AI agents have graduated from being an experimental technology. Companies are increasingly exploring agents that can qualify leads, fulfill support requests, take advantage of internal data, update code, carry out tasks and even make decisions under constraints. According to McKinsey's 2026 research, 40% of respondents from companies with more than $1 billion annual revenue said they had scaled an AI agent, compared with 27% the previous year.
But you need to think about this more than simply selecting a model and developing a chat interface for it.
Before launching into hiring a team or internally building an AI agent, there are some questions worth answering:
**
1. What Problem Are We Actually Trying to Solve?
**
This may sound like a no-brainer, but that's where a lot of AI projects go wrong. "Build an AI agent" is not a business objective. Cutting the time to qualify leads is.
Automating the more frequent customer-support workflows is.
Finding information for employees in your internal systems is. Responding to mundane requests without manual oversight is.
The more specific the problem, the more straightforward the answer becomes, and the more likely it is you'll be able to assess whether or not an agent truly has a place in your workflow. An agent should be doing a job, not just existing because the technology is there.
**
2. Does This Workflow Actually Need an AI Agent?
**
Not every automated workflow requires a bot, Not every workflow is a place where traditional automation is more straightforward, less costly, and easier to maintain - if your workflow is a simple, repetitive set of procedures, then that workflow is likely simpler to automate without an AI agent.
AI agents are most compelling when a workflow: Has inputs that evolve in some way, or loads of different routes it could go down; Is filled with unstructured data; Has decision points that require a little reasoning.
For instance, if your workflow involves processing a form with 10 fixed fields, you don't necessarily require an agent. But if a customer request involves understanding the message, verifying account info, referencing internal policies, and determining next actions, well…that's a different matter.
**
3. What Should the Agent Be Allowed to Do?
**
This is one of the questions business should answer upfront of development. There is a huge difference between an agent that can advise and one that can act.
- Should it only read information?
- Update records?
- Send emails?
- Approve?
- Make financial decisions?
- When does it handover to a human? These boundaries need to be built in from the start rather than bolted on after go-live. As agents come to greater autonomy, mistakes also have a greater impact. According to McKinsey's 2026 AI Trust study, security and risk was the most cited barrier to scaling agentic AI.
**
4. What Systems and Data Will the Agent Need?
**
An agent doesn't usually operate on its own. It might need to pull information from a CRM, database, ERP, helpdesk, email system, internal documents, or other business tools. So this needs to be part of the discussion early on. Ask:
- Which other systems does the agent need to tap into?
- What info can it read?
- What info can it update?
- Are APIs already in place?
- Is the data clean and ready to access?
- How will we manage permissions?
This can be one part of the project that ends up surprisingly substantial. In fact, McKinsey found that data limitations are one of the biggest barriers to scaling agentic AI (eight in ten companies said they faced data limitations as a barrier in McKinsey's 2026 research). The agent may be smart, but if the data fed to it is partial or incomplete, its output can only go so far.
**
5. How Will We Know If the Agent Is Actually Working?
**
A B2B AI demo doesn't always indicate a successful product. An agent can generate good samples in testing but fail to handle edge cases, bad requests, or missing data. Or it might fail to integrate with other systems.
Or it might return responses but haven't fundamentally solved a problem, created a new revenue stream, or handled edge cases.
Define what it means to succeed before you begin to build. Depending on your use case that might be:
- Less time to answer a user query
- More queries successfully handled
- Less manual work for human agents
- Better resolution rate
- Fewer manual errors
- More qualified leads
- Better worker productivity
That gives the dev team metrics to aim for, rather than asking if the model looks great.
**
6. What Happens When the Agent Gets It Wrong?
**
This is something we don't tend to think about when things work during a demo. But production systems are different. The agent might misinterpret a command, get bad data, run into a system failure, or try to do something it shouldn't.
That's not good; so you need to have a plan.
That plan might be to get human approval, a confidence threshold, fallback workflows, limited capabilities, activity tracking, or a simple system escalation.
The more critical the task, the more delicate these preventative measures must be. An AI agent should not only know what to do when everything is working well. It should also know when to stop.
**
7. Are We Building a Prototype or a Production System?
**
This is a difference that could dramatically alter the scale of an AI agent project. A prototype may indicate that an agent can complete a workflow. But a production system must complete that workflow reliably in the hands of real users, with real data, security, monitoring, failures, updates, maintenance, and other features.
That's why organizations should articulate what they want upfront:
- Do you require a proof of concept?
- An internal pilot?
- A product for your end users?
- An agent that will operate in multiple functions of your organization in the future?
These are vastly different initiatives.
According to McKinsey's 2026 research, although we are seeing an increasing number of organizations adopt agents, organizations are still in different phases of scaling them across functions.
**
The Better Way to Start an AI Agent Project
**
The tendency with AI agents is to begin with the technology.
- Which model?
- What framework?
- How autonomous can we make it?
These are questions, but they shouldn't be the first questions. Start with the workflow. Understand the problem.
- What is it that you want the agent to do?
- What systems, data, and decisions need to be in scope?
- What should be out of scope?
- How will success be defined?
- Where do you draw the line?
Only then, build your technical architecture. Whether you work with an in-house development team or leverage AI agent development services, the best projects start with a business problem, not an AI feature. Our goal is not to build a general agent that can do everything.
Our goal is to build an agent that can do something useful, reliably, and within its defined scope.
Top comments (0)