What to test before an agent goes live
Good answers are the easy part. The release decision rests on what the agent does when the answer is not in its knowledge, when the request is ambiguous or out of scope, when it calls an action, and when the person asking should not see the data.
| Category | What it proves | Copilot Studio test method | Blocks release? |
|---|---|---|---|
| Grounded answer | Answers match the source | Compare meaning, Text similarity or Exact match | No, scored |
| Knowledge gap | It admits what the source does not cover | Custom | Yes, if it invents an answer |
| Ambiguity | It clarifies instead of guessing | Custom | No |
| Tool use | It calls the right action, after confirming | Tool use | Yes |
| Escalation | Sensitive requests reach a person | Custom | Yes |
| Authorization | Restricted users see nothing restricted | Custom, run as a restricted user and an allowed user | Yes |
| Privacy and safety | It declines personal or harmful requests | Content safety and Custom | Yes |
| Prompt injection | Instructions in a prompt do not change its behavior | Custom | Yes |
Running the pack in Copilot Studio
- Import the test set. On the agent’s Evaluation page, create a single response evaluation and import the CSV. Copilot Studio expects the headings Question and Expected response, up to 100 cases and up to 1,000 characters per question.
- Pick test methods per case. Compare meaning, Text similarity and Exact match need an expected answer; Tool use needs the expected capability; Keyword match needs the required words; Custom needs your label definitions.
- Run authorization cases as the right people. Evaluations run as the currently signed-in user, so sign in as a restricted test user for the refusal cases and as an allowed user for the positive ones.
- Keep the evidence. Results stay in Copilot Studio for 89 days; export them and attach them to the release decision.
- Mind generated cases. Generated test cases can include sensitive data the connected account can read, and any maker with access to the agent can view the test set; generate with a scoped account.
Acceptance rule
Set pass scores and labels before the first run. Any failed authorization, privacy, escalation, injection or invented-answer case blocks release whatever the other scores say; scored categories then go to the agent owner and business sponsor for sign-off. For the wider pilot decision, pair this pack with the Copilot pilot acceptance scorecard.
Download the test set (CSV, import-ready) and the acceptance plan (CSV). The example agent is a synthetic HR policy assistant on a fictional handbook; its questions and expected answers are illustrations, not client data. Replace them with cases from your own agent.
Primary sources
- Create a single response test set (Copilot Studio) — Microsoft Learn, accessed 2026-10-28
- Choose evaluation methods (Copilot Studio) — Microsoft Learn, accessed 2026-10-28
Frequently Asked Questions
How many test cases can a Copilot Studio test set hold?
Up to 100 test cases in a single response test set. You can write them by hand, generate them from the agent design or its knowledge, or import a CSV or text file whose first row has the headings Question and Expected response, with each question up to 1,000 characters. Results stay in Copilot Studio for 89 days, so export them if you need them as release evidence.
How do you test that an agent respects permissions?
Copilot Studio runs an evaluation as the currently signed-in user, so the evaluation sees what that user can see. Run the authorization cases twice: once signed in as a user who should be refused and once as a user who should get the data. A response that shows restricted data to the restricted user fails the release, whatever the other scores say.
Which Copilot Studio test methods fit which cases?
Compare meaning, Text similarity and Exact match need an expected answer and suit grounded factual questions. Tool use checks that the expected capability was called. Keyword match checks for required words or phrases. Content safety flags harmful content. Custom lets you define labels such as declines and stays in scope, which suits refusals, escalation and injection cases. General quality needs no expected answer.
Can generated test cases expose sensitive data?
Yes, and Microsoft says so: when Copilot Studio generates test cases it uses the connected account to read knowledge sources, generated cases can include sensitive data that account can access, and any maker with access to the agent can view its test sets. Generate with an account scoped to the data the agent should use.
Talk to a senior architect
A one-hour call with a senior EPC Group architect covers your agent, the cases that should block its release and how to run them as the right identities. Email contact@epcgroup.net or call 888-381-9725.