Expect five follow-ups: who owns the policy, how it is enforced technically, when it was last reviewed, how employees are trained, and what happens when someone breaks it. Prepare one specific sentence for each. A policy that exists only as a document will not survive the second question.
Why the policy answer is only the start
"Do you have an AI policy?" is easy to answer yes. Reviewers know that, so the real questions come next. They are checking whether the policy is a living control or a file on a shared drive.
The five follow-ups
1. Who owns it? Name a role. "The security lead owns it and reports changes to leadership" is a full answer. "Everyone" is not.
2. How is it enforced? This is the big one. Separate rules enforced by tools from rules enforced by trust. Examples of technical enforcement include an allowlist of assistants distributed through managed settings, repository exclusions, tool scoping for agents, and merge gates on pull requests. Rules that rely on people reading and remembering should be labelled that way.
3. When was it last reviewed? Give a date and a cadence. AI tools shift quickly, so a policy untouched for a year looks stale. A twice yearly review is a believable rhythm for many teams.
4. How do people learn it? Onboarding, short refreshers, and examples. Say how you know people read it, such as a signed acknowledgement or a quiz.
5. What happens on a breach? Describe the path: report, assess, contain, and a proportionate response. Mention that pasted secrets are treated as exposed and rotated. See what to do after pasting a secret into an assistant.
A table to prepare from
| Follow-up | Evidence to hold |
|---|---|
| Owner | Policy header with role and date |
| Enforcement | Screenshot or config of managed settings |
| Review | Change log or ticket |
| Training | Acknowledgement record |
| Breach path | Incident procedure |
Common weak spots
- A policy that names tools no one uses and omits the ones people do.
- No mention of agent features, which can run commands.
- No exception route, so people work around the policy quietly. A short exception process is healthier. Related: what an AI coding assistant policy should include.
Answer template
"The policy is owned by the security lead and was last reviewed on a date in the policy header. Approved assistants are distributed through managed settings, agent actions that change systems need approval, and merges require review and scanning. Employees acknowledge the policy at onboarding and annually. Violations follow our incident procedure."
Fill every blank with a fact. If you cannot, plan the fix and say so.
Safeguard can help with the enforcement side through policies, gates and guardrails, and with proof through compliance and evidence automation.
Quick check
Is a signed acknowledgement enough proof of training?
It proves people were told. Add a short example based refresher for stronger evidence, and say how often it runs.
What if the policy is new?
Say so and give the date. A recent, specific policy is better than an old vague one.
Last reviewed: 19 October 2026.
