AI Governance for Finance Teams
Your CFO asks what controls you have around AI. You have about a day to answer, and the answer is that four people on your team are pasting numbers into a chat window and nothing is written down anywhere.
That is where most finance functions are right now. The tools arrived bottom-up, one person at a time, and the governance is showing up afterwards.
Here is the structure I use. Three layers: the data going in, the models doing the work, and the approvals coming out. Write those three down and you have something real. Skip one and you have a document that falls over at the first audit question.
If you came here for data governance in the older sense, ownership, lineage, and who gets to change a definition, that is still in here. It sits in the first layer, and AI raised the cost of getting it wrong, because a model will average across bad data and hand you a confident number in fluent English.
The three layers you are governing
Most AI policies fail because they treat AI as one thing. In a finance function it arrives as at least four: a chat window someone uses to draft an email, Copilot summarizing a variance, an agent that reads invoices and proposes journal entries, and a model buried in a vendor product that turned itself on in the last release.
Those need different rules. Splitting the policy by layer instead of by tool is what makes it survive the next tool.
| Layer | The question it answers | Who owns it |
|---|---|---|
| Data | What is allowed to go in, and where does it physically go | Whoever owns the data today |
| Models | Which tools are approved, for what, and who can see the logs | IT with your finance systems lead |
| Approvals | Who signs off on output before it moves, and against what | The process owner |

Layer 1, the data going in
The rule that does the most work here is to classify by what happens if it leaks, rather than by which system it lives in. Three buckets is enough.
- Internal-safe. Your chart of accounts structure, a variance you already published, a template, a formula. This can go anywhere you would send an internal email.
- Confidential. Results before release, headcount, customer names, unit economics, anything a competitor would pay for. Approved enterprise tools only.
- Restricted. Material non-public information, personal data, anything under an NDA, anything with a salary attached to a name. This does not go into a model at all, in any tier, until legal has signed off on a specific use case.
The failure I see is teams spending three weeks perfecting the buckets and never telling anyone which tool belongs to which one. Write the buckets on one page, then put the tool names next to them. That second step is the whole document.
The other thing most people on your team do not know: consumer tiers of most models train on what you type. Enterprise and team tiers do not. That single difference is most of the reason a policy needs to exist, and it takes one sentence to explain.
If you work at a public company, add one more line. Pre-release results are material non-public information and your model vendor is a third party holding it. That is a conversation with legal, not a policy line you write on your own.
Why the old data governance work suddenly matters more
Definitions are the part that bites. Ask a model why margin is down and it will answer from whatever you pointed it at. If gross margin is calculated one way in the ERP and another way in the reporting layer, you now get a confident, well-written, wrong explanation, which is far harder to catch than a broken formula.
So the boring work still applies. One owner per dataset, one definition per metric, a documented source for anything that reaches a board slide. If you have not done that yet, financial data quality is where I would start, and data cleaning covers the checks worth running before anything is pointed at a model.

Layer 2, the models doing the work
Keep the approved list short. Three questions get a tool onto it.
- Does the contract say your data is not used for training? Get the clause, not the marketing page.
- Can an admin see what was done in it? Ask for the specific mechanism, an audit log or an export API.
- Where does the data physically sit, and for how long?
Question 2 is where the answers get uneven, and it is worth knowing the real state before you write a rule that assumes visibility you do not have. From what I can see across the tools I use daily:
- ChatGPT Enterprise has the most complete logging of anything here. Admin console, retention controls, and a Compliance API that returns conversation content.
- Claude has a Compliance API that covers Cowork sessions, but local session history sits outside that retention path, and Claude for Excel appears in neither the audit logs nor the Compliance API today.
- Copilot is split. Purview covers a lot of the Microsoft 365 surface, but Researcher has no usage dashboard and session content is not visible to an admin.
Write the approval layer differently for each one. Where you cannot see what people did, be stricter about what they are allowed to do it with. A tool with no audit trail can still draft your email. It should not be anywhere near a journal entry.
Shadow AI, and the approval path that prevents it
Every finance team has people using tools that were never approved. What works is a fast approval path. If getting a tool cleared takes six weeks, people use the free tier on their phone and you never find out, which is the exact outcome the six-week process was built to prevent.
My version is a one-page request, a named reviewer, and a five-day clock. Most requests come back as a yes with a data-bucket condition attached.
The AI your vendors turned on without asking
Your close software, your expense system, and your planning tool all shipped AI features in the last eighteen months, and a fair number of them defaulted to on. Send the same three questions to each vendor. The answers arrive slowly, and the ones that arrive last are usually the ones you needed.

Before you write the policy
The AI library for finance teams
The prompts, templates and checklists I use on real finance work, including the data-bucket page this article describes. Free, and it lands in your inbox in about a minute.
Layer 3, the approvals coming out
This is the layer people skip, and it is the one that decides whether the policy is worth anything.
The signature does not move. A model drafting the commentary changes what the reviewer is checking, and that person used to verify the number was pulled correctly. Now they are verifying it was pulled from the right place and that the sentence sitting next to it is true.
Tier the review by consequence rather than by tool.
| Where the output goes | What has to happen | Documented? |
|---|---|---|
| Stays with you. A draft variance comment, a formula, a summary of your own meeting. | Normal review. Nothing extra. | No |
| Leaves your team. A board slide, commentary a business partner will act on, a forecast someone budgets against. | A named person checks the output back to the source, at a level of detail they could show you. | Note the reviewer |
| Leaves the company, or posts to a system of record. A journal entry, a filing, anything a customer sees. | A control step, with the reviewer named and the evidence retained. Treat it like any other manual entry control. | Yes, as a control |
Do you have to say AI wrote it
For internal work, I would say no, in the same way you do not disclose that Excel did the arithmetic. For anything going to your auditors, your audit committee, or into a filing, yes, and you want to raise it before they ask. They will ask.
Whatever you land on, write down that you landed on it. An auditor’s question is almost never whether you used AI. It is how you know the number is right. A policy that cannot answer that second question is decoration.
The one-page version
If you only publish one thing, publish this. It fits on a page and it answers the question people are asking, which is what am I allowed to do on Monday morning.
| Question | The answer you write down |
|---|---|
| Which tools can I use? | The named list, with the tier. Everything else goes through the five-day request. |
| What can I paste in? | The three data buckets, with tool names beside each one. |
| What about pre-release numbers? | Restricted. Nothing before release goes into any external model. |
| Who checks the output? | Whoever signed it before. Tier 2 and 3 work gets checked back to source. |
| What gets logged? | Tier 3 approvals, with reviewer and evidence. Nothing else needs a record. |
| Who do I ask? | One name. If the policy has no name on it, it has no owner. |
What to do this week
- Ask, do not guess. Send one question to your team: which AI tools have you used for work in the last month? Attach no consequences to the answer or the list comes back empty and useless.
- Check the tier on each one. Consumer or enterprise. This takes twenty minutes and usually turns up at least one surprise.
- Write the three data buckets on one page and put the tool names beside them.
- Pick your consequence tiers for review, and name the reviewer for the top one.
- Take the draft to whoever owns risk before they come to you. Walking in with a page is a different conversation than being asked what you have.
That is a week of work. The version that never ships is the one that tried to cover everything, and I have watched that draft sit in someone’s documents folder for two quarters while the team kept pasting.
For the wider picture of where this fits, AI in finance is the map, finance automation covers what the work looks like once the tools are in, and will AI replace finance jobs is worth reading if the question behind the governance question is really about jobs.






