Blog /Guides

The CFO's Guide to AI Governance in the Finance Function

Anirudh Badam

Anirudh Badam

Co-Founder and CAIO, Adopt AI

16 November 2026

An empty boardroom with a long table, chairs and microphones

Your external auditor asks a plain question in the planning meeting: "Which entries did the AI prepare, and who approved them?" If your answer is a shrug, the audit committee will hear it too. AI governance in finance exists so that you answer that question in one sentence, with a log to back it up.

This guide is for a CFO or controller who has to say yes or no to AI in the finance function. It lays out six controls to require before any agent touches the close, and the question to ask a vendor about each.

What AI governance in finance covers

Governance in finance has a narrower job than governance for AI in general. You do not need an ethics framework for a bank reconciliation agent. You need to know four things: what it can read, what it can change, who signs off, and what record it leaves.

Your team already runs this discipline on people. A preparer drafts a journal entry, a reviewer approves it, and the general ledger keeps the history. An AI agent gets the same treatment. It is a preparer with a named reviewer, a defined scope, and a log.

That framing helps with the room, too. IT hears access control. Audit hears segregation of duties. The board hears a policy it already understands.

Control 1: A sign-off gate before anything posts

Require that no agent writes to the ledger without a named person's approval. Read the vendor's "autonomy" settings. Some tools post on high confidence and call it a feature. For the close, set every step to "suggest" or "act with approval" until a step earns more room on a track record you can see.

Ask the vendor: Where does the review gate sit, and can I put one at every step?

At Adopt, nothing posts without sign-off. The agent prepares the work, and a person approves what ships.

Control 2: A firewall between the language model and the numbers

A language model predicts text. It does not add. Any design that lets the model produce a balance, a variance, or a tax amount puts a guess where an audit expects a calculation.

The control to require has three parts:

  • Deterministic code computes every figure.
  • The model reads documents, drafts explanations, and proposes matches.
  • A second model from a different family checks the first before a person looks at the result.

The two-model check matters because two copies of the same model share the same blind spots. Different families disagree in useful ways. A tie-out against your ERP's own report should sit on top of both, since a model checking a model is not independent validation.

Ask the vendor: Which step in your workflow could produce a number from a model's output alone?

Control 3: An audit trail down to the source

Every number should link to the transaction or document behind it. Every proposal, override, and approval should record who or what acted and when. Preparer and reviewer should appear on every row.

Test this the way your auditor will. Pick one reconciling item from last month. Ask the vendor to show the source document, the agent's proposal, the reviewer's decision, and the timestamp, all from one screen. If you have to export four reports and stitch them together, the trail is thin.

Audit trail and audit readiness are worth a refresher if your team has not written the requirement down.

Control 4: Narrow access and a read-only start

Give the agent the least access that does the job. Start with read-only connections to the bank feed and the ERP. Put credentials in a secret vault, and use the same login a staff member would use for portals, not a shared password in a script.

Adopt never moves money. That boundary belongs in your policy as well as in the vendor's product. Write down which systems an agent may read, which it may write to, and which are off limits.

Ask the vendor: What happens to my credentials, and can I revoke them in one step?

Control 5: Data residency and model choice

Your IT lead will ask where the data goes and which models see it. Have the answers before the meeting. The terms to require:

  • Deployment in your own cloud, a private VPC, or on-premises if your policy demands it.
  • The option to use your own model keys.
  • Zero data retention at the model provider.
  • No training on your data.

For regulated or sensitive entities, this control often decides the purchase. AI accounting data privacy covers the questions in more depth, and AI finance security questions gives your IT team a list to send the vendor.

Control 6: Ownership of what gets built

Governance includes what happens if the relationship ends. If the agent definitions, skills, and context live inside a vendor's product, your controls depend on the vendor's roadmap and survival.

Require that you can export the skills and agent definitions and run them elsewhere. At Adopt, forward-deployed engineers build the first workflows with your team, and you own what gets built.

Ask the vendor: If we cancel, what leaves with us, and in what format?

A one-page policy to write before the pilot

You do not need a forty-page policy to start. One page, signed by you, covers it:

  1. Scope. The workflows in the pilot, and the systems each may read or write.
  2. Owners. A policy owner (you), an operator (the controller), an access owner (IT), and a tester (internal audit).
  3. Approval rules. Which steps need a person's sign-off, and who.
  4. Evidence. What the log must show, and how long you keep it. Set the retention period with your auditor, not the vendor.
  5. Exit. What you export and how, if you stop.

Then prove the controls on a month you already closed. Nothing posts, and your auditor can read the log and tell you what is missing before real entries depend on it. The 90-day AI pilot plan shows how to set the go or no-go criteria.

Where to go next

If your team is still choosing tools, evaluating AI finance tools turns these controls into vendor questions, and internal controls for AI goes deeper on testing them.

To see the controls on your own data, book a pilot. We start with a closed month, nothing posts, and your team and your auditor decide what happens next.

FAQs