rules. An invoice can be generated from approved fields. A notification can be sent when a threshold is crossed. These processes benefit from consistency more than judgement.

OpenAI recommends checking whether a deterministic solution is sufficient before building an agent. Agents are more appropriate when the work requires contextual decisions, rules have become difficult to maintain or the input is largely unstructured. Adding an agent to a stable process increases the testing surface, latency and possible failure paths without necessarily improving the result.

A workflow is usually enough when:

the trigger, inputs and expected output are known;
important branches can be written as explicit conditions;
the same conditions should lead to the same action;
failure must be easy to reproduce and audit;
an LLM is only needed for a bounded step such as extraction, classification or drafting.

When does an AI agent earn its complexity?

An agent earns its complexity when the correct next step depends on information discovered during execution. The model may need to interpret a request, select one of several data sources, compare conflicting evidence, retry a failed approach or decide that human judgement is required.

The agent should still operate inside a designed process. Tools define what it can access. Instructions define its role. Guardrails restrict risky actions. Logs preserve evidence. Exit conditions stop the run when the goal is reached, the attempt limit is exceeded or the system cannot continue safely.

An agent may be justified when:

inputs arrive as emails, documents, images or free-form requests;
exceptions are common and hard to express as a manageable rule tree;
the choice of tool depends on the meaning of the case;
the system must adapt after a tool returns missing or unexpected information;
success can be evaluated and unsafe actions can be blocked or approved.

Figure 1. Choose by where the uncertainty lives

1. Conversation: Do people mainly need to ask questions or provide information in natural language? Choose a chatbot interface.
2. Sequence: Can the important steps and branches be specified before the process starts? Choose a deterministic workflow.
3. Decisions: Must the system choose tools or next steps from changing, unstructured context? Add a bounded AI agent.

Start with the user interface, keep predictable work deterministic, and introduce agent control only where contextual judgement is required.

The diagram's text equivalent is simple: use a chatbot for natural-language interaction; use a workflow when the sequence can be predefined; use an agent only for the part that must select actions from changing context. A system may move through all three layers without giving the agent control over every step.

How can the three approaches work together?

A hybrid design separates interaction, execution and judgement. Consider a supplier-onboarding process. A chatbot answers common questions and collects missing details. A workflow validates tax fields, creates folders and routes documents. An agent reads non-standard documents, identifies inconsistencies and prepares a risk summary. A procurement manager approves the supplier before the workflow updates the system of record.

This architecture keeps routine operations visible and testable. The agent receives only the tools and data needed for the judgement step. A wrong summary cannot create a supplier because the approval and final write remain separate controls.

How do you choose the architecture step by step?

Define the outcome. Name the business result, the owner and the evidence that proves completion.
Map the current process. Separate stable steps from decisions, exceptions and human judgement.
Choose the interface. Add chat only when conversation improves data collection or access to the result.
Automate the fixed path. Keep validation, permissions and predictable integrations deterministic.
Bound the agentic step. Give the model the smallest useful set of tools, explicit instructions and clear exit conditions.
Protect consequential actions. Require human approval before external communication, deletion, financial changes or other hard-to-reverse actions.
Evaluate the complete process. Test normal cases, missing data, conflicting inputs, tool failures, unsafe requests and escalation.

Criteria for success

The system completes the defined outcome, not merely a convincing response.
Every external action has an owner, permission and audit trail.
The agent stops or escalates when evidence is missing or risk exceeds its authority.
Deterministic steps produce repeatable results under the same conditions.
The hybrid system performs better on agreed tests than the simpler baseline it replaces.

Learn to design agents and workflows as one system

Barcelona Code School's current AI Agent & Automation Bootcamp is a four-week, full-time, no-code programme. The live curriculum uses n8n and covers connected workflows, tool-using agents, business data, human approvals, RAG, QA and production readiness.

The programme fits people who want to build complete business automation systems. A short chatbot tutorial is the better choice if your only goal is to create a conversational demo.

See the current bootcamp curriculum

The live course page is the source for current dates, format, tools and tuition.

Frequently asked questions

Is a chatbot the same as an AI agent?

No. A chatbot is primarily a conversational interface. An AI agent manages work toward a goal, selects tools or actions, observes results and decides what to do next within defined boundaries. A chatbot can be the interface to an agent, but conversation alone does not make a system agentic.

What is the difference between an AI workflow and an AI agent?

An AI workflow follows a path designed in advance, even when one step uses a language model. An AI agent uses a model to control at least part of the execution, such as choosing the next step, selecting a tool or recovering from an unexpected result.

Can one system contain a chatbot, a workflow and an agent?

Yes. A customer can speak through a chatbot, a deterministic workflow can validate and route the request, and an agent can handle the part that requires interpreting unstructured information or choosing among several tools.

When is a fixed workflow better than an AI agent?

A fixed workflow is usually better when inputs are structured, the rules are stable, the sequence is known and the same conditions should produce the same action. It is easier to test, faster to run and simpler to audit.

Does an AI agent remove the need for human approval?

No. Human approval should remain before consequential or hard-to-reverse actions such as sending external messages, moving money, signing contracts, deleting data or publishing content.

Sources

OpenAI: A practical guide to building agents — agent definition, workflow control, tools, guardrails and architecture selection.
Google Cloud: AI chatbot — chatbot definition and conversational use.
n8n documentation: Workflows — workflow and connected-node model.
Barcelona Code School: AI Agent & Automation Bootcamp — current course format and curriculum.

Continue with Barcelona Code School

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

How an AI Agent and Automation Bootcamp Works

George K. — Sat, 10 Oct 2026 18:32:15 +0000

Format, prerequisites, the daily routine, the final project, group size, remote study and job support, answered by the team that runs the AI Agent & Automation Bootcamp at Barcelona Code School.

The AI Agent & Automation Bootcamp runs four weeks full-time, Monday to Friday, 9:00 to 18:00 CET, on campus in Barcelona or remote live in the same classroom.

Groups are capped at six students. Mornings are lectures and live demonstrations, afternoons are building. No prior coding is required, and every student leaves with finished agents and automations as portfolio proof.

Key takeaways

The programme is four weeks full-time, Monday to Friday, 9:00 to 18:00 CET; students attend on campus, remote live, or a mix of both.
Roughly 60 to 80 per cent of each day is spent building, not listening. Lectures come in the morning, implementation after lunch.
No coding background is needed. The course teaches business requirements, tool choice, system design, orchestration, deployment and maintenance.
The final project connects agents built for separate company departments into one enterprise system that exchanges data and coordinates work.
Groups are capped at six students so an instructor can see individual progress and work one-on-one with anyone who is falling behind.
Barcelona Code School gives job support, not a job guarantee: portfolio, mock interviews, a career advisor session and introductions to hiring partners.

Questions answered here

The bootcamp at a glance

The table below summarises the answers in this Q&A. Dates, tuition and the current syllabus live on the AI Agent & Automation Bootcamp page, which is the source of truth for anything that changes between cohorts.

Question	Answer
Duration	Four weeks, full-time
Schedule	Monday to Friday, 9:00 to 18:00 CET
Where	Barcelona campus, remote live, or a combination of both weeks
Group size	Six students maximum
Coding required to join	No. Coding helps but is not a prerequisite
Split of the day	Lecture and live demonstration in the morning, building after lunch; about 60 to 80 per cent build time
Final project	A professional dashboard with departmental agents connected into one enterprise system
Career support	Portfolio, mock interviews, career advisor session, hiring partner introductions
Job guarantee	No. Support is given, outcomes are not guaranteed

Figure 1. How one bootcamp day is structured

09:00 Lecture · New concepts, patterns, methods and approaches explained by the instructor.
Late morning Live build · The instructor implements the same thing in front of the group, step by step.
After lunch Your build · You design, implement, test, iterate, fix and deploy the system yourself.
Until 18:00 Feedback · Work is checked, questions answered, and struggling students get one-on-one time.

Figure 1. Daily sequence in the AI Agent & Automation Bootcamp: morning lecture, instructor demonstration, student build, feedback. Text equivalent: classes run 9:00 to 18:00 CET; the morning covers new material and a live demonstration, the afternoon is spent building, and work is checked the same day. Source: bootcamp explainer Q&A, questions 2 and 4.

What is this bootcamp and who is it for?

Hi! In this video we're going to answer some frequent questions about our AI Agent and Automation Bootcamp. This bootcamp is essentially for people who want to enter this growing area of the job market of AI agents building and AI automations. A lot, a lot of companies more and more are looking for someone who can make and operate and manage the system of AI agents for them to run certain processes in their companies and build those automations. So a lot of companies are thinking about AI transformation, becoming AI-first, AI-native, and so on.

This bootcamp is a preparation for these kinds of jobs. It's suitable both for people with some experience in IT or for people who are complete beginners in that and want to make a career switch from a completely different industry. This bootcamp is possible for both categories of students.

What is the duration and schedule of the bootcamp?

This bootcamp is a four-week, full-time course, which means that for one month, every day, Monday to Friday, from 9:00 in the morning till 6:00 in the evening, you will be learning and building AI agents and automation systems.

What will I learn, and is it a practical course?

It's not just about learning some tools, but it's about understanding the whole process, the whole job: how to translate business needs and business goals into a map of automation, how to implement it, how to run it, and how to deploy it.

You will leave not with just some abstract theoretical knowledge, but you will leave, first of all, with done projects which you can show to prove your skills. You will leave with an understanding of how all of that actually works in the company, and you will learn all these skills starting with understanding the business and going all the way to implementation, testing, deployment, maintaining, optimizing, and expanding the system.

So you will leave with very concrete skills which are right now in much demand in the job market due to this massive AI transformation wave which goes globally, affecting all types of companies. It's not specific to some industry; it's not only like software development. No, it's for everything, for every company. Every company feels that, "Okay, competitors are using AI and we don't use AI. We need to do it fast. Who can do it?" You!

What is the daily routine and how are the classes conducted?

Our classes are Monday to Friday. We start at 9:00, we finish at 6:00 p.m. In the morning, we give a lecture explaining some new materials, new concepts, new patterns, new methods, new approaches, new solutions, etc., etc. An instructor shows you how to implement it in practice. And after lunch, it's your time to build it.

So now you sit and actually make it. That's why it's not just listening or watching some video. No, it's actually listening, understanding, seeing people doing it in front of you, and then doing it by yourself. This is how you learn and this is how you become capable of delivering it at actual work.

Plus, you finish the course with a set of complete projects which are agents and automations. So you can always show and you can always prove your skills and your knowledge. All the bootcamps we do are practical and project-based because that's the only way to actually learn it - by trying and sometimes failing and then fixing it and then understanding how to do it properly next time.

Same here in this bootcamp: it's not just theory or some history of AI or whatnot. No, you spend at least half of a day building the actual products, understanding the business goal, making a map, thinking about what tools to pick, what would be the structure, what would be the system design of your agent, what it needs to know, what it needs to be connected to, and then you implement it, and then you test it, and then you iterate, and then you fix everything, and then you deploy it, and then you have it. So the course is at least 50% of practice, but in reality, it's a little bit more. It's closer to like 60% to 80% of the time you will be building, and the rest you will be listening to the lectures and explanations.

Do I need to know how to code before I join this bootcamp?

No, you don't need to know how to code before you join this bootcamp. If you know how to code, it will help you a lot. But if not, you don't need that. If you will need to build something by coding, you can always build it with AI, and we will explain how to do it properly, how to understand what AI will do, and how to control it.

The focus of this course will be on understanding the bigger picture and more important skills. You will be training to understand the business requirements and business goals. You will be training to understand how to choose the appropriate tools and stack of tools. You will be training on understanding things like system design and your data design, and how to connect to any possible data sources or external platforms - some APIs, or some file storage, or cloud storage, or whatnot. You will be able to understand how to create the system of agents, how to orchestrate them, how they will communicate with each other, and how to deploy them.

So all of that is not really coding. Coding is a very small part of this process, and if you will need to code something, you can always code it with AI, and under our supervision, we will explain how to do it properly and how to do it right and correctly.

What is the main goal or final project of the course?

The goal of this course is to build a professional dashboard with a set of connected agents. Initially in the course, you will be building separate agents - they will be disconnected. Then later in the course, you will build the agents for different departments of a company. So they will all operate on their own areas and they will all do their own job.

And then you will connect all of them into one enterprise system which will basically run the whole company to a certain degree. It will perform processes of different departments and also exchange data, communicate, and have some coordination between them so everything is running smoothly with the goal to achieve the business requirements which are given to this company.

Why is this bootcamp better than self-learning and online tutorials?

We get this question a lot through all the years we run our bootcamps. Usually, students who come in here and tried some self-learning and tutorials before, they all mention the same problem: a tutorial shows you how to make one specific thing. If you are following a tutorial step-by-step, you make the same project the tutorial is describing. But the problem starts when you want to build something different or something of your own, and then there is no tutorial for that.

So you have no idea which tools to pick, how to combine them, how they communicate, and all of that is just falling apart because by following a tutorial, you understand just how to make one exact thing.

But by doing the bootcamp, you will get an understanding not of how to make a specific thing, but you will get the understanding of all the available tools, possibilities, and functionality. You will learn how to make educated decisions about structuring your project, picking the tools, combining them together, what kind of logic to put in, and what kind of external tools to connect. You will know your tools and you will know how to pick them. This would be the outcome of the bootcamp.

With a tutorial, the outcome is that you will know how to build exactly the same thing as in the tutorial, but step to the left, step to the right, you are lost, because this is not how tutorials work.

Do you guarantee a job after graduation? What career support is provided?

None of any school can guarantee a job to any of the students because, first of all, it depends on us and also on you - how much effort you will put - plus it will also depend on the job market, the companies, how you will go for interviews, and so on. So we never gave a job guarantee, but we always gave job support, and this works very well, and there is a big difference.

After the bootcamp, the most important thing to understand is that you will be completely prepared to start a job in this industry, in this area. You will be completely prepared to start the job of AI agent developer, AI agent engineer, AI agent implementer specialist, or AI business automation expert - doesn't matter the words, as long as it's about building what we're teaching: building AI agents and building AI automations. You will be completely qualified to do this job.

You will have a portfolio with actual projects done during the bootcamp which will be proof of that. You can show them to recruiters, to the clients, to wherever you're going to apply. It will be your hard and solid proof of the skills you have and the work you did.

And we will train you to do the job interviews. We will do some mock interviews with you, and we will give you a career advisor session to optimize your LinkedIn, take a look at your portfolio, see maybe some skills are transferable, etc., etc. And then we will also connect you with hiring partners we have. We will send your profile to all the hiring partners we have, to all the new incoming requests. If you need any second opinion during the technical job interviews, you can also get in touch with us - we can also give you our opinion on that.

All of that, all this set of measures, is focused on helping you to find a job as soon as possible. But we cannot really guarantee that you will be hired in one week after the course - no, because it's not only up to us, of course. But the rest, yes, all the possible job support the school can provide to someone, we will do that for you, same as we have been doing it for the last 10 years for the graduates of our school.

Do you accept international students? Can I take the course remotely?

Yes! We accept international students. A lot of our students - more than half of them - are international students, either coming here in Barcelona for the course or doing it remotely.

In the case of this bootcamp, the groups are hybrid. You can be here on campus every day from 9:00 to 6:00, or you can do it remotely live by connecting to the live classroom at 9:00 in the morning CET time, and then we will be finishing at 6:00 in the evening. You will be joining the same classroom, the same group, same students, and same instructor.

All the lectures will be streamed live. You will be able to pose and ask questions same as if you are in the classroom, and all the work during the day will be checked in the same way as if you are here on campus.

Or you can also combine: you can take the course and maybe, let's say, spend one week on campus, then one week remotely, then again on campus, or spend half of the course here and half of the course at home. It's not really a problem.

Nowadays, a lot of companies work remotely. So not only will you not miss anything, but you will actually learn the workflow of how these teams are working. If they are remote teams, you will be already prepared and feeling comfortable with that if you do, let's say, the whole course or part of the course in remote mode.

And for international students, it's the same - no problem. There is no difference between registering for the course and taking it if you are in Spain, or if you are in another country, or if you're in another continent. Of course, there is a time difference you need to consider, but apart from that, it's all the same.

What is the group size and why is it so small?

The group is six people maximum. Why? Because we believe that a bootcamp is quite a serious investment and there are serious things at stake. This is your career. We cannot afford for it to be just a one-way style of communication where we have 20 students in the group and we don't really have time to talk with each one of them per day because we have to deliver materials and then they have to stand in line to ask a question. This all affects the outcomes and efficiency in a negative way.

Instead, what we discovered through the years is that six students per group is the optimal number where we have enough time to actually work with you individually if needed. We know the progress of each student. We know if someone is struggling with something or if someone is going faster, and you don't even need to tell it to us because we can see that. We can see it by the work you're doing, by the questions you're asking, and of course, we can see it when we check your work to give you feedback.

If we feel that someone is lagging behind a little bit, then we have a chance to actually sit with this student one-on-one and explain something which is misunderstood or show them better ways of doing something, and so on and so on. So we feel that this size of the group is very comfortable for the students, and we do not compromise the quality of delivering these courses if we have up to six students maximum.

How early should I register for the bootcamp?

What it also means is that if you're thinking of signing up for this bootcamp or any other bootcamps, you need to do it basically as soon as you decide so that you will reserve your place in the group you want. Otherwise, you might have to wait until the next cohort if the one you wanted is filled up.

Is this the right course for you?

Pick by what you want to run after the course, not by how much AI experience you have now.

You want to build and operate AI agents and automations for companies, as a job, a freelance service or an internal role. The four-week AI Agent & Automation Bootcamp is built for that.
You only want to automate your own daily work and do not need a company-wide system or a career change. The shorter Work Automation with AI Agents course is a better fit and a much smaller commitment.

See the AI Agent & Automation Bootcamp

Dates, tuition, payment terms and the full syllabus are on the course page.

Sources

Official AI Agent & Automation Bootcamp explainer video, Barcelona Code School. The answers on this page come from it.
AI Agent & Automation Bootcamp course page, Barcelona Code School. Format, schedule, group size, cohort dates and tuition.

Published 8 September 2026. Last substantive update 8 September 2026. Author: Barcelona Code School. The answers come from the school's own explainer video and were not rewritten.

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.

What Is Agentic AI? Agentic AI vs Generative AI

George K. — Fri, 09 Oct 2026 08:38:26 +0000

AI agents explained

Generative AI produces an answer. Agentic AI manages a goal: it chooses steps, uses tools, checks results and adapts within defined boundaries.

Published and reviewed 26 August 2026 · Barcelona Code School

Agentic AI is an AI system that can pursue a goal through a sequence of decisions and actions. Instead of producing one response and stopping, it can plan the next step, select an approved tool, observe what happened, update its state and continue until the task is complete or human judgement is required. Generative AI is usually one component inside that system: it supplies language understanding and generation, while the agentic layer manages the work.

Key takeaways

Generative AI creates content; agentic AI coordinates actions toward a goal.
An agent normally combines a model with instructions, tools, memory or state, an execution loop and guardrails.
Not every automation needs an agent. Fixed processes are often safer and cheaper as deterministic workflows.
Useful autonomy is bounded autonomy: permissions, validation, logs and human approval remain part of the system.
The important skill is not merely prompting a model, but designing and testing the complete process around it.

Generative AI creates. Agentic AI acts.

Generative AI refers to models that create new content from patterns learned during training. Given a prompt, a model can produce text, code, images, audio or another output. A familiar example is asking a language model to draft a reply to a customer.

Agentic AI starts with a different question: what must happen for the customer's problem to be resolved? The system may need to identify the customer, check an order, read a policy, decide whether a refund is allowed, update a record and prepare a response. If the amount exceeds a threshold, it may have to stop and request approval.

The model still generates language and makes bounded decisions, but generation is no longer the end of the process. It becomes one capability inside a system that can interact with external data and software.

Dimension	Generative AI	Agentic AI
Primary job	Create an output	Move a task toward a defined goal
Typical interaction	Prompt → response	Goal → plan → actions → observations → adjustment
Tool use	Optional	Usually central to reading data or taking action
State	Often limited to the current context	Tracks progress, intermediate results and task status
Control flow	Mainly directed by the user or application	The model can choose the next step within boundaries
Risk	An inaccurate output	An inaccurate output plus a wrong or unintended action
Example	Draft a follow-up email	Find stalled leads, choose who needs a follow-up, draft messages and request approval before sending

The boundary is architectural, not a competition between two model types. The same language model can power a simple generator, a fixed AI workflow or an agentic system.

How does an agentic AI system work?

An agent begins with a goal and a set of permitted capabilities. It decides which action is useful, calls a tool, receives information from the environment and uses that observation to decide what to do next.

The agent loop

1. Goal · Understand what outcome the user or process requires.
2. Plan · Choose the next useful step rather than scripting every path in advance.
3. Act · Use an approved tool to search, calculate, update or communicate.
4. Observe · Read the tool result and compare it with the expected state.
5. Adjust · Continue, retry, change approach, stop or ask a person for help.

An agent repeats the plan–act–observe cycle until it reaches an exit condition. Guardrails, permissions and human checkpoints limit what it may do.

The minimum useful architecture normally includes:

A goal and instructions: what the system is responsible for, what success means and what remains outside scope.
A model: the reasoning and language component that interprets context and selects among allowed actions.
Tools: controlled interfaces for retrieving data or changing an external system, such as reading documents, querying a CRM or creating a task.
State or memory: the facts and intermediate results required to continue a multi-step task.
Guardrails: validation, access controls, risk rules, iteration limits and conditions for stopping.
Evaluation and logs: evidence of what the agent attempted, what happened and whether the result met the acceptance criteria.

A practical example: handling an incoming sales lead

Generative AI version

A salesperson pastes a new enquiry into a chat and asks for a reply. The model drafts the text. The person still checks the CRM, verifies the details, creates the task and sends the message.

Agentic AI version

A form submission triggers the system. The agent validates the contact details, checks for an existing CRM record, reads the enquiry, retrieves the relevant service information and proposes the next step. It updates the record and creates a task. Before any external message is sent, a person reviews the draft and approves or rejects it.

Both versions use generative AI. Only the second one owns part of the workflow, interacts with business systems, keeps task state and responds to intermediate results.

Agentic AI is not the same as automation

A deterministic workflow follows paths defined in advance: when event A happens, perform steps B and C. It is usually the right choice when inputs are structured, rules are stable and exceptions are known.

An agent is useful when the correct path cannot be fully specified beforehand. It can interpret an ambiguous request, search different sources, select a tool and change its approach when the first attempt fails.

Use the simplest architecture that reliably solves the problem. Adding an agent increases flexibility, but it can also increase latency, cost and the number of ways the process may fail.

Prefer a deterministic workflow when:

the sequence is stable and every important branch can be written as a rule;
the same input should always produce the same action;
speed, cost and predictability matter more than flexibility;
a wrong decision would be difficult to detect or reverse.

Consider an agent when:

the task requires interpreting natural language or unstructured documents;
the next step depends on information discovered during execution;
several tools may be valid and the system must select among them;
the work benefits from planning, iteration or recovery after a failed attempt;
you can define clear permissions, stopping conditions and acceptance tests.

Does agentic AI mean fully autonomous AI?

No. Agency is a spectrum. One system may only choose which document to retrieve. Another may complete a long sequence of tool calls before it returns to the user. The amount of autonomy should follow the risk of the task, not the ambition of the demo.

Low-risk, reversible actions can often run automatically. Actions involving money, contracts, deletion, publication, sensitive data or communication with customers should use explicit permissions and human checkpoints. A production system also needs records of every significant action and a clear way to stop.

What agentic AI is good at—and where it still fails

Agentic systems are useful for processes that combine judgement with action: researching across sources, classifying and routing requests, preparing documents, updating records, monitoring changing information and coordinating work across tools.

They can also fail in more complex ways than a chatbot. An agent may misunderstand the goal, choose the wrong tool, carry an early error into later steps, loop unnecessarily, use stale information or take a valid action at the wrong time. For that reason, a convincing prototype is not the same as a reliable production system.

Before deployment, teams should test normal cases, missing information, conflicting sources, tool failures, permission boundaries and handoff to a human. The system should know not only how to proceed, but when not to proceed.

What should you learn to build agentic AI?

Prompt writing is only one part of the work. Building an agentic system requires process mapping, data design, tool integration, structured outputs, memory, retrieval, human approval, failure handling, evaluation and production monitoring.

A useful learning project should finish with more than a chatbot demonstration. It should include a process map, acceptance criteria, test cases, an action log, failure scenarios and a runbook explaining how the system is operated safely.

Learn to build AI agents around real business processes

Barcelona Code School's AI Agent Engineer programme teaches agentic systems through complete business workflows. The current internal curriculum covers workflow design, tool-using agents, structured business data, human approval, RAG, evaluation and production readiness.

It is intended for people who want to design and build working AI automation systems—not only use a chat interface or collect prompts.

Explore the AI Agent Engineer programme

See the live course page for the current syllabus, format, dates and tuition.

Frequently asked questions

What is agentic AI in simple terms?

Agentic AI is an AI system that can pursue a goal through multiple steps. It decides what to do next, uses approved tools, observes the result, adjusts its plan and stops or asks for human input when necessary.

What is the difference between agentic AI and generative AI?

Generative AI primarily creates content in response to an input. Agentic AI uses a generative model inside a larger system that manages a task, chooses actions, calls tools, maintains state and reacts to results.

Is ChatGPT agentic AI?

A basic prompt-and-response chat is generative AI, not an agent. A ChatGPT-based system can become agentic when it is given tools, state, a goal, an execution loop and rules for acting or asking for approval.

Does agentic AI work without humans?

It can complete bounded low-risk steps autonomously, but consequential actions should use permissions, guardrails, logs, stopping conditions and human approval.

When should a business use agentic AI?

Agentic AI is useful when a task is multi-step, varies from case to case, requires interpretation of unstructured information and benefits from choosing among tools. A deterministic workflow is usually better for stable, predictable processes with fixed rules.

Sources

OpenAI — A practical guide to building AI agents. Agent definition, tools, workflow control and guardrails.
Anthropic — Building effective agents. Distinction between workflows and agents, agent loops and architectural trade-offs.
Anthropic — Trustworthy agents in practice. Human control, permissions, transparency and security.
Anthropic — Demystifying evals for AI agents. Evaluation of multi-turn tool-using systems.
Barcelona Code School — internal AI Agent Engineer curriculum, reviewed 26 August 2026.

From generating answers to building systems that complete work

Learn how to map a business process, connect tools and data, add human approval, test failures and turn an AI model into a controlled agentic system.

See the AI Agent Engineer programme

Continue learning about AI systems

How to Build an AI Agent: A Simple Step-by-Step Guide

Major Curriculum Upgrade: Building AI-Powered Apps

Originally published at Barcelona Code School. Learn more about our AI, coding and design programmes at barcelonacodeschool.com.