Why AI business cases fail the CFO's test, and what a fundable one contains

In brief

Most AI business cases count hours, skip the baseline and leave out run costs. What finance looks for, a one-page template and a worked full-cost example.

Steve Elliott

Steve Elliott · · 6 min read


On this page
  1. What does finance find wrong with most AI business cases?
  2. Which benefits will a CFO accept?
  3. What does a fundable AI business case contain?
  4. A worked example: an accounts payable exceptions assistant
  5. What should the kill criteria say?
  6. Does every AI initiative need a hard-dollar case?
  7. Sources

In our view, most AI business cases fail the CFO's test for five reasons: they count hours saved as money, they have no measured baseline, they price the build but not the run, they never say when to stop, and nobody owns the benefit.

A fundable case closes all five. It states a baseline measured before launch, a three-year cost that includes running and evaluating the system, a benefit range rather than a point estimate, a named owner who will give up budget when the benefit lands, and kill criteria agreed before the money moves.

What does finance find wrong with most AI business cases?

In McKinsey's 2026 global survey of 1,719 participants, 37% report that AI has contributed positively to their organization's EBIT, essentially unchanged from the year before. Gartner said in November 2026 that productivity gains in most organizations are too marginal to materially affect financial performance or consistently lower headcount. Neither figure says why. In our view, five gaps explain much of it:

  1. Benefit counted as hours. Minutes saved per task, times volume, times a loaded rate, produces a large number that never appears in the P&L. Gartner senior specialist analyst Grace Gstalder says: "AI-generated capacity is a resource, not a return."
  2. No baseline. If cycle time, error rate and cost per transaction weren't measured before launch, any later improvement is an assertion. The FinOps Foundation's February 2026 paper on AI value puts cost and value baselines at the MVP gate, before the pilot.
  3. Run costs left out. Inference, hosting, monitoring, evaluation, model upgrades and support recur every year. About 20% of McKinsey's respondents say AI-related operating costs, including token costs, have constrained their AI use. We covered why labor, not tokens, is usually the bigger line in AI initiative costs: tokens vs labor.
  4. No kill criteria. Without agreed thresholds, weak pilots get extended.
  5. No owner for the benefit. Someone in the business has to own the budget line the benefit lands on. We think the test is simple: will they agree to have their budget reduced when the case says so?

Which benefits will a CFO accept?

Ones with a route to cash and a budget line that moves: contractors released, planned hires not made, revenue a controlled test can attribute. Hours saved drive those benefits but are not one themselves; Hours saved aren't dollars saved sets out the routes and how to measure each. At funding time, the case needs each benefit's route, the evidence that will prove it and the owner whose budget changes.

Give every number a low, expected and high range. HM Treasury's Green Book guidance on optimism bias says appraisers tend to be overly optimistic and that estimates of costs, benefits and duration should be adjusted explicitly, using evidence from past projects.

What does a fundable AI business case contain?

This fits on one page:

SectionWhat it statesThe CFO's question it answers
Decision requestedAmount, stage and date of the next gateWhat am I approving now?
Problem and baselineCurrent volume, cycle time, error rate, cost per transaction, measured over a stated periodBetter than what?
Benefit mechanismWhich route each benefit takes to cashHow does this reach the P&L?
Benefit rangeLow, expected and high, by yearHow wrong could this be?
Benefit ownerNamed person and the budget line that movesWho gives up money when it works?
Full cost, three yearsBuild, run, change and evaluation, by yearWhat will this really cost?
Unit costCost per task before and after, at the vendor's claim and at the kill thresholdDoes it scale, and how far can it slip?
Time to valueRamp assumption and when run rate startsWhen do we see it?
Gates and funding tranchesWhat each tranche buys and what it must proveWhat do I risk before I learn?
Kill criteriaThresholds and the review where they applyWhen do we stop?
Dependencies and risksData access, integration, workforce changes, vendor termsWhat else has to go right?

The FinOps Foundation's paper proposes the same discipline at portfolio level: each approved initiative gets a decision record with scope, budget, success metrics, review cadence and sunset criteria, and funding moves in stages. Its phrase is "Fund the next steps, not entire projects."

A worked example: an accounts payable exceptions assistant

These numbers are illustrative.

An accounts payable team of 40 analysts handles 108,000 invoice exceptions a year, measured over the last two quarters at a median 20 minutes each: 36,000 hours, about half the team's time. At a fully loaded $85,000 per analyst and 1,700 productive hours, an hour costs $50. A vendor demo suggests an AI assistant halves handling time.

The case that fails. 18,000 hours saved at $50 is $900,000 a year, or $2.7 million over three years, against a $400,000 build. It counts every saved minute as cash and has no run cost or owner.

The case that could be funded. In our view, saved minutes spread across 40 analysts mostly get absorbed. Cash comes from two routes. The approved headcount plan adds four analysts next year for 12% volume growth, and the team uses three contractors at $108,000 a year each.

Annual benefit at run rateLowExpectedHigh
Hires avoided ($85,000 each)2: $170,0003: $255,0004: $340,000
Contractors released ($108,000 each)1: $108,0002: $216,0003: $324,000
Total$278,000$471,000$664,000

Year 1 realizes a quarter of the run rate, an assumption. For comparison, Gartner's September 2026 survey of 160 senior finance function leaders found that data extraction, AP and AR automation and report creation generally deliver returns within nine to 10 months. If value arrives around month 9, about a quarter of year 1 runs at full rate.

Three-year costYear 1Year 2Year 3
Build: integration, engineering, security review$400,000
Change: process redesign, training, parallel running$120,000
Run: inference, hosting, observability$50,000$50,000$50,000
Run: support and model upgrades (0.5 engineer)$90,000$90,000$90,000
Evaluation: test runs, 2% human review sample$50,000$50,000$50,000
Total$710,000$190,000$190,000

Evaluation recurs: each model upgrade or prompt change means re-running the test set, and the human review sample never falls to zero. Change cost lands in year 1, when parallel running means some work is done twice.

Three-year cost is $1.09 million. Three-year benefit is about $626,000 in the low case, $1.06 million expected and $1.49 million high. Net, undiscounted: a loss of about $465,000, a loss of about $30,000, or a gain of $404,000. At a 10% discount rate, with cash flows discounted at year end, the expected case's net present value is about minus $95,000.

So the expected case loses money over three years. It pays back only if the system runs into a fourth year at full benefit. Unit cost shows how thin the margin is: about $16.67 per exception today, about $9.90 if the vendor's halving holds, and about $13.24 at a 30% cut (14 minutes of analyst time plus about $1.57 of run cost spread over 121,000 exceptions).

In our view, that is an argument for buying evidence before buying the system. Approve a $150,000 first tranche for a 12-week pilot on 10% of exceptions. The AP controller signs off that, if the pilot passes, the headcount plan drops by the expected three hires, two contractor contracts end on stated dates and the system is planned to run at least four years. On those terms the expected case is worth about $97,000 at 10%. If the pilot points to the low case, the loss stops at $150,000 rather than reaching $465,000.

What should the kill criteria say?

Kill criteria should be numbers set before the pilot and checked at a named review. For the example above, we would write:

  • Handling time. Median exception handling time down at least 30% against the baseline, on the pilot sample.
  • Quality. At least 85% of AI-proposed resolutions accepted without rework, and an error rate no worse than the baseline.
  • Unit cost. Variable inference and hosting cost per exception under $3 at pilot volume.
  • Owner commitment. The headcount and contractor changes and a four-year run plan signed off before the second tranche is released.
  • Time box. A decision at week 12; any extension needs a revised case.

The FinOps Foundation paper gives two rules worth borrowing: sunset when criteria are met or missed for two consecutive reviews, and escalate automatically when a project exceeds budget by more than 15%. We read the rule's first half as a prompt to decide: a pilot that passes should move to the next tranche or be wound down, not linger.

Does every AI initiative need a hard-dollar case?

No. Gartner's March 2026 guidance warns that a narrow focus on traditional financial metrics will significantly undervalue many AI investments, and suggests treating AI as a portfolio of productivity, decision and transformational bets with different cost curves.

We think the ask should match the evidence. A small exploration can be funded on learning criteria alone, with a cap and a date. A case to scale, hire against or sign a multi-year commitment for should meet the full template. The AI total cost of ownership doesn't shrink because the benefit is hard to measure.

A business case is only as good as the link between its benefit, the budget line it changes and the person who answers for it, which is why that link belongs in the operating model rather than in a slide deck.

Sources

Frequently asked questions

Why do CFOs reject AI business cases?

In our view, mostly because the benefit is counted as hours saved rather than cash, with no measured baseline to compare against. Run costs such as inference, evaluation and support are often missing too. Without kill criteria or a budget holder who agrees to give up money, each gap either overstates the value finance can bank or leaves no one accountable for it.

How much should an AI business case ask for up front?

In our view, only enough to reach the next decision. Fund a time-boxed pilot as a first tranche, with kill criteria and the owner's budget commitment agreed before it starts, and release the build and scale money only if the pilot passes. The FinOps Foundation's guidance makes the same point: fund one stage at a time.

What costs belong in an AI business case besides tokens?

Build costs such as integration, engineering and security review; change costs such as process redesign, training and parallel running; and recurring run costs such as hosting, observability, support, model upgrades and ongoing evaluation. Show them by year for at least three years. We think labor usually outweighs token spend, so a case priced on tokens alone will understate the total by a wide margin.

What are good kill criteria for an AI pilot?

Numeric thresholds set before the pilot and checked at a named review: improvement against a measured baseline, a quality bar such as acceptance without rework, a ceiling on variable cost per task, and a signed owner commitment. The FinOps Foundation suggests sunsetting when criteria are met or missed for two consecutive reviews and escalating at more than 15% over budget.

Free playbook

Run the enterprise from any AI.

Read The Headless Enterprise
Steve Elliott

Steve Elliott

Published November 22, 2026

More on AI Economics

Tokens, inference cost, cost per task.