How to test an AI agent before it talks to your customers

Trying an agent means asking it a few questions and judging the answers by feel. Testing means writing down, before you look, what a good answer is for each case, and checking the agent against that list after every change. It matters for three reasons.

• Customers are wary of bots. In a Gartner survey of 5,728 customers, 64% said they would prefer companies did not use AI in customer service; their top concern was that reaching a person would get harder (Gartner, survey of December 2023, published 9 July 2024).
• What the agent says, your company says. In February 2024 a Canadian tribunal ordered Air Canada to pay a passenger 650.88 Canadian dollars in damages after the chatbot on its website suggested a bereavement fare could be claimed retroactively, which the airline did not allow. The airline was “responsible for all the information on its website”, the tribunal found, chatbot included (Moffatt v. Air Canada, 2024 BCCRT 149).
• The same question does not always get the same answer. Anthropic’s engineers run each test several times because model outputs vary between runs (Anthropic, January 2026).

You do not need hundreds of cases. Anthropic’s engineering guide to testing agents says 20 to 50 simple tasks drawn from real failures is a great start. We work the same way: about 20 real conversations to learn how your team writes, turned into around 50 cases, and more whenever a situation is not covered. Pick them from your WhatsApp, inbox or call notes:

• The questions that repeat. Opening hours, documents, published prices, where an order stands.
• The ones that went wrong. A customer who misunderstood, a document that had to be asked for three times.
• The ones a person had to decide. An exception, a discount, a technical call.
• The awkward ones. An angry customer, a photo nobody can read, a voice note, a question in another language, a request for something you do not offer.

Take out names, phone numbers and anything personal first. You need the situation, not the person.

For each case, write three things before you see the agent’s reply: what the customer writes, what a good answer must do, and what it must never do. Judge the outcome, not the wording: Anthropic’s guide advises grading what the agent produced, not the path it took. There are many good ways to ask for a clearer photo; there is one outcome you want.

It helps to sort every case into one of three zones:

• It answers. Facts from your sources it can give with certainty: hours, requirements, documents.
• It informs, but does not decide. It explains how something usually works and says clearly that a person confirms it.
• It hands over. Prices outside your rules, technical verdicts, complaints: it passes the conversation to a person, with the history.

• Facts only from your sources. If the answer is not in your information, it says so and hands over. It never fills the gap.
• The right request. It asks for the document or photo the procedure needs, one at a time, and asks again when one cannot be read.
• Its zone. It answers, informs or hands over, as the case says.
• That it is an AI. Since 2 August 2026, the EU AI Act requires providers to design systems that talk to people so that people are told they are dealing with an AI, unless it is obvious (Article 50).
• A way to a person. WhatsApp’s Business Messaging Policy, updated on 23 September 2026, allows automated replies in the 24-hour window but requires “prompt, clear, and direct escalation paths”, such as a transfer to a human agent in the chat (WhatsApp).
• Your voice. Short sentences, the customer’s language, no filler. Most bots that sound robotic are polite in a way nobody on your team is: judge tone against two or three real replies from your team.

Most failures trace back to a rule that is missing, vague or contradicted. We explain how to keep them clean in how to write instructions for AI agents.

• Every hand-over and every complaint becomes a case. Anthropic’s guide recommends turning the failures users report into test cases. Old cases should keep passing, close to 100%: Anthropic calls these regression evals.
• Run the whole set after every change, including a new model. Models change often: Anthropic released Claude Haiku 5.5 on 7 October 2026, and its release notes warn that code written for the previous Haiku can break on it (Anthropic). Before you switch, run the set.
• Read a sample every week. Anthropic suggests sampling transcripts to read weekly, and since 14 October 2026 its platform can send traces of API requests and Managed Agents sessions to a company’s own monitoring tools (OpenTelemetry export, in beta). Whatever you use, keep a record of what the agent did, and look at it.

Every agent we build is tested with the client’s real questions and goes live when it passes. If you already run your own agent, built with ChatGPT, Claude, n8n or anything else, we can build that test set with you, fix what fails and leave you the tests: that is our AI agent setup help.