AI governance guide

A one-page AI policy

Long policies do not get read, and an unread policy is not a control. Here is the short version, and what each part is for.

Why most AI policies fail

They are written to satisfy an auditor rather than to guide a decision. "Exercise appropriate caution when using AI tools with sensitive information" tells a member of staff nothing actionable, so they fall back on their own judgement — which is precisely the situation the policy was meant to remove.

A working policy answers one question: can I put this into that tool? Everything that does not help answer that is padding.

The structure that works

  1. Data classes, named in your terms. Four is enough: public, internal, confidential, restricted. Use real examples from your business — "client contracts", "patient records", "supplier pricing" — not abstract categories.
  2. Approved tools, listed. By name. A policy that says "approved tools" without listing them is not a policy.
  3. The routing table. Which class may go into which tool. This is the whole policy; everything else supports it.
  4. What is never permitted, concretely. Credentials, personal data of customers, anything under a specific restriction.
  5. What to do when unsure. Name a person. Not a mailbox.
  6. Verification requirements. Which outputs need checking before use, and how.
  7. Disclosure. When to tell a customer AI was involved.

Worth saying: One page. If it runs longer, you are writing for an auditor rather than for the person deciding whether to paste a contract into a chat window.

The routing table is the policy

Everything else is context. A member of staff holding a document should be able to find its class, find the tool, and get a yes or no in fifteen seconds.

Where the answer is no for a confidential class and the work genuinely needs doing, that is the business case for a private deployment. A policy that forbids useful work without offering a compliant route creates shadow usage rather than preventing it.

Keeping it alive

  • Review when a tool is added, not on a calendar.
  • Make policy sign-off part of purchasing, so new tools cannot arrive unreviewed.
  • Re-read vendor terms annually. They change.
  • Revisit after any incident or near miss, and write down what you changed.

Questions people actually ask

Before you call

Can we just use a template?

As a skeleton, yes. The routing table and the data examples have to be yours — a generic template fails precisely at the point where a member of staff needs an answer about a specific document.

Who signs it off?

Whoever can say no and make it stick. In a smaller business, an owner.

Do we need staff training on it?

A short session, on real examples from your own work. Handing out a document achieves very little. Corporate training covers this properly.

Still have the question?

Ask us directly. We answer these on calls all day.

Certified across the platforms we build on

AWSGoogle CloudMicrosoft AzureAnthropicOpenAI