How to measure the ROI of AI — so it shows up in the P&L

The moment the question lands

I have sat in enough boardrooms to know the exact moment the energy shifts. The CFO — or a non-exec — asks "What is the return on this?" and the room goes quiet. Not confused quiet. The quiet of people who hoped that question wouldn't come up yet.

The management team reaches for the dashboard. The dashboard is green. Adoption rate: up. Usage frequency: steady. Hours saved: estimated. The question is not answered — it is moved past.

By the time it reaches the board, it has been whitewashed. Not dishonestly. The team genuinely believes the activity data tells the board something useful. It does not. The board is seeing what the programme is doing, not what it is changing. Those are different things — and only one of them ends up in the P&L.

Why traditional ROI can't see AI's return

Only 29% of executives report measuring AI ROI confidently, according to McKinsey's State of AI survey. If 71% of management cannot measure it, the board is governing blind. That is not a measurement execution problem. It is a tool selection problem.

ROI was built for capital projects with a defined input, a defined output, and a time horizon short enough to close the loop. Replacing a machine. Building a factory. Running a campaign. The return is measurable because the boundaries are clear.

AI investment does not work that way. Value accumulates in changes to buyer position, decision quality, and pipeline. Pilot output captures only a fraction of it. Gartner estimates that 95% of generative AI pilots fail to scale to production. If most pilots never reach the stage where ROI is measurable, then measuring ROI on the pilot is measuring the wrong thing — you are defending a capital allocation on a metric that applies to one programme in twenty.

Two failure modes repeat. In the first, management defers metrics until "we know more." Nobody sets the trigger. By month 18 the programme is infrastructure, the ROI question no longer applies, and the board never had a stop-condition. In the second, activity stands in for value: adoption rate, usage frequency, hours saved — real numbers, none of which connect to a commercial outcome the CFO can audit. I have seen six AI pilots running in one organisation, no shared intent layer, each technically live and operationally inert. The board was handed a dashboard of activity dressed as a dashboard of value.

Meaningful AI has to show up in the P&L. Without the right signals, organisations keep funding the wrong pilots and kill the right ones.

The Three Leverage Questions

The replacement is not a more sophisticated ROI model. It is a different question — three of them. Each maps to an outcome the board already governs. The board does not need to understand the AI capability in detail. It needs to ask: which of these three signals is this investment moving? If none, that is the answer.

1. What changes in buyer position?

Does this capability change how customers evaluate or select you — before the RFP, before the shortlist, before the first meeting? Buyer-position change maps to market share risk, which boards already own. It is invisible to an ROI framework because it registers in pipeline dynamics, not pilot output — but it is real, and it compounds.

2. What changes in decision quality?

Does it change the quality of decisions made in the business, by people or by systems — faster, better-informed, more consistent? Poor decisions are expensive, and the cost stays invisible until it accumulates. Decision quality maps to operational risk. A board can hold management to account on it without understanding the underlying technology.

3. What changes in pipeline?

Does it change the volume, speed, or conversion of commercial opportunities? Pipeline is the most commercially legible signal — it connects directly to numbers the CFO already measures, and it exposes fastest whether a pilot is creating genuine leverage or just activity that looks like leverage from the dashboard.

If an investment is not moving at least one of the three, that is the board's answer. Don't get stuck in pilots waiting for ROI to emerge.

The questions in practice

Take a manufacturer running an AI document-processing pilot. Score it:

• Buyer position — no change. It is an internal process capability; it does not affect how customers evaluate the organisation.
• Decision quality — yes. Compliance review cycles are faster and better-informed.
• Pipeline — no direct effect. Opportunity volume, speed, and conversion are unchanged.

One signal of three is moving. The board with this framework can say: "That justifies continued investment at current scope. It does not justify scaling." That is a defensible position. "The dashboard is green" is not.

The governance the board has to require

The Three Leverage Questions only work if the board asks for them before the programme is approved. The most common reason AI ROI cannot be measured is not missing data — it is that no measurement architecture was ever required.

Before approving any AI programme, a board needs three anchors from management: which leverage signal this investment targets, a measurable definition of what movement looks like at six months, and a stop-condition if the signal does not move. Without them, there is no defensible basis for a keep, kill, or scale decision when the programme comes back. For that call, see when to kill an AI pilot; for accountability, who owns the AI outcome.

Most boards I work with have named the problem accurately: there is an altitude problem. Management measures activity. The board approves programmes. Nobody requires governance-grade signals at the point of approval. The Three Leverage Questions close that gap — without the board becoming technical.

Frequently asked questions

How should a board measure the ROI of AI?

Not with a pilot-level ROI figure. Ask what the investment changes in three places the board already governs — buyer position, decision quality, and pipeline. Those signals connect AI spend to commercial change, and to the P&L, without requiring the board to understand the technology.

Why is measuring ROI on AI pilots the wrong approach?

ROI assumes linear, discrete value: a defined input produces a defined return. AI investment creates compounding capability whose value accumulates in decision quality, buyer position, and pipeline. Forcing a linear ROI frame onto a non-linear investment systematically undervalues the right pilots and sustains the wrong ones.

How does a board know if AI investment is making money if ROI is the wrong measure?

If the investment is not moving at least one of the three leverage signals, it is not making or saving money either — regardless of what the activity dashboard shows. Movement in the signals is what eventually shows up in the P&L.

What is the most common reason AI ROI cannot be measured?

The measurement architecture was never required before the programme started. ROI was deferred until "we know more," nobody set a trigger, and the deferral hardened into doctrine.

Are the Three Leverage Questions a replacement for financial metrics?

No. They are a diagnostic layer above financial metrics that connects AI investment to outcomes the CFO already measures. Pilot ROI can still be tracked internally; the board's governance question is which investments are moving the signals that matter commercially.

Stefan Finch is an AI board advisor and the founder of Graph Digital, working with boards and non-executive directors across UK mid-market and enterprise organisations.

If you are navigating this in your boardroom, Subscribe to Agentic Leaders — a weekly brief for board directors and executives governing AI — or work with Stefan directly.