Most enterprises now use AI, yet few can show its effect on the income statement. In McKinsey's 2025 global survey, 88% of respondents used AI in at least one function, but only 39% reported any enterprise-level EBIT impact, and most of those put it below 5% of EBIT.
Teams report hours saved and adoption rates, while finance needs cash, margin and risk numbers that survive an audit committee. This paper sets out a method for closing that gap: baseline, attribute, convert, cost fully and govern.
Its evidence comes from analyst surveys, peer-reviewed economics research and accounting firm guidance from 2024 to 2026. Survey figures are self-reported, the worked example is illustrative, and the accounting content is general information rather than advice for any specific company.
None of this requires a finance team to become data scientists. It asks finance to check whether an AI business case can be audited, and that is visible on the page. Three questions do it, each answered by one named artifact the sponsoring team should be able to produce.
Ask which single financial outcome the use case is accountable for, such as cost per transaction or incremental gross margin. A KPI tree that ends in several dollar metrics lets owners pick the flattering one after the fact. Revenue lift counts at contribution margin, and capacity freed counts only once converted. The artifact is a KPI tree ending in exactly one dollar metric, with guardrails that carry veto power.
Ask what the AI result is being compared against, and accept only a group that ran the old process over the same period. A before-and-after chart cannot separate the AI effect from seasonality, volume shifts and other projects. The method should be chosen and written down before launch, not after the numbers arrive. The artifact is a documented attribution design: an A/B test, a holdout or a difference-in-differences plan.
Ask whether the cost line covers build, data, people and change, human review, governance, run and exit over three to five years. Licenses and tokens are often a minority of the total, and BCG reports that more than 80% of AI spend now sits outside IT budgets, so an IT-only view understates the denominator. The artifact is a full lifecycle cost inventory per use case.
Record unit cost, cycle time, error rate and volume for at least one full business cycle, and name an owner accountable for a single dollar outcome.
Use a randomized A/B test where volume allows, and a holdout or difference-in-differences design for phased rollouts. Treat simple before-and-after comparisons as acceptable only for stable processes.
Release discovery, pilot, scale-up and rollout budgets only when each gate's evidence is met. Pre-commit kill criteria at the first gate and return money from killed use cases to the pool.
Review leading indicators monthly, realized value and gate decisions quarterly in a CFO-chaired council, and the portfolio twice a year, with capacity freed and capacity converted shown side by side.
Set a pre-AI baseline for each use case, attribute change with a control group or holdout, convert operational gains into dollars that show up in a ledger, and subtract the full lifecycle cost.
Report realized ROI separately from projected value and strategic option value. Review the portfolio quarterly against stage gates, so measurement drives funding decisions instead of sitting in a slide deck.
Because productivity gains rarely reach the P&L on their own, several initiatives share credit for the same outcome, and returns arrive late. In McKinsey's 2025 survey only 39% of respondents reported any enterprise-level EBIT impact from AI.
Each of the three problems is solvable, but only when measurement is designed before deployment rather than reconstructed afterwards.
Use a comparison group. Randomized A/B tests give the strongest evidence, while holdout groups and difference-in-differences work for rollouts phased by team, site or region. Research on 5,172 customer support agents used a staggered rollout to estimate a roughly 15% gain in issues resolved per hour. Simple before-and-after comparisons are confounded by seasonality, volume and other projects.
Typically two to four years, according to Deloitte's 2025 survey of 1,854 executives in Europe and the Middle East. Only 6% reported payback within a year, against the seven to 12 months expected of other technology investments.
Productivity J-curve research explains why early returns look weak. Budgets should build that delay in rather than expect first-year returns.
Every dollar needed to build, run, govern and eventually exit the system over three to five years. That covers engineering, data work, model and platform fees, training and change management, human review, governance and exit or migration.
Gartner estimated some generative AI deployment approaches cost $5 million to $20 million, and BCG reports more than 80% of AI spend sits outside IT budgets.
Often yes, for internal-use software. Under ASU 2025-06, capitalization starts when management has committed funding and completion is probable, and not while significant development uncertainty remains. The update applies to annual periods beginning after December 15, 2027, with early adoption permitted.
Experimental AI work may stay expensed longer, so confirm treatment with your auditors before setting policy.
It gives a six-part value taxonomy, ranked attribution methods, a full cost inventory, a four-gate funding model, KPI trees by use case, ASC 350-40 basics, a review cadence and a CFO checklist with a gate scorecard.
The evidence is published surveys, economics research and accounting firm guidance. Survey figures are self-reported, the worked example is illustrative, and none of it benchmarks your company.
Drop your details and we'll send Measuring AI Transformation ROI: A CFO-Ready Framework straight to your inbox - no spam, unsubscribe anytime.
Bring one AI use case your finance team is being asked to fund and the numbers currently used to justify it. We will work through the baseline, attribution design and instrumentation needed to report its value in dollars that hold up at a gate review. SECTION 7 - FAQ - 5 to 8 questions
Book a value measurement session