AI for CFOs: where it cuts cost, and where it does not

The short answer

Applied AI pays where it removes losses or work from a process that is already measured: reconciliation, exceptions, claims, credit, tax, document review, service. It rarely pays as a general tool rolled out without a target. The CFO's job is to make every AI line in the budget answer to a baseline.

Where AI tends to pay

• Loss reduction: divergences, overpayments and errors caught before they cost money.
• Operating cost: hours of manual review and investigation removed.
• Cycle time: faster decisions, when speed converts into revenue, working capital or fewer penalties.
• Risk: fewer incidents and better traceability, valued at what an incident costs.

Where it usually does not

Broad licences with no process attached. Pilots without success criteria. Projects where the data does not exist yet. And chatbots that answer questions nobody was paying to have answered. These are not technology failures; they are investments without a target.

The costs that show up after launch

The build is the visible cost. After launch come the running cost of models and infrastructure, which grows with use; the people who review and operate the system; evaluation and monitoring, so quality does not silently decline; and changes when the rules or the data change. Ask for a cost per decision, not only a project price.

How to keep it accountable

Agree three numbers before approval: what the problem costs today, what solving it costs in full, and the net benefit over a period you accept. Review them on the same cycle as any other operating line. Set in advance the date and the threshold at which the project is stopped if the benefit does not appear.

A review template for any AI line in the budget

• Problem: which process, and what it cost before the project.
• Target: the metric, the baseline and the agreed target.
• Actual: the measured result to date, from a system finance trusts.
• Full cost: build, running cost per month, people operating and reviewing.
• Cost per decision: running cost divided by the decisions or cases handled.
• Status: on track, adjust, or stop, with the date of the next review.

Asking for this on one page every quarter is usually enough to separate the AI spend that pays from the spend that does not.

Questions to ask the team proposing an AI project

What does the problem cost today, and where does that number come from? If the answer is an industry benchmark, ask for your own.

What will it cost per month to run, at full volume? Model and infrastructure costs grow with use.

Who reviews the system's mistakes, and how much time does that take? Review time is part of the cost.

What result would make us stop? If nobody can answer, the project has no exit.

Build, buy or partner

Buy when the problem is generic and the tool fits your process. Build or partner when the logic that makes you money is specific to your operation, because whoever owns that logic owns the advantage. In either case, make sure the code, the data and the decisions remain yours.