How we measure the AI value gap

The three numbers on our homepage — 88%, 39%, ~6% — are the basis of most of what we say about AI value. This page is the footnote: where each figure comes from, exactly what it measures, where the evidence disagrees with itself, and what we'd want to check before believing any of it. Every figure is defined in the table at the end.

The three numbers, defined

All three come from McKinsey's State of AI survey — 1,993 respondents across 105 countries, fielded 25 June to 29 July 2025.

88% is the share of organisations that "regularly use AI in at least one business function." It is a low bar by design. One team using a copilot clears it.

39% report any EBIT impact attributable to AI. Some measurable effect on operating profit, of any size.

~6% is what McKinsey calls AI high performers, and the definition is specific: respondents who both say their organisation has seen "significant" value from AI and attribute EBIT impact of 5% or more to AI use.

That last definition is what makes the number small. It is not "6% of AI projects work." It is "6% of organisations can point to AI moving five percent or more of operating profit." Those are very different claims, and the gap between them is where most of the confusion about AI ROI lives.

Why we show all three

Quoted alone, 88% and 6% describe a cliff, and cliffs read as marketing. The middle number turns it into a funnel:

The middle row is the one that matters, because it describes most organisations we meet. They are not failing. Something is working — a support workflow that's faster, a document process that used to take three people. What they can't do is size it. And an effect you can't size is an effect you can't defend at budget time, can't scale with confidence, and can't tell apart from the thing next to it that's quietly losing money.

Why the gap stays open

The obvious explanation — the technology isn't good enough yet — is not what the evidence says.

Solvd surveyed 500 US CIOs and CTOs in December 2025. 80% attributed AI project failures to inadequate visibility or oversight rather than to the technology. 52% said their organisation is still spending on projects it already knows are underperforming. Not projects that might be underperforming — ones they have already concluded aren't working, still being funded.

Harness put a cost on the same problem in July 2026: roughly one in four pounds spent on AI is wasted. More than half of organisations have nobody specifically accountable for monitoring AI spend — it's scattered across engineering, platform and finance. Only one in five can explain an unexpected AI cost spike within an hour. More than 40% still track it in spreadsheets.

Those findings describe a measurement failure, not a capability failure. Spend is committed before anyone can say what it should return, and reviewed long after the point where the answer could have changed the decision.

Boards have worked this out faster than most delivery teams. In the same Solvd survey, over 80% said their board is now questioning AI spending levels — while 71% still plan to increase AI investment. The scrutiny isn't scepticism about AI. It's scepticism about funding something nobody can measure.

Where the evidence disagrees with itself

We would rather flag this than hope nobody checks.

Domino Data Lab's Enterprise AI Report (July 2026, 639 senior enterprise AI leaders) found 57% of enterprises have ROI that fails to outpace their AI investment — 66.9% in the UK, 67% in Europe, 51% in North America. Which implies roughly a third of UK organisations are ROI-positive.

A third, against six percent. Those look irreconcilable. They aren't, for three reasons:

1. Different bars. Domino asks whether returns outpace investment. McKinsey's 6% requires significant value plus 5%+ EBIT impact — a materially higher threshold. Domino's ~43% ROI-positive figure sits close to McKinsey's 39% reporting any EBIT impact. Compare like with like and the two surveys broadly agree.
2. Different samples. Domino surveyed senior AI leaders at enterprises — organisations invested enough to employ them. McKinsey sampled far more broadly across 105 countries. A population selected for AI maturity reports better AI outcomes.
3. Different dates. McKinsey's fieldwork was mid-2025; Domino's is mid-2026.

The useful conclusion isn't that one survey is wrong. It's that "is AI paying off?" is not a well-formed question until you say what threshold you're measuring against. Most internal AI business cases we review have the same defect.

What we'd want to check before believing any of it

All four sources share weaknesses worth naming:

• Every figure is self-reported. Nobody audited these EBIT attributions. Respondents are describing their own programmes, and the incentive to round up is obvious.
• Attributing EBIT to AI is genuinely hard. If support handling time falls 20% in a year when AI was deployed, headcount changed and volumes shifted, the honest answer to "how much was AI?" requires a counterfactual most organisations never established. Some of that 39% is confident guessing.
• Survey populations skew to larger organisations. Findings may not transfer to a 200-person business.
• The McKinsey figures are over a year old. They remain the current State of AI edition, but a year is a long time in this market. We'll update this page and the homepage when the next wave lands.

None of that changes the picture. The direction is consistent across four independent sources using different methods: adoption is near-universal, measurable returns are a minority, and the binding constraint is measurement and ownership rather than model quality. But "consistent direction across imperfect sources" is a more honest description than a single confident percentage — and it's the standard we'd want applied to our own numbers.

Every figure we cite

Sources: McKinsey, The State of AI · Domino Data Lab, Fifth Annual Enterprise AI Report · Solvd AI Research 2026 (via CIO Dive) · Harness AI cost report (via CIO Dive)

Last reviewed: 19 August 2026.