DEV Community

Cover image for Five Agentic AI Examples for Business, From Returns to Reporting
Ahsan Nadeem
Ahsan Nadeem

Posted on Originally published at znovastudio.com

Five Agentic AI Examples for Business, From Returns to Reporting

“The mug broke, the towels are fine, and I already sent the photo. Can I have my money back?” A person reading that ticket knows to look for an earlier conversation. A refund form needs an order number and a line item. Between the two sits one of the clearest agentic AI examples for business. An agent can decide where to look next, gather the missing evidence and bring a specific refund proposal back to the shop.

The same pattern appears in late deliveries, invoice discrepancies, software incidents and unusual sales figures. In each case, someone knows what a finished investigation looks like, but the route depends on what the first lookup reveals. In our illustrative blind comparison of 100 closed cases, the agent saves ten handoffs and makes two mistakes the existing workflow avoided.

AI agents vs workflow automation in one decision

Ask the person who handles the work, “If the first answer changes, what would you do next?” For a dispatch notification, the answer may be identical every time. Read the carrier event, find the customer, send the tracking link. A conventional workflow already does the job. It can have branches, wait overnight and call five systems without needing a model to choose a single step.

Reading an email adds another option. A model can extract the order number and classify a request as a return, exchange or delivery query. Ordinary code can then send each class down a fixed path. That is a useful design when the words vary and the procedure stays the same. A more flexible agent becomes interesting when it has to choose between looking for a prior refund, tracing a replacement parcel or asking which of two purchases the customer means. Anthropic uses this distinction between predefined steps and model-directed tool use in its engineering guide. Building effective agents

Before commissioning workflow automation, write down one recent request that followed each path. If everyone reaches for the same three screens in the same order, build a single view of those screens first. Owners deciding which queue to tackle can also use the shortlist in AI agents for small business.

Five agentic AI examples for business

Business task Decision the agent can make Evidence of a useful result Action with a separate owner
Customer returns Which order, item or prior ticket explains the request? Correct item, remedy and supporting records Issuing the refund
Supplier delays Which shipment or alternative affects the promised date? Updated quantities and dates tied to supplier records Changing the purchase order
Invoice exceptions Which receipt or credit explains a mismatch? Reconciliation down to the invoice line Releasing payment
Software incidents Which log, deployment or service should be checked next? Reproducible timeline and a proposed next step Changing production
Business reporting Which comparison could explain an unexpected movement? Reproducible figures with the metric defined Changing the operating plan

The returns, supplier and invoice cases below are worked examples. Brex and OpenAI have published first-hand engineering accounts of the incident and reporting examples. Their implementations show what teams have put into use. Your team still needs to measure how much of the same work exists in its own queue.

1. Customer service when a return spans two conversations

The simplest returns case starts with an authenticated customer who picks one delivered item and a reason from a list. A workflow can check the return window and create the request. An agent earns a place when the case reaches support because that form no longer describes what happened. A customer may call the replacement “the second one”, attach a photo to yesterday’s chat and ask about a different item in the same message.

The agent’s job is to reconstruct the case. It can read the customer’s permitted orders, follow the reference to the older conversation and ask one question if two items remain plausible. A support lead can judge its output against facts: the right customer, the right line and the right proposed remedy. The worked refund below follows that ticket through to the point where money can move.

2. Procurement when a supplier changes the delivery

Suppose a supplier emails that “the remaining blue units will leave Friday”. The buyer needs to know whether a customer’s installation date will slip. The agent can identify the purchase order, read quantities already received and see which installation depends on the outstanding units. If a second supplier has stock, the next lookup is its lead time. If the first delivery already covered the installation, the investigation ends there.

For example, a purchase order contains 40 units, 24 have arrived and the first installation requires 20. A reply about the remaining 16 can sound urgent even though the first job is covered. The conclusion changes if eight of the 24 received units are reserved elsewhere. Show quantities received, reserved and available in the recommendation. A buyer can then authorise an order change knowing which commitment is actually at risk.

3. Accounts payable when three records disagree

An ordinary three-way match compares the purchase order, goods receipt and invoice. A clean match can follow the finance team’s approval rules. Imagine the invoice lists 100 units and the receiving screen shows 96. An investigator can check whether the supplier invoiced too many, whether a second receipt exists or whether four damaged units are waiting for a credit note.

The finish line is an explanation with record IDs and quantities. “Invoice looks right” leaves the reviewer with all the original work. “96 accepted on receipt A, four rejected on receipt B, supplier credit still outstanding” gives the reviewer a concrete next step. Keep arithmetic and approval thresholds in the accounts system. The model’s contribution is finding the records that belong together, particularly when an emailed credit cites a supplier reference instead of the internal purchase order.

4. IT operations when an alert has several plausible causes

Brex described an engineer investigating a failed accounting export. Its agent gathered logs, identified a relevant code path and posted its findings for an engineer to confirm. Brex reports that gathering context on this incident took about three minutes, compared with its estimate of 30 to 45 minutes for the manual work. That is one incident at Brex. Its more transferable design choice is that investigations start with read access, and reports include evidence, assumptions and a suggested next step. Brex’s engineering account

For a first pilot, choose a recurring alert and make the handoff useful to the next engineer on call. Ask for the first observed failure, the last known success, the affected service and a link that reproduces each observation. If the logs only establish when the failure started, the report should say that. Mistaking the most recent deployment for a proven cause can send a whole response team in the wrong direction.

5. Reporting when someone asks why a number moved

A scheduled sales report follows a fixed query. An investigation starts when a manager asks why orders rose while revenue fell. The answer could involve discounts, smaller baskets or a change in which orders qualify for the report. An agent can check the measure’s definition, inspect the relevant tables and choose its next comparison from the first result.

OpenAI describes this problem in its own data platform. Its internal data agent combines table usage, engineers’ annotations, code, shared business knowledge and live queries. Employees can follow links to the underlying results. The team tests answers to known questions by executing queries and comparing their results with expected outputs. OpenAI’s account of its in-house data agent

For the sales question, ask for the same week, currency and order status on both sides of the comparison. Then ask how much of the change remains after removing cancellations. A generated chart can look convincing with inconsistent definitions. A colleague who can rerun the comparison has a way to catch the error before it reaches the leadership meeting.

A worked refund from customer message to receipt

Suppose an online homeware shop sells two mugs at £18 each and a towel set at £42 in a single order. The mugs ship together and the towels travel in a second parcel. One mug arrives broken. The customer already sent a photo, wants a refund for that mug and intends to keep the other items. The total merchandise value was £78. For this example, the shop’s pilot policy allows a refund for the paid price of one damaged item after a reviewer confirms the evidence. Assume no discounts, tax adjustments or shipping refund. These are design assumptions for a fictional shop.

Establish which order the customer owns. The signed-in account determines which orders the support tool can read. An order number typed into chat is a search hint. Once the order is found, the agent reads the line items and the earlier ticket, then links to the photo already supplied. Asking for a second photo would make an easy case feel difficult while doing nothing to clarify the purchase.

Find the one detail that could change the outcome. If the mug line has quantity two, “the mug broke” leaves the affected quantity open. A focused question such as “Was one mug damaged or were both damaged?” resolves that ambiguity. Once the customer confirms one, the proposal names the mug line, quantity one, £18 and the original payment method. A separate exchange request for a towel on the same order should remain intact.

Give the reviewer a decision they can make in one screen. Show the original message, linked photo, paid price, policy version, remaining refundable quantity and earlier refunds against the line. Present the exact amount and currency before the reviewer approves. Support and self-service must recognise the same damage claim. Assign it one operation ID. In a single database transaction, check the balance and reserve one unit against that claim. If another channel already reserved it, return the existing operation and its status. Two channels can otherwise both read “one available” before either records a payment.

Keep the return and the money movement distinct in the record. Shopify’s return-processing operation confirms which quantities are returned or exchanged and how the merchant handles them. A refund is an optional part of that operation. A support message can therefore correctly say “your return request is open” at a point when “your refund has been issued” would be false. Drive the customer update from the recorded transaction and communicate its actual status. Shopify’s return-processing reference

Recover if the reply goes missing. Suppose the payment service accepts the £18 refund but the network drops before support sees the result. A fresh refund could pay the customer twice. Record a stable identifier for the approved operation before sending it. Use the provider’s documented retry mechanism and reconcile its transaction record before starting any new operation. For example, Stripe’s API v1 can return the first saved response when the same idempotency key and parameters are reused. Each provider has its own rules on retention and errors. Stripe’s idempotent request documentation

Now change one fact. The first £18 refund already succeeded yesterday, and today the customer is asking when it will appear on the card. The same message history and photo may still be relevant. The correct outcome is a status explanation tied to yesterday’s transaction. A pilot that awards credit whenever a refund appears somewhere on the order will mark a second, mistaken payment as a success. This is why the expected change to the record belongs in the test case.

Which business process should you automate first?

Pick a queue whose supervisor can answer four questions with last week’s work in front of them.

  1. What varies between cases? Highlight the exact step at which a person chose a different source or question. Variations in spelling, tone or email layout may justify a classifier. Cases that take genuinely different investigative paths are better agent candidates.
  2. Which record settles the answer? Name the owner of the record and how fresh it needs to be. A carrier scan establishes a parcel’s location at a particular time. A warehouse reservation determines whether an item is available to promise now. If two departments disagree on the authoritative figure, settle that disagreement before asking a model to interpret it.
  3. How would an error be discovered and reversed? A mistaken internal summary can be corrected before the next shift. A customer who receives an incorrect payment confirmation may already have acted on it. Give the pilot only the authority justified by its least reversible action.
  4. What can a reviewer actually save? Time a person assembling the evidence and then time the same person reviewing an assembled case. If the reviewer opens every source and repeats every calculation, improve the handoff before expanding coverage. Budgeting belongs after you know the handling time that remains. Our worked year-one AI agent budget

How to test an AI agent against your current workflow

Create a small evaluation from closed cases. Remove unnecessary personal details, preserve the order in which information arrived and freeze the source records as they stood then. Using today’s ledger to replay yesterday’s enquiry can leak the answer through a refund that had yet to happen. Have a supervisor label the right proposal, any required escalation and the records that should change. The researchers behind τ-bench compared the database after a conversation with the intended end state. That method catches an assistant that writes a reassuring reply but changes the wrong order. The original τ-bench paper

Suppose a fictional test set contains 100 tickets. Seventy contain enough evidence for a standard proposal and 30 require a supervisor. Give each system a fresh, isolated copy of each case. Disable production writes and let simulated actions update only those copies, so you can compare the resulting records. Hide the expected answers from both systems and hide each proposal’s origin from the reviewer. These illustrative results show why you need more than a single completion percentage.

Result on the same 100 cases Fixed workflow Agent
Correct standard proposal 50 60
Standard case unnecessarily handed to a person 20 10
Correctly escalated exception 30 28
Exception incorrectly treated as standard 0 2

The agent would save ten unnecessary handoffs. It would also present two exceptions as routine. Its overall correct decisions rise from 80 to 88, yet only 60 of its 62 routine proposals are correct. If either error would issue money twice, investigate both before granting payment access. You may still have enough evidence to release the read-only investigator and keep every payment with the existing reviewer.

Repeat ambiguous cases several times. Keep the expected result fixed and check whether the assistant arrives there consistently when the order of lookups changes. Also time the humans. Record how often a reviewer reopens a source, how long they spend correcting proposals and whether the customer contacts support again about the same decision. A shorter first reply can hide a longer second conversation. Anthropic’s agent evaluation guide

If you have a queue in mind, bring one routine case, one exception and a record showing how each ended. We can use those to scope a pilot with a baseline your operations team can verify. Book a free consultation.

Frequently asked questions

What is a good first AI agent for a small business?

A read-only investigator for one recurring support issue gives you a finish line you can inspect. Choose a task where records already exist and a named employee can verify the answer. Start with one order system and the relevant support history. Add another source only when reviewed cases show what information was missing.

Can I use an existing helpdesk or automation platform?

Yes, if it can restrict access to the correct account, expose the source records and preserve the decision history. Demo the hardest recent ticket in the existing platform. Ask the reviewer which steps still require switching tools. That gap helps you decide whether configuration, a small integration or a custom agent is justified.

When should an agent ask a person for help?

Give it explicit handoff conditions such as conflicting records, an unverified account or an exception to the published policy. Have it pass the evidence, the unresolved question and any action already recorded. A colleague taking over should be able to continue from that point without asking the customer to tell the whole story again.

Sources

Cover illustration generated for ZNova Studio.


Originally published on znovastudio.com.

Top comments (0)