Reading an AI policy for what it actually requires

An AI policy is a document written to be approved. That is not a criticism; it is a description of the conditions it was produced under. It had to satisfy legal, risk, IT and a sponsor, and it had to pass a committee. What survives that process is often a text that states values clearly and obligations vaguely.

Your job as a reader is narrower than the document. You need to extract the sentences that tell you what to do differently on Monday, and to notice which of your actual questions the policy does not answer.

Sort the sentences into three piles

Rules. Sentences with a subject, an action and no escape. Personal data must not be entered into tools that are not on the approved list. Any AI system used in a customer-facing decision must be recorded in the register before first use. You can comply with these or fail to; there is no third option.

Aspirations. Sentences about being responsible, ethical, transparent and human-centred. They are not useless, because they tell you what argument will be persuasive when you raise something. But no behaviour follows from them, and no one can be found to have breached them.

Delegations. Sentences that move the decision somewhere else. Use of generative AI must be approved by the relevant business owner. High-risk use cases require review by the AI governance forum. These are the most important sentences in the document and the ones most often skimmed, because they are the ones that tell you a decision exists and has an owner.

Go through your organisation's policy with three highlighters. If the rules pile is thin and the aspirations pile is thick, you have a statement of intent rather than a policy, and you should read the next lesson.

The seven questions a usable policy answers

These are the questions that come up in practice. A policy that answers them is doing its job whatever its length.

• Which tools may I use? A list, or a named place where the list lives. Not a principle.
• What may I put into them? Specifically: customer data, personal data, health data, confidential commercial information, code, draft documents.
• What must I not do with the output? Send it to a customer unreviewed; use it as the reason for a decision; paste it into a regulated filing.
• When must something be recorded, and where? Ideally with the trigger being use rather than approval, because approval processes are where entries go to wait.
• Who approves a new use, and how long does it take? A named role and an indicative time. A policy that requires approval but does not say from whom produces unapproved use, not compliance.
• What do I do when something goes wrong, and how fast? A named destination and a deadline.
• What happens if I get this wrong in good faith? The answer determines whether anyone tells you anything.

Three failure patterns worth naming

The prohibition with no alternative. Public AI tools must not be used for work purposes, with nothing approved in their place. People still have the work to do, so use moves out of sight. This policy has not reduced risk; it has made risk invisible, which is worse because it also removes the register entry.

The approval process nobody can complete. Approval requires a form, a risk assessment and a forum that meets monthly. For a two-week pilot, the rational response is to skip it. Approval friction above the value of compliance converts into shadow use reliably.

The policy that predates the tools. A 2023 policy that addresses machine learning models in production and says nothing about assistants embedded in everyday software describes a world that no longer matches the organisation. Most staff AI use is now through features that arrived by vendor update.

What to do with what you find

When the policy does not answer your question, do not treat silence as permission and do not treat it as prohibition. Both are guesses, and one of them is expensive. Ask the delegation sentences who owns the decision, and put the question to them in writing, briefly: what you want to do, what data is involved, what the output feeds. A written question creates a record that you asked, which matters more than the answer in the cases where this goes wrong.

As of 24 September 2026.