A chatbot answering a few familiar questions convincingly is not enough evidence for launch. Real customers omit details, change their minds, make incorrect assumptions and request actions the system cannot perform. Answer quality testing defines acceptable behaviour in advance and gives your team a repeatable way to evaluate those situations.
Define correctness in terms of the customer task
For each important task, record the expected outcome and the information that an answer must include. Exact wording is often unnecessary; explaining the correct service boundary or obtaining required approval matters more. Precise fields such as an order reference need stricter checks: a similar-looking but incorrect value is still wrong.
Assess clarity separately from factual correctness. A fluent answer may contain an invented delivery promise, while an accurate answer may bury the useful information in irrelevant detail. Combining both issues into a single score can make the required improvement harder to identify.
Include difficult cases in the test set
Use this matrix as a starting framework and replace the examples with your own service conditions. Remove personal information when adapting customer conversations. Avoid a set that simply repeats the assistant's prepared example answers.
| Test area | Example situation | Acceptance condition |
|---|---|---|
| Source accuracy | Two services have similar names | Keeps their scope and conditions separate |
| Missing information | No delivery date exists in the source | Does not invent a date; explains the next step |
| Conversation context | The customer changes their selection | Does not act on the previous choice |
| Access boundaries | A different customer's record is requested | Does not reveal unauthorised information |
| Action outcome | A connected system rejects the request | Does not report an incomplete action as successful |
| Human handoff | The customer asks for a colleague | Explains the transfer and its actual status |
Test the conversation beyond its first answer
Imagine a hypothetical appointment assistant where a customer chooses a day and then requests a different one. It must recognise the latest preference, check availability in the appropriate system and report the confirmed outcome accurately. A correct first reply does not demonstrate that the rest of this journey works.
Test everyday phrasing, typing mistakes and alternative ways of expressing the same need. Check that conditions and exceptions remain accurate in every supported language. On mobile, confirm that answers, source links and required controls are usable. The text output alone does not show whether the customer can complete the task.
Do not let an average score hide a critical failure
Separate results by task and failure type. A confusing product explanation and exposure of another customer's information require different release decisions. Define clear blocking conditions for unauthorised access, unapproved actions and reporting an unsuccessful operation as complete.
Microsoft explains that Copilot Studio agent evaluation measures correctness and performance rather than replacing reviews of ethics and safety. Use automated scores as one useful check, then verify critical scenarios through authorised human review and actual system behaviour. A high overall result should not overrule a failed requirement that protects customer accounts or controls business actions.
Compare fixes and releases using the same evidence
Record the assistant version, knowledge version, question, response, expected condition and review decision for each case. After correcting a failure, rerun related scenarios as well as the original question. Changes to sources, models, instructions or integrations can create new verification needs.
After launch, extend the test set using unresolved questions and unsuccessful handoffs. Give each open issue an owner and a clear decision about the next change. Ongoing quality review should establish whether customers can obtain the right information and take the right next step, rather than simply counting the conversations the assistant handles.