The AI use policy that forgot the basics
Most published AI policies skip the plain security rules every policy used to have.
I ran a checker over 34 real, published AI acceptable-use policies, looking not for AI rules but for the ordinary ones: keep your account safe, use approved software, report incidents. Most were missing. 94% had no safe-computing duty. Two in three said nothing about account security or approved tools. The AI clauses were there. The floor under them wasn't.
Companies wrote new rules about AI. But they forgot the old rules, like "use a strong password" and "only use tools we said are okay." A rulebook with only the new rules has a hole in the bottom.
What I measured
- Collected 46 published policies that called themselves AI acceptable-use policies. A classifier kept the 34 that read as internal policies for staff and dropped the marketing pages and customer terms.
- Wrote down twelve things any acceptable-use policy should say, none of them about AI. Account and password duties. Approved software. Safe computing. Monitoring notice. Incident reporting. Consequences. Each one traces to CIS Controls IG1 and ISO 27001.
- Ran a checker over all 34. Present or missing, by pattern match. No judgment calls.
What was missing
| Basic rule | Missing from |
|---|---|
| Safe computing and malware duties | 94% |
| Account and credential responsibilities, including MFA | 67% |
| Approved-software clause | 67% |
| Monitoring notice | 64% |
| Internet and email acceptable use | 61% |
| Acknowledgment and training | 50% |
| Incident reporting | 44% |
| Enforcement and consequences | 44% |
| Prohibited and unlawful use | 32% |
| Acceptable use | 26% |
| Scope and applicability | 23% |
| Data handling by classification | 23% |
The pattern is the point. The most-missed items aren’t AI items. They’re the plain security hygiene that policies carried for twenty years, dropped when someone wrote a fresh document about the new thing.
What to do about it
Don’t write an AI policy. Extend the one you have. The AI rules belong next to the account rules, the approved-tools rule, and the consequences. A standalone AI policy is how the basics fall out.
Run the twelve basics against yours. It takes ten minutes with the list above. If a basic lives in another document, say so in the AI policy and link it.
Approved tools is the one that matters most for agents. An agent is software that acts. If your policy can’t say who approved it and what it may touch, nothing downstream can either.
The honest limits
Thirty-four public policies, not a random sample of anyone. Detection was pattern matching, not a lawyer reading each one; a couple of false negatives were found and fixed during validation, and a small residual error remains, so a few percentages are marginally overstated. Missing from the AI policy doesn’t mean the organization doesn’t do it somewhere else.
Ask your team
- Does our AI policy say who is allowed to use which tools, and who approves a new one?
- Does it still carry the basics, or did those move to a policy nobody reads next to it?
- When someone breaks it, what happens, and does the policy say so?