AI strategy for CFOs: where to spend, what to measure, when to stop

Most AI business cases that land on a CFO’s desk are engineering enthusiasm dressed up in a spreadsheet. The numbers are real in the sense that someone typed them; they are fiction in the sense that nobody has pressure-tested a single assumption. Your job is not to kill the enthusiasm. Your job is to make it survive contact with reality.

The summary you can act on: treat AI like any other capital allocation problem — fund it in tranches, tie each tranche to a metric you would defend to your board, and pre-agree the number that makes you stop. The firms that get burned are the ones that approved a lump sum against a payback claim, then discovered eighteen months later that the “savings” never left the slide deck.

A sound ai strategy for cfo purposes is less about the technology and more about the discipline you already apply to every other investment: staged funding, honest measurement, and the willingness to walk away from a sunk cost.

Where AI spend actually pays back

AI does not create value evenly. It clusters. After enough post-mortems you can predict, before a dollar is spent, whether a use case sits in the payback zone or the money pit.

Spend pays back when the work is high-volume, language-shaped, and currently done by expensive people who hate doing it. Think document review, first-draft correspondence, claims triage, contract summarisation, code assistance, tier-one support deflection. The pattern is the same: lots of repetitions, a tolerance for “good enough plus human review”, and a labour cost you can actually see on the payroll.

Spend burns when the use case is bespoke, low-volume, or requires being right every single time with no human in the loop. A model that is 92% accurate is a gift for drafting and a liability for a regulated decision. If the last 8% costs more to close than the first 92% saved — and in regulated settings it usually does — the business case inverts.

The uncomfortable truth for finance leaders: the use cases the business is most excited about are frequently the ones with the worst economics. The flashy customer-facing agent carries brand and compliance risk that dwarfs its efficiency upside. The boring internal document-processing pilot nobody wants to present at the town hall is where the return lives. Push your team toward boring.

The metrics a CFO should demand

The fastest way to expose a soft business case is to ask what it measures. Weak cases measure activity — prompts sent, documents processed, users onboarded. Strong cases measure outcomes tied to a line in your P&L.

Demand these:

• Cost-to-serve per unit, before and after — per claim, per ticket, per contract, per report. This is your headline number and the only one that survives an audit.
• Cycle time, if speed converts to revenue or working capital. Faster settlement, faster onboarding, faster invoice processing all have a dollar value; make someone attach it.
• Fully-loaded cost of the AI itself — not just tokens or licences, but the human review layer, the integration build, the ongoing evaluation, and the internal time to run it. This last category is where hidden cost accumulates.
• Quality and error rate, measured continuously, because a degrading model quietly transfers cost from the AI budget to the rework budget where nobody is watching.
• Adoption, expressed as the share of eligible work actually flowing through the system. A tool used for 20% of eligible volume delivers roughly 20% of its business case, no matter how good the demo was.

If a proposal cannot state its cost-to-serve baseline, it does not have a business case — it has a hypothesis. Fund the measurement first. Our CFO framework for measuring AI ROI walks through building that baseline before you commit capital, and the ROI calculator lets you sanity-check the arithmetic in an afternoon.

The honesty discount on claimed savings

Every projected saving needs a haircut, and you should apply it openly rather than pretend the forecast is gospel. Three discounts matter.

The realisation discount. A projected 40% reduction in handling time does not become a 40% reduction in cost unless you actually remove the cost — reduce headcount, redeploy people to revenue work, or absorb growth without hiring. If the hours saved are simply reabsorbed as slack, the saving is real for the employee and invisible to you. Ask directly: what happens to the freed capacity, and who owns making that happen?

The residual-work discount. AI rarely removes 100% of a task. It removes 60% and leaves a supervision and exception-handling tail. Model the tail. A case that assumes the human disappears entirely is almost always overstated by the cost of the humans who stay to check the machine.

The ramp discount. Savings do not begin on go-live. Adoption climbs over months, models get tuned, edge cases surface. Front-loaded benefit curves are a tell that nobody has run one of these before. Push the benefits right and the payback period lengthens honestly.

A useful rule of thumb: take the vendor’s or the internal team’s projected saving, discount it by a third for realisation, and see if the case still clears your hurdle rate. If it only works at the undiscounted number, it does not work. The AI business case template builds these discounts in so the optimism is visible rather than buried.

Stage-gating so you can stop early

The single most valuable thing a CFO can impose on AI investment is the option to stop. Most organisations structure AI spend so that stopping is embarrassing — a big approval, a named executive sponsor, a public commitment. That is how good money follows bad for a year past the point of obvious failure.

Structure it the opposite way. Fund in tranches, each with a gate and a pre-agreed kill criterion.

1. Discovery gate — small money, weeks not months. Establish the baseline, confirm the data exists and is usable, and validate that the use case is technically feasible. Kill criterion: the baseline cost-to-serve is smaller than assumed, or the data is a mess. Most doomed projects die cheaply here if you let them.
2. Pilot gate — real users, narrow scope, live measurement. Prove the outcome metric moves on a contained slice of real work. Kill criterion: the metric does not move, or adoption stalls despite the tool working. This is where why most enterprise AI pilots fail becomes required reading before you approve the next tranche.
3. Scale gate — the expensive one. Only fund the enterprise rollout, integration and change management once the pilot has produced a defensible number. Kill criterion: the pilot economics do not hold when you add integration and governance cost at scale.

The gates do more than protect capital. They convert a scary all-or-nothing bet into a series of small, informed decisions — which is exactly how you already run every other capital programme. A staged structure also gives you honest optionality: the right to double down when the number is real, and the right to stop without anyone losing face.

Reading an AI proposal as a CFO

You do not need to understand transformer architecture to read one of these proposals well. You need to read it the way you read any investment paper, looking for the tells.

• Where is the baseline? If the current-state cost is asserted rather than measured, the whole return is unanchored.
• What is the human-in-the-loop cost? A proposal that shows the AI cost but hides the review-and-exception cost is understating the denominator.
• Is the saving a headcount decision or a productivity hope? Only the former shows up in your accounts. Make the sponsor own which one they are claiming.
• What is the run cost, not just the build cost? AI systems are not “done”. They need ongoing evaluation, model updates and monitoring. A capital-only view will surprise you in year two.
• What is the exit? If the proposal has no defined stop point, the author has not thought about failure — which means you have to.
• Is the risk priced? In regulated sectors, a wrong output can cost far more than a right one saves. That asymmetry belongs in the model, not in a footnote.

A proposal that answers these cleanly is not necessarily a good investment, but it is an honest one — and honest proposals are the only kind you can actually govern. The dishonest ones are not lying to you so much as lying to themselves, which is worse, because there is no bad actor to catch.

None of this requires you to become the sceptic who says no to everything. The AI opportunity is real, and the firms that allocate to it well will pull ahead of the ones that either ignore it or spray money at it. The CFO’s edge is neither enthusiasm nor refusal — it is discipline: fund small, measure honestly, and keep the right to stop.

If you want a second set of eyes on a business case before it goes to your board, or help building the staged funding structure and measurement baseline, our AI strategy practice does exactly this work. Get in touch and we will tell you honestly whether the numbers hold.