Questions Your CFO Will Ask About AI

AI proposals fail in finance review for predictable reasons, and almost none of them are about whether AI works. They fail because the baseline is missing, the run cost is absent, the savings rest on hours that will never leave the cost base, and nobody has said how the claim will be verified. This guide gives the questions a competent CFO asks and what a strong answer contains.

The questions, and what a strong answer looks like

Question	Weak answer	Strong answer
What does this process cost today?	"It's very manual"	A measured figure per unit, with who measured it and when
What is the build cost, and what is included?	A single number	Discovery, build, integration, evaluation, and deployment, itemized
What is the run cost per task?	Not addressed	Inference, tools, platform share, monitoring, and residual review
What happens at three times the volume?	"It scales"	Cost per task at that volume, with the pricing sensitivity
What controls apply?	"It's secure"	Identity, thresholds, segregation, audit trail, change control
What happens to the freed capacity?	"Staff focus on higher-value work"	Named reinvestment, redeployment, or removal with a mechanism and date
How will we verify the return?	"We'll track benefits"	The specific comparison, the data source, and the review date

Why does the baseline question come first?

Because everything else depends on it. A CFO cannot approve a claim of improvement against an unmeasured starting point, and a proposal that asks for money to improve something nobody measured is asking for faith. If the baseline does not exist, the right proposal is a small discovery tranche to establish it, which is a far easier approval than a full build.

This is also protective for the sponsor: with a baseline, an improvement is provable and creditable; without one, a successful deployment produces an argument about whether anything changed. The how to measure AI success guide covers establishing baselines.

Why does run cost matter more than build cost?

Because it recurs and grows. Build cost is one-off and bounded; run cost scales with volume and continues for the life of the system. Within a year or two it typically exceeds the build. A proposal presenting only build cost is presenting the smaller half of the investment. The AI total cost of ownership guide lists the components that get omitted: monitoring, scheduled evaluation, maintenance, and the human review that persists.

Present the sensitivity too: what happens to cost per task if model pricing changes materially in either direction, and what routing or alternatives exist. That answer demonstrates the risk is understood. The model routing explained for executives piece covers the mitigation.

What does finance require on controls?

If the agent acts in financial processes: a scoped identity per agent, approval thresholds set by finance rather than by engineering, segregation of duties, a complete audit trail, reconciliation of agent activity to source records, and change control covering provider model updates. These are conditions of approval rather than enhancements, and auditors will ask about them. The head of finance operations' guide to AI agents covers the control set.

Why is the capacity question decisive?

Because it determines whether the savings are real. Hours freed and absorbed into the working day produce no financial effect, and a CFO who has seen several efficiency programs knows this. The honest answers are reinvestment in named work that will be measured, redeployment to a specific function with a plan, or removal from the cost base by a stated mechanism and date.

Proposals that present freed hours multiplied by a loaded hourly rate as savings are the single most common reason AI business cases lose credibility. The how to think about AI and headcount guide covers making the decision properly.

How should verification be proposed?

Specifically: the comparison to be made, the data source, who produces it, and the review date. For example, cost per invoice processed from the finance system, compared monthly against the pre-deployment baseline at comparable volume, reviewed by the business owner and finance at the monthly AI review, with a formal reconciliation at two quarters.

Offering verification unprompted changes the character of the conversation, because it signals that the sponsor expects to be held to the number.

What should sponsors prepare before the meeting?

• The measured baseline and its source.
• Build cost itemized, and run cost per task at two volumes.
• The control set, reviewed with finance in advance where agents touch financial processes.
• The capacity disposition, agreed with the function leader.
• The verification method and dates.
• The answer to "what if the provider raises prices?"

How can FISTA Solutions help?

FISTA Solutions establishes baselines before builds, presents run cost per task at multiple volumes, and builds AI agents with the identity, threshold, segregation, and audit controls finance requires, through its AI enablement practice. Since 2017, FISTA has delivered 150+ projects for 50+ companies across 12+ countries; clients report efficiency gains of up to 47% on automated processes.

To prepare a case your CFO will approve, talk to FISTA on WhatsApp, or read the CFO's guide to AI and agentic AI.