McKinsey's 2025 survey reports 88% of organisations using AI somewhere, but 39% attributing any enterprise-level EBIT impact to it. That is a gap in reported enterprise-level financial impact, not evidence that other company-wide benefits are rare.
Organisational barriers recur in the research, alongside technical limits. I call the work, data and accountability around the tool the "last mile". That explanation and the sequence below are my synthesis, not a cause established by every source.
Is AI use actually normal now?
McKinsey's State of AI survey (1,993 respondents, November 2025) found 88% of organisations using AI in at least one business function, up from 78% a year earlier. The Stanford AI Index 2026 puts it at the same 88%.
Official statistics count lower because they count everyone, not mostly large firms. The US Census Bureau counted 18% of US firms using AI in a business function, 32% when weighted by employment, for its November 2025-January 2026 reference period. Eurostat counted 20% of EU enterprises with ten or more employees, up from 13.5% in 2024.
These numbers measure different populations and define use differently. Adoption alone does not tell us the enterprise-level financial impact.
How common is enterprise-level financial impact from AI?
McKinsey's 2025 survey puts reported enterprise-wide EBIT impact at 39%. It also reports broader innovation and customer benefits, so EBIT is not a measure of all company-level value. The other figures below measure different outcomes, not one common failure rate:
- McKinsey: 39% of respondents attribute any level of enterprise-wide EBIT impact to AI, and most of those put it below 5% of EBIT.
- IBM's CEO study (2,000 CEOs): 25% of AI initiatives delivered the expected return, 16% were scaled enterprise-wide.
- Gartner, January 2026: at least 50% of generative AI projects had been abandoned after proof of concept by the end of 2025.
- The US Census Bureau, November 2025-January 2026 reference period: among US firms using AI, 57% use it in three or fewer business functions, and 44% report task augmentation. Augmentation can coexist with substitution or task creation. Among firms reporting any task effect, 66% (70% employment-weighted) report augmentation exclusively.
The McKinsey gap is between reported AI use and reported enterprise-level EBIT impact. These other findings add context; they do not establish that AI has little value generally.
How can an AI pilot get parked?
One possible sequence is:
- A tool is bought for a visible use case, often a chatbot or an assistant licence.
- It is piloted inside a process that nobody changes.
- Individuals get real gains, which is genuine and easy to demonstrate.
- Nobody redesigns the workflow, so the gain stays with the individual instead of reaching the process.
- Nobody fixes the data for that process, so results are not reliable enough for work that matters.
- Nobody owns the business outcome, so when a cost or risk question appears, the project has no defender.
- The pilot is parked, and the organisation concludes that AI was not ready, or that a better model is needed.
Gartner's analysis of hundreds of implementations puts poor use-case selection and missing business value first, and poor change management in its top five. RAND's interviews with 65 AI engineers put leadership and organisational causes first, starting with optimising for the wrong problem.
Do technical limits matter too?
Yes. RAND names tasks that remain too hard even for advanced models. Its study excluded projects that simply used a pretrained LLM without further training. MIT also identifies output-quality, memory and adaptability limitations. Organisational barriers recur too, without a shared causal ranking across these sources:
- The work is not redesigned. Deloitte: 84% of organisations have not redesigned jobs around AI. McKinsey, BCG and Bain all find that value comes from redesigning workflows, not from deploying tools.
- The data is not ready. Gartner, surveying 1,203 data management leaders in July 2024: 63% report their organisations lack or are unsure they have the right data practices for AI.
- People and change are skipped. IBM's 2026 study: 83% of CEOs say AI success depends more on people's adoption than on the technology.
- Costs and risk are handled late. Total cost of ownership and responsible AI arrive as afterthoughts, and they kill technically successful projects.
What can't the loudest numbers tell you?
They cannot tell you the real failure rate, because the two most-quoted figures do not hold up well. You have probably seen the claim that 95% of AI projects return nothing. It comes from MIT's GenAI Divide report, which is worth reading and worth quoting carefully: it rests on 52 interviews, 153 survey responses and an analysis of 300 public initiatives, was published as preliminary findings, and is contested by researchers who argue the sample does not support a general rate. S&P Global's 42% abandonment figure is quoted everywhere, but I could not verify its primary page directly.
Do not treat these reported figures as one general failure rate. The McKinsey finding concerns enterprise-level EBIT impact; the studies also describe technical limits and recurring organisational barriers.
What does this mean if you run a smaller business?
You have the most to gain and the least support, and the fix does not require an AI team. The second divide is quieter. Across the OECD, 40% of large firms use AI against 11.9% of small ones. In the EU, Eurostat counts 55% of large enterprises against 17% of small ones. Skills, finance and awareness are the named barriers. The companies that could gain most from automating repeated work are the least equipped to absorb it.
A smaller business cannot fund an internal AI team, so the choice usually looks like a subscription nobody absorbs, or waiting. But the fix is the same at any size: one repeated task that hurts, one number it has to beat, one named owner, agreed checks, and a judgement made on the business metric. That does not require headcount. It requires the last mile to be done on purpose, by someone, for one process.
That is the work I do through CPLT: I scope one process, build the assistance or automation around it with agreed human checks, and hand it over. If you want the method first, I turned the winner patterns from this research into eight questions to ask before approving any AI spend, and the trial itself is covered in how to scope an AI trial before buying tools.
Frequently asked questions
Do most AI projects really fail?
There is no single authoritative failure rate. McKinsey's 2025 survey reports 88% AI use but 39% attributing enterprise-level EBIT impact to AI; that does not measure all forms of company-wide benefit. MIT's 95% comes from 52 interviews and 153 survey responses and is contested, and S&P Global's 42% primary page is not independently verifiable.
Is the AI model the reason projects fail?
Technical limits and organisational barriers both matter. RAND names tasks that remain too hard for advanced models; MIT also identifies output-quality, memory and adaptability limitations. Workflow redesign, data readiness, adoption and costs recur in the research, but the sources do not establish a shared causal ranking.
What separates the companies that get value from AI?
They pick one narrow process that hurts, define the business outcome before the pilot in numbers they already track, name the person who owns that outcome, redesign the workflow including who checks, fix the data for that one process, train the people, and judge the pilot on the business metric rather than on demo quality or usage counts.
What does this mean for a small business without an AI team?
Across the OECD, 40% of large firms use AI against 11.9% of small ones. A smaller business does not need an internal AI team to start; it needs one repeated task, one named owner, one number to beat and agreed checks. That is a scoped engagement, not a hiring programme.
Bring me the task that keeps coming back
If one repeated task is costing you people and hours every week, that is the one to look at, whatever the research says about everyone else. The first 45-minute remote call is free, and you receive a one-page note on fit and next steps. Book the call, or start with the free task questionnaire.