The ROI Number in Your AI Business Case Is Probably Wrong

Every AI business case I’ve reviewed arrives with one number circled at the top. Sometimes it’s a multiple — 3x, 3.5x. Sometimes it’s a dollar figure with real confidence behind it — $4.2M in year one. That number gets debated in the room, gets approved or doesn’t, and gets remembered long after the deck is gone. What almost never gets debated is what’s actually inside it.

The headline number is usually a survey, not an audit

The most-cited AI ROI figure in circulation is $3.50 back for every $1 spent — from a Microsoft-sponsored IDC study of 2,109 organizations, published in November 2023. It’s a real number from a real research firm. It’s also, on inspection, a self-reported average built from a multiple-choice question: respondents picked a bucket — 2x, 3x, 4x, 5x, no ROI, not sure — and anyone claiming more than 5x was asked to specify further. Five percent of respondents reported as much as $8 back per $1. Nobody audited the underlying P&L. Nobody controlled for who chose to answer the survey in the first place. None of that makes the number fraudulent. It makes it a number that was never built to survive the weight a slide deck puts on it.

That’s the pattern, not the exception. Most of the AI ROI figures executives repeat to each other are self-reported, vendor-adjacent, or both. The gap between what gets repeated and what gets independently verified is where most business cases quietly go wrong.

The gap between adoption and return is real, and it keeps showing up

Four independent research organizations, running different methodologies on different populations in different years, have converged on close to the same finding: most enterprise AI spending isn’t showing up as return.

MIT’s Project NANDA reviewed more than 300 publicly disclosed AI initiatives, ran 52 structured interviews, and collected 153 survey responses from senior leaders across four conferences between January and June 2025. Their finding: despite $30–40 billion in enterprise generative-AI spending, 95% of organizations saw no measurable P&L return. Just 5% of integrated pilots were extracting real value — millions of dollars — while the rest stayed stuck in what the report calls the “GenAI Divide.” I’d be doing this piece’s own argument a disservice if I didn’t flag that the 95% figure has itself been challenged — Futuriom’s critique of the report points out that a conference-intercept survey of 153 leaders isn’t a representative sample of enterprise AI, and treats the number with real skepticism. I think that critique is fair. I’m still citing the finding, because the direction it points is corroborated independently, repeatedly, by firms with very different incentives and methods.

Gartner predicted in mid-2024 that at least 30% of generative-AI projects would be abandoned after proof-of-concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, and unclear business value. Distinguished VP Analyst Rita Sallam’s framing was blunt: executives are impatient for returns, and organizations are struggling to prove the value even exists.

BCG surveyed 1,000 CxOs across 59 countries and found 74% of companies had yet to show tangible value from their AI investment. Only 4% had built AI capability mature enough to generate significant value across functions.

McKinsey’s 2025 State of AI survey — 1,993 respondents across 105 countries — found that only 39% of organizations attribute any EBIT impact to AI at all, and most of those attribute less than 5%. Roughly 6% clear McKinsey’s bar for “AI high performer.”

IBM’s Institute for Business Value surveyed 2,000 CEOs across 33 countries and 24 industries in its 2025 CEO study. Only 25% said AI initiatives had delivered the ROI they expected. Only 16% had scaled past a pilot. Only 29% said they could measure ROI with real confidence. IBM’s own read on the cause: not a technology problem. A culture, governance, workflow, and data-strategy problem.

Different firms, different years, different methods — and they keep landing in roughly the same place. That convergence is stronger evidence than any single number in it.

Three specific errors, not one vague one

“The ROI is wrong” isn’t an argument. It’s a complaint. The argument is that almost every AI business case I’ve seen makes the same three specific mistakes, and they compound.

The denominator is priced against the model, not the system

BCG has been telling clients some version of the same ratio for a while now: in AI transformations that actually work, roughly 10% of the effort goes to the algorithm, 20% to the surrounding data and technology, and 70% to people, process, and organizational change. Deloitte’s global CTO, Bill Briggs, put a sharper number on how far off most companies actually are from that ratio — speaking to Fortune about Deloitte’s 17th annual Tech Trends report, he said companies are spending 93% of their AI budget on technology and 7% on people. He compared it to trying to make paella and ending up with a plate of cilantro: technically an ingredient, not the dish. A business case that divides projected value by the license fee is dividing by roughly a tenth of what the system actually costs to stand up and run.

The numerator is a pilot metric wearing a production number’s clothes

This is the mechanism behind MIT NANDA’s Divide and McKinsey’s EBIT gap at once. A pilot shows a real, local, honestly-measured time saving — an analyst finishes a report in half the time. That number gets multiplied across a headcount and presented as the AI’s value. It never touches the P&L, because the workflow around that analyst was never redesigned, so the saved time gets absorbed into slack rather than reallocated into output. The gap between “users report time saved” and “the business realized more value” is exactly the gap these five research firms keep independently measuring.

The calculation happens once, at launch, and never again

I wrote about this same failure mode from the governance side in Guardrails Are Not Governance — a system’s risk profile doesn’t freeze at launch, and neither does its economics. Usage patterns shift. The workflow it was built for changes. Maintenance and monitoring costs accrue quietly, the same way model drift does. Deloitte’s own 2025 survey of 1,854 executives names this directly: rising AI spend, elusive returns — a real paradox, right there in the report’s title. The same survey found 74% of companies say their advanced AI initiatives meet or exceed ROI expectations — which is a self-report, from the people who championed the initiative, measured however each of them chose to measure it. Both things are true at once. That’s not a contradiction. That’s what happens when ROI is treated as a single number computed once, instead of a trajectory that has to be re-measured.

What rigor already looks like, if anyone reaches for it

None of this requires inventing a new discipline. Forrester’s Total Economic Impact methodology has been doing exactly this for over twenty years — a composite model built from independent customer interviews, decomposed into cost, benefit, flexibility, and risk, with a risk-adjusted ROI at the end rather than a single confident multiple. It’s not a perfect instrument, and Forrester is usually commissioned by the vendor whose product it’s evaluating — that’s a real conflict worth naming. But the structure itself is the right shape: it forces cost, benefit, and uncertainty to sit next to each other instead of collapsing into one number nobody can interrogate. Most AI business cases I’ve seen skip straight to the multiple and skip the structure entirely.

What I’d ask for instead of a single number

Four questions, asked before a number gets approved, do most of the work:

• What’s the loaded cost — not the license fee, the whole system: data readiness, integration, governance, human oversight, change management?
• What fraction of the claimed benefit has actually shown up in production, audited, versus projected from a pilot?
• When does this number get re-measured — not if, when? On what cadence, and who owns it?
• What result would make us kill this? If there isn’t one, the ROI number isn’t a forecast. It’s a decision someone already made, dressed up as evidence.

A number nobody could be wrong about isn’t a measurement. It’s a commitment with better production values.

Part of a larger framework — The System Around the Algorithm