How to measure the ROI of AI initiatives.

By Justin Moore · Founder & CEO, COGScontrol · June 12, 2026

To measure the ROI of AI initiatives, attribute the fully loaded cost of each initiative, choose one value denominator per initiative — the business metric it exists to move — measure cost against that metric continuously, and reconcile both sides to the P&L. Treated this way, AI ROI is an operating discipline, not a one-time business case.

That sequence has a name in this article: the AI Value Ledger. It is a framework a finance team can run a quarterly board cycle on, and it sits inside the broader discipline of AI Value Management — the practice of measuring what AI spending buys, not merely what it costs.

Why can't most companies measure AI ROI?

Because the two halves of the calculation live in different systems, and one half is usually not measured at all. AI costs arrive scattered across model providers — OpenAI, Anthropic, AWS Bedrock, Azure OpenAI, Google Vertex AI — and buried inside cloud bills alongside everything else the company runs. AI value, meanwhile, rarely has an owner or a metric. The result is an asymmetry any controller will recognize: spend known to the cent, value known by anecdote.

This is not a problem confined to companies that under-invest. Uber exhausted its 2026 budget for AI coding tools by April, and its chief operating officer, Andrew Macdonald, was candid that the link between token consumption and shipped value “is not there yet” — the company is still working to draw, in his phrase, “a direct line” from spend to useful features. Further down the org chart the picture is starker: the chief executive of Faros AI told TechCrunch of a CTO whose engineer spent $40,000 on tokens in a single month — “I genuinely don't know whether I should stop him” — while Priceline saw a four-to-five-fold cost increase at its Cursor renewal, and OpenAI's head of enterprise reports that customer conversations now open with how much is being spent and what visibility anyone has into it.

The aggregate evidence points the same way. MIT's NANDA initiative reported in 2025 that about 95% of enterprise generative-AI pilots deliver no measurable P&L impact — a finding contested on methodology, but consistent with what operators describe in public.

The counterexample is instructive. Asked about AI inference costs, ServiceNow's CFO, Gina Mastantuono, could answer in a sentence: “AI reasoning is less than 10% of our cost to serve.” That is what a measured position sounds like — a unit cost, attributed, with a denominator. The rest of this article is about getting there.

None of this is an argument against cost tooling. FinOps teams and cloud cost platforms are genuinely good at their job — rate optimization, commitment coverage, tagging hygiene — and that work keeps paying for itself; the boundary between the disciplines is set out in FinOps vs AI Value Management. But it answers a different question. Cost tools tell you what you spent. AI Value Management tells you what it bought.

The AI Value Ledger: a five-step framework

The AI Value Ledger is a per-initiative operating loop: cost in, value out, reconciled to the P&L, reviewed on a cadence. Run it on every material AI initiative individually — never on AI spend in aggregate.

Step 1 — Attribute the fully loaded cost of each initiative

Start with everything the initiative consumes, not just its model bill. The fully loaded cost of an initiative includes tokens across every provider the initiative touches, plus the cloud infrastructure that serves it: GPU instances, vector databases, orchestration, logging, retries, and evaluation runs. Normalize all of it into one ledger, then classify it by initiative using rules across the dimensions finance already thinks in — cost center, P&L category, product line, environment, project. Two properties make attribution finance-grade rather than engineering-grade: rules reapply retroactively when classifications change, and every change leaves an audit trail. Keep production and development environments separate — experiment spend belongs in the initiative's ROI as cost, but not in the unit economics of serving customers.

Step 2 — Choose the value denominator

Each initiative gets exactly one value denominator: the business metric it exists to move. A support automation exists to resolve tickets, so its denominator is resolutions. An AI product feature exists to retain or monetize users, so its denominator is monthly active users of that feature. Import the metric — revenue, headcount, DAU/MAU, transactions, queries — by CSV or API, and divide. If nobody can name an initiative's denominator, the measurement exercise has surfaced a strategy problem rather than a data problem: the initiative has no defined job. That is worth discovering before the next invoice, not after.

Step 3 — Measure continuously, not as a launch-day business case

AI unit costs drift in a way conventional software costs do not. Model prices change, prompts grow, agents retry, usage scales non-linearly with adoption. Priceline's four-to-five-fold jump at a single renewal is the visible version; the quiet version is the per-unit cost of an unchanged feature creeping upward for weeks before anyone looks. A business case approved at launch is typically stale within a quarter. Track cost per unit on a continuous basis, set budgets with alerts on the dimensions that matter, and treat a moving unit cost as a signal to investigate rather than noise to tolerate.

Step 4 — Reconcile to the P&L

An ROI figure survives scrutiny only if both of its terms tie to statements finance already trusts. On the cost side, reconcile attributed spend to provider invoices daily, not quarterly, so the dashboard number and the invoiced number are the same number. On the value side, use the revenue, usage, and headcount figures the business already reports — not a parallel set of estimates. Then map AI serving costs into COGS so gross margin actually reflects them. An AI feature that looks profitable on a slide and dilutive in the P&L is the expensive kind of surprise, and it is detectable months early if the ledger reconciles.

Step 5 — Decide and report

The ledger exists to force a decision per initiative: scale, hold, or stop. Each quarter, put every material initiative on one line of a board pack — fully loaded cost, value denominator, unit cost, trend, decision — and present it the way any other operating review is presented. The discipline lies in keeping the third option live: a measured initiative can be stopped with a number attached, which is precisely the accountability unmeasured initiatives avoid.

Which value denominator fits which AI initiative?

Match the denominator to the job the initiative was funded to do. The four most common categories map as follows.

The fourth row deserves candor. Internal copilots are the largest AI line at many companies and the hardest to value: cost per active user is computable to the cent, but the value side — output, cycle time, quality — requires a pre-rollout baseline or a holdout group, and most teams have neither. Uber's exhausted coding-tools budget is the cautionary case precisely because the spend side was tracked and the value side was not. The honest treatment is to report the unit cost exactly, report the productivity effect with stated uncertainty, and resist netting the two into one confident figure.

What are the formulas for AI ROI?

The headline formula is conventional; the work is in making its terms honest.

Beneath it sit the unit-economics formulations the board pack is built from:

Measured value is the demanding term. For revenue-side initiatives, value is the uplift against a baseline or holdout — retention, conversion, ARPU — not the gross revenue the feature happened to touch. For cost-displacement initiatives, value is the fully loaded cost of the work displaced, not a list-price salary. The layer beneath these formulas — cost per interaction, per customer, per MAU — is covered in AI unit economics, and you can run your own figures through the AI unit economics calculator.

What are the common mistakes when measuring AI ROI?

Five recur, and each one flatters the result.

• Counting model list price, not fully loaded cost. Tokens can be a minority of an initiative's true cost once GPU serving, vector stores, orchestration, retries, and evaluation are included. ServiceNow's finding that AI reasoning is under 10% of its cost to serve shows how misleading the model bill alone can be.
• Productivity proxies without baselines. Hours-saved surveys taken after rollout, with nothing measured before it, produce numbers no CFO should sign.
• One-time snapshots. A business case at launch, never re-measured, ages badly in a market where model prices and usage patterns shift monthly.
• Ignoring margin drift after model or prompt changes. A model swap or a longer system prompt changes unit cost overnight; without continuous measurement, it surfaces at invoice time, after a quarter of quiet margin leakage.
• Averaging across initiatives. A portfolio-level AI ROI hides one initiative subsidizing four. The ledger is per-initiative or it is decoration.

Where does the AI Value Ledger fit?

The Ledger is the operating core of AI Value Management, and it can be run on spreadsheets — many finance teams start there. COGScontrol exists to run it as software: provider and cloud costs normalized into one ledger and reconciled to invoice every 24 hours, rule-based attribution across five dimensions with an audit trail, business-metric imports, margin-leakage detection, and board-ready reporting — the full scope is on the features page, and pricing is a fixed subscription, never a percentage of AI spend. However you run it, the test of the framework is the same: when the board asks what this quarter's AI spend bought, the answer is a number with a denominator, not an anecdote with a roadmap.