Add Content Safety Checks to Copilot Studio Agent Evaluations, One Content Type at a Time

Each type gets a pass or fail against a threshold from 0 to 7, so add one method per type and decide the permitted level first.

How does the content safety evaluation decide pass or fail?

Test question → agent answer → severity per content type → pass if within the permitted level

For each selected type, the evaluator assigns a severity level based on how much detail the answer contains. If that level is within the threshold you set, the test passes; otherwise it fails. No percentage score is reported.

These evaluation pages are written for agents and agent flows on the standard harness. They do not say whether the same applies to agents on the GitHub Copilot harness.

How do I choose among the four types and levels 0 to 7?

Based on Microsoft Learn, checked on 2026-10-18.

When you create or edit a test set, select Add test method, choose Content safety, pick one content type and a permitted level, then select OK. One method covers one type, so checking all four types means adding the method four times.

Set the level to fit the agent's purpose. For an internal work assistant, starting every type at a low level is a safe default. For an agent whose subject naturally involves violence or self-harm, such as a support guide, use different levels per type and write down why you chose each one.

How do I test questions the agent should refuse?

Split refusal cases across evaluation methods. General quality, which every test set starts with, checks whether the agent attempted to answer, and the documentation says it can flag an appropriate refusal for improvement.

• For questions that should be refused, define the expected refusal in a Custom method with evaluation instructions and labels so it counts as a pass.
• Add content safety to the same questions to confirm the refusal itself contains no harmful content.
• Review refusals with low general quality scores separately instead of treating them as failures right away.

When Copilot Studio generates test questions for you, generation can fail if a question violates the agent's content moderation settings. For sensitive topics, writing the questions yourself or importing them from a file is more reliable.

How many test cases can I add, and in what format?

A single response test set holds up to 100 test cases. For a file import, use a CSV or text file whose first row has the headings Question and Expected response in that order, and keep each question within 1,000 characters including spaces.

In the documentation's table of methods, content safety only needs a content type and a permitted level. The test set creation page also has a general note that every method except general quality needs expected responses or keywords. On the first run, check whether cases without an expected response come back as Invalid.

What should I watch for in results and connected accounts?

Evaluations connect to knowledge sources and tools with the currently signed-in user's account. Sensitive data that account can reach may end up in generated test cases, and any maker with access to the agent can view those test sets. Run evaluations with an account that has only the permissions it needs.

Test results stay in Copilot Studio for 89 days. To compare results before and after changing thresholds, exporting them to CSV after each run is the safer habit. The documentation says content safety evaluations do not guarantee an agent is safe and recommends continuing responsible AI reviews and content safety filters.

For adding human review steps to flows, see what to know before adding human review and long-running work to Copilot Studio agent flows. For moving evaluations off another platform, see the plan for the OpenAI Agent Builder and Evals shutdown.

Start with just 10 questions the agent should refuse, set all four types to level 0, and run it once. Read the answers for the types that fail first, and adjust only those levels.

Sources: What's new in Copilot Studio, Choose evaluation methods, Create a single response test set. Checked on 2026-10-18.