Finance teams are not waiting to be convinced. Gartner's survey of 121 finance leaders found 58% of finance functions using AI in 2024, up 21 percentage points on 2023, with intelligent process automation (44%) and anomaly and error detection (39%) the most common uses. Deloitte's CFO Signals survey of 200 North American CFOs at companies with at least $1 billion in revenue found 87% expect AI to be extremely or very important to finance operations in 2026, and 48% of those in a leading strategy role said they had already deployed specific AI agents within finance. The question for a smaller firm is not whether, but what an "AI CFO" should be allowed to do.
1. What an AI CFO Is Good For
The repetitive half of the finance lead's week. Turning the pipeline into a revenue forecast and sending it on a schedule. Answering budget-versus-actual and runway questions from the connected books instead of from memory. Flagging deals that have sat in a stage too long, invoices that are late, or a number that moved more than it should have. Preparing the pack for a meeting so the human CFO spends the hour deciding rather than assembling. Each of these is preparation: it produces something a person reads.
2. Where It Has to Stop
Moving money, changing the books, approving spend and signing off numbers are decisions with a named owner, and they should stay that way. OWASP's guidance on excessive agency, the risk of an LLM application taking damaging actions on unexpected or manipulated output, is to limit permissions to the minimum necessary and have a human approve high-impact actions before they are taken. An AI CFO that can wire funds on its own is the textbook case. It should be able to see the books and draft the entry, and a person should post it.
3. Scope It Like a New Hire
Give an AI CFO the finance tools and the reports, and leave out outbound email, membership and administration until you have a reason. In Kavanah that is a capability scope: an AI employee limited to Finance and Reports never sees another tool, whether it is chatting, running on a schedule or working an assigned task. Then set the action policy for moving money and changing the books to wait for approval, or block it outright, for that employee alone. Pick a stronger model for this role if you like; each employee carries its own.
4. Make It Show Up
An AI CFO earns its keep by arriving without being asked. Bind a weekly or daily schedule to it that sends the forecast email, with charts, to the partners. Seat it in the daily executive meeting, where it gives its update from read-only tools, so the numbers in the room come from the system of record rather than from the model. When it needs a decision, it files an approval request on the task and the person sees it in a daily digest, if the workspace has turned the decision window on. It cannot approve its own request; only a person can.
5. Know What It Costs
A forecast schedule and a daily meeting spend AI every day whether anyone reads them. In Kavanah that is metered in credits at about a cent each, at the price of the model the employee runs on, and an admin can set one daily ceiling across every unattended run. Price a month of the schedules you plan before you turn them on.
An AI CFO is a very good finance analyst that never gets tired, and a terrible finance officer if you give it the authority of one. Scope it, schedule it and keep the decisions. For the full picture of AI employees and how vendors compare, see our AI employees page.
