Answer with specific statements about controls you run today: each agent has its own identity, tool access is allowlisted and reviewed, high impact actions need approval, and calls are logged. Offer evidence, state what is still in progress, and avoid any claim you cannot show. This is not legal advice.
What buyers are really asking
They want to know whether an AI system acting in their data could do more than it should. The questions vary, but they cluster around identity, scope, approval, logging and review. A short, consistent answer set saves time on every questionnaire.
Question patterns and answer shapes
| Question | Answer shape | Evidence to offer |
|---|---|---|
| Do AI agents have separate identities? | State yes or no, and how identities are issued | Registry extract |
| How is agent access limited? | Allowlist per agent, narrow scopes, argument limits | Redacted manifest |
| Are risky actions approved by a person? | Name the actions and the approver role | Policy page, sample approvals |
| Are agent actions logged? | Which fields, retention in your policy | Sample log with data removed |
| How often is access reviewed? | Cadence and last review date | Review minutes |
| Can an agent act on our data without us knowing? | Describe data scope and notification | Data flow summary |
Steps to build the answer set
- Write each answer once, from facts verified with the engineers who run the controls.
- Date it. "As of October 2026" beats an undated claim.
- Mark anything partial plainly: "approval for send actions is live, review cadence began this quarter".
- Link each statement to an artifact you can share under NDA.
- Have a second person check each line against reality before it goes out.
Words to avoid
Avoid "fully secure", "cannot be bypassed" and "industry standard" without a named standard. OWASP describes excessive agency as arising from too much functionality, permission or autonomy, so a credible answer addresses all three, not one. Do not describe any audit, attestation or certification as complete unless you hold it.
When you do not have a control yet
Say so, give the date you plan to have it, and describe the compensating step today. Buyers accept honest gaps better than answers that fall apart in a follow up call.
Keep the answers current
Review the set after every change to your agent controls. A stale answer in a signed questionnaire is a risk of its own.
Safeguard offers MCP Guardrail and agent identity for applying this kind of policy to agent tool calls, and you can run a free scan of your own repositories at gold.safeguard.sh/scan. Autonomous fix pull requests are off by default, and every category starts on Manual.
Quick check
Should I share my full tool manifest?
Usually a redacted extract under NDA is enough. Share enough to show scope, not internal system details.
What if I have no approval step yet?
Say so, give a target date and describe what limits the risk today, such as narrow read only tools.
Sources
Last reviewed: 19 October 2026.
