AI Governance for Finance Teams: A Practical CFO Framework

I suspect a surprising number of finance teams already have an AI policy.

It is just unwritten and different for every employee.

One person will not put a customer name into anything. Another has uploaded half a forecast model because the tool promised to make the formulas nicer. Someone else is using an approved enterprise product but has no idea whether the output can be sent to the board.

This is how governance tends to arrive: after adoption.

I would rather Finance reverse the order.

The first meeting is not about AI

I would put a list of finance information on the screen.

Payroll. Compensation. Customer economics. Pricing. Bank information. Forecasts. M&A work. Investor material. Employee data.

Then I would ask which of it is allowed to leave our controlled environment, which tools are approved to receive it, how the data is retained and who can access it.

That conversation is much less exciting than “What can the model do?” It is also the one I want before somebody pastes a confidential forecast into the shiny new thing.

Those of us who grew up in the 1990s were repeatedly told not to tell strangers on AOL where we lived. Apparently corporate finance now needs the grown-up version of the same talk.

Shadow AI is the real inventory

I would not ask only what tools the company bought. I would ask the team what they are actually using.

Forecast commentary. Spreadsheet formulas. Contract summaries. Variance analysis. Reconciliations. Board-deck drafts. Research. Ad hoc analysis.

If unofficial use is widespread, a policy that simply says “don’t” is probably going to become decorative.

People route around controls that make useful work impossible. Good governance needs safe lanes.

I want three things to be obvious

What data is allowed. What actions require approval. What evidence we keep.

That covers more ground than a 46-page AI policy most people will open once and then admire from a distance.

The approval rule should follow consequence. Drafting internal commentary is different from changing an approved forecast. Recommending an accrual is different from posting one. Preparing an external response is different from sending it.

If an error can move money, alter an official financial number, expose sensitive data or reach someone outside the company, my tolerance for unsupervised execution drops quickly.

“The AI said so” is not an audit trail

For a material output, I want to be able to walk backward.

What source data went in? Which assumptions mattered? What did the system produce? What did a person change? Who approved the result?

Finance numbers travel. An analysis becomes a slide, the slide becomes a sentence in a meeting, and three weeks later the sentence has somehow become institutional truth.

Traceability is how we get back to the beginning.

I would test the embarrassing cases

Most pilots are built to prove a tool works.

I want somebody assigned to make it uncomfortable.

Give it incomplete data. Give it conflicting sources. Change a permission. Ask for something outside policy. Feed it a stale assumption. See what happens when the model is very confident and wrong.

A control environment designed only for the happy path reminds me of parenting advice written for children who are rested, fed and delighted to cooperate.

I am sure it works beautifully somewhere.

Somebody in Finance has to own this

Enterprise AI governance may involve IT, Security, Legal, Risk and Internal Audit. Fine.

Finance still needs a person accountable for translating those rules into finance work: approved tools, access, exceptions, training, monitoring and periodic review.

“Everybody owns it” is usually corporate for “we will identify the owner during the incident review.”

I do not want governance to slow useful AI down. I want it to make useful AI easier because the team knows where it can move without renegotiating risk every Tuesday.

These tools are approved. This information is allowed. These actions can run. These require review. These outputs need documentation.

Then experiment.

I am very comfortable letting Finance move quickly inside a lane we actually understand.

I am less interested in discovering the lane after the robot emails the board.

For the workflow side, I wrote more broadly about what CFOs should actually let AI agents do. The forecast-specific trust problem is different enough that I covered it separately in AI-generated financial forecasts.