AI Policy For Finance Teams (Template Included)

Your finance team is already using AI. Somebody pasted a variance table into a chat window last week to get the commentary started, and it worked, and they did not tell anyone.

That is what happens when a tool is useful and nobody has written down where the line is.

The reason most finance teams have no AI policy is that writing one looks like a legal project, and legal is busy. So here is the version that takes an afternoon: four questions you have to answer, a clause-by-clause template you can copy, and the audit trail that turns all of this from a risk into evidence you ran a controlled process.

Nothing here needs a budget or a committee. Checked 29 August 2026.

Why finance needs its own, on top of the company one

Plenty of companies now have a general AI policy. It says be careful with confidential information, use approved tools, and remember you are responsible for the output. All true, and all useless the moment somebody asks whether a draft budget counts as confidential.

Finance has three problems a general policy never addresses. What counts as financial data, in your terms rather than in the abstract. Who signs off on a number a model produced, and what “signing off” means when the arithmetic came from somewhere you cannot see. And what you can show an auditor in March about work you did in October.

The demand for this is not theoretical. AICPA and CIMA surveyed 1,446 senior finance and accounting leaders in December 2025 and found 88% expect AI to be the biggest technology shift in the field over the next 12 to 24 months, while 8% called their organization very well prepared for it. Most of that gap is process, not technology.

The four questions your policy has to answer

Everything else is detail. If a policy answers these four in language your team recognizes, it works. If it answers them in language a lawyer recognizes, it gets filed and ignored.

Which tools, on which accounts

The boundary is the account, not the brand. The same tool can be covered by an agreement your company signed on one login and covered by nothing at all on another, and that is the distinction people get wrong most often.

Microsoft is explicit that Microsoft 365 Copilot inherits your organization’s existing commitments and that prompts and responses are not used to train the foundation models. Anthropic publishes its position on training data separately for consumer and commercial plans. Both are worth reading once, because the answers differ by plan, and the plan is the thing your policy names.

Write the approved list as a table with three columns: the tool, what it is approved for, and the account type required. Three rows is a fine policy. Fourteen is a wish list.

What data can go in

“Do not put anything confidential in” fails because everybody draws that line somewhere different. Name the categories in your own words and sort them into three buckets: fine on anything, fine on approved tools only, and never.

The never list is short and it is worth being specific. Employee personal data including salary. Customer-identifiable detail. Anything under an NDA. And unpublished results, which matters most in a listed company and is the line an auditor will go looking for.

Who checks it before it counts

This is the clause your auditor cares about, and the one most policies leave vague. Name a role, and define review as checking the calculation against the source rather than reading the output for plausibility.

The reason to be precise is the failure mode. Ask a model to categorize 400 expense lines and it will categorize all 400, including the twelve it had no basis for, and it will not tell you which twelve. A person reading for plausibility will not catch those. A person checking a sample against the source will. That difference is the entire control.

What gets written down

The audit trail, which is the next section, because it is the part everyone underestimates and the part that pays you back.

The audit trail, and why it is the useful part

Most people treat logging as the compliance tax at the end of the policy. It is the clause that changes the conversation, because it converts AI use from something you have to defend into evidence that you ran a controlled process.

Record five things wherever your working papers already live: what was produced and for which period, which tool, the prompt or saved instruction set, who reviewed it and when, and what the reviewer changed. That last one is the one people skip and the one that shows judgment was applied.

Keep it for the same period as the rest of your working papers. There is no need to invent a new retention rule for this.

The practical version: in March somebody asks how the October classification was done. “I asked an AI” is a bad answer. A saved prompt, a named reviewer and a note saying twelve lines were reclassified on review is a fine one. If you are already using Skills or saved instruction sets, you have most of this for free, which is covered in Claude Skills for finance work.

Where to point when somebody asks for a framework

At some point this goes to legal, or to an audit committee, and the question is whether you invented it or based it on something. The answer that ends that conversation is the NIST AI Risk Management Framework, which is free, widely referenced, and organized around govern, map, measure and manage.

You do not need to adopt it formally. Most corporate AI policies borrow their shape from it, and being able to say yours is consistent with it is usually enough to move on to the substance.

Getting it adopted, which is the harder half

A policy nobody helped write is a policy nobody follows. Two things make the difference.

Say out loud that there is no penalty for reporting a mistake promptly, and mean it. The alternative is that people hide it, and something you find in week one is a fraction of what you find in month six. This is the clause that decides whether your policy describes reality or fiction.

Then put one name against each responsibility: who maintains the approved list, who approves additions, who answers questions, who reviews the whole thing in six months. An unowned policy drifts within two quarters, and the vendors will change what they do with your data faster than that.

If you need to make the case upward before any of this gets attention, the argument that works is a number attached to a process rather than an opinion about the future. That conversation is in getting buy-in for finance automation.

What to do this week

Download the template and fill in the brackets. Thirty minutes.

Walk it past whoever owns security or legal before you circulate it. Ten minutes of their time now is worth a great deal later, and going to them with a draft gets a much better reception than going with a question.

Then send it, put the approved tool list somewhere visible, and diarise a re-read for six months out. That is the whole project.

Where to go next

The wider career and governance picture, including what happens to finance roles, is in will AI replace finance jobs. The data side of the same problem sits in data governance in finance. And if you are still choosing which tools to put on the approved list, AI tools for finance has the verdicts and AI in finance is the map of everything else.