Skip to content
Chat with AI Agent
Business & governance

Six AI Statistics Your Board Quotes, and What They Actually Measure

Orange ITS — AI engineering team 11 min read

A board member forwards a headline on Monday morning: “95% of AI pilots fail.” On Wednesday someone else quotes a different one: “more than half of AI-native startups earn over $400,000 per employee.” Both numbers were published by real organisations. Both are used to decide how much your company should spend on AI, and neither was measured on a company like yours.

This article takes the numbers most often quoted in business AI discussions and asks three questions of each: who measured it, on which population, and what does the figure leave out. If you are deciding where AI belongs in your operations, our AI strategy work begins from the same habit of reading the evidence before spending.

Optimistic and pessimistic AI statistics describe different populations, with different definitions and different incentives. An SME owner should budget against their own measured baseline rather than any headline.


The MIT 95%: Perception, Pilots and 52 Organisations

The figure comes from “The GenAI Divide: State of AI in Business 2025”, published in July 2025 by MIT NANDA. The report describes its method as a review of over 300 publicly disclosed AI initiatives, structured interviews with 52 organisations and survey responses from 153 senior leaders collected at four industry conferences. It covers January to June 2025.

That is a small, non-random sample. The report labels itself preliminary and says its figures are “directionally accurate” because they rest on individual interviews rather than company reporting. It adds that success definitions may differ between organisations.

What counts as success

The report defines a successfully implemented task-specific tool as one that users or executives have remarked as causing a marked and sustained productivity or P&L impact. Its appendix uses a second, stricter definition (deployment beyond the pilot phase with measurable KPIs and ROI measured six months later), so the report does not apply one test throughout.

The headline also blends units. The executive summary speaks of organisations getting zero return, and of integrated pilots extracting millions in value. A later page calls the figure a failure rate for enterprise AI solutions. A reader quoting “95% of AI projects fail” has merged three different denominators into one.

What the same report also says

Another finding in the document rarely travels with the headline. Workers at over 90% of the surveyed companies reported regular use of personal AI tools for work tasks, while only about 40% of companies said they had bought an official LLM subscription. People are getting value from AI. The corporate programmes around them are the part that stalls.

Where the report reads against itself

The report states that external partnerships see about twice the success rate of internal builds, a useful hint for a firm deciding between buying and building, and it labels that finding as self-reported outcomes that may not account for confounding variables. The authors also describe their own work as preliminary. A reader should hold the 95% with the same caution the authors apply.

The fair reading is narrow. In a 2025 MIT NANDA study of 52 organisations interviewed and 153 leaders surveyed, most task-specific or integrated AI pilots had not produced sustained P&L impact as judged by the executives and users interviewed. For why real projects stall, see why AI agent projects fail.


The Gartner 40%: A Forecast Nobody Has Counted Yet

On 25 June 2025 Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027. Reuters, carried by The Star, quoted the reasons as escalating costs and unclear business value, and trade coverage adds inadequate risk controls.

Read the verb. “Will be cancelled” is a forecast about the future. A cancellation that has not happened yet cannot be a failure rate, and the only data point in Gartner’s release is a January 2025 poll of 3,412 webinar attendees on agentic AI investment, which measures spending intentions rather than cancellations. Some commentary nevertheless repeats the 40% as if it were a measured rate.

Agent washing

Gartner also warns of “agent washing”, vendors relabelling assistants and chatbots as agentic products, and estimates that only about 130 of the thousands of agentic AI vendors are real. Our article on agent washing shows how to test a vendor’s claim before you sign.


The AWS 55%: A Sponsor’s Survey of Founders

On 30 June 2026 AWS released a startup trends report. The press release says AI-native startups report 156% average annual revenue growth against 65% for startups overall, and that 55% generate more than $400,000 in revenue per employee.

The population is more than 3,400 founders and senior leaders across 20 countries. These are people who answered a survey, and the release does not describe how they were found, how many declined, or whether revenue per employee means recognised revenue or an annualised run rate. The full methodology sits in a separate report.

AWS sells the cloud and AI services that these startups buy. That does not make the numbers false. It means the sponsor chose to publish them, and the method is not visible in the release. A 55% share of respondents is also no market-wide rate, and a young startup tells you little about a 30-person manufacturer in Aargau.


Revenue per Employee: The Most Quoted, Least Checked Number

Charts of revenue per employee at AI companies circulate widely, mostly from newsletters, trackers and social posts. None of the sources reviewed describes an audited dataset. The numerator might be annual revenue, annualised run rate, ARR or recognised revenue, and the denominator might be full-time staff, contractors or everyone on the payroll. The source rarely says which.

Selection adds another layer. Lists feature companies that raised money, made headlines or went viral. An average over such a list describes the famous outliers, and survivors dominate it.

Midjourney is the usual example. Published estimates of its revenue per employee run from about $3 million to about $5 million, depending on which revenue guess and which headcount (from roughly 40 to over 160 people) a tracker picks. All of them are unaudited, and the same company moves by tens of percent between sources. A chart of fifty such companies deserves caution.

The takeaway for an SME is simple. Do not use any of these figures as a benchmark for your own firm. If a competitor claims a staggering ratio, ask what was divided by what.


METR: A Result That Was Revised by Its Own Authors

METR’s 2025 study is a randomised controlled trial: 16 experienced open-source developers, 246 real tasks on their own repositories, AI tools allowed on some and not on others. Tasks with AI allowed took 19% longer. The developers had predicted a speedup before the study and afterwards still believed AI had made them about 20% faster, which METR describes as mistaken.

Two lessons sit in that result. Perception and measurement can diverge sharply, which is why a survey of how productive people feel is weak evidence. And the sample was 16 people using early-2025 tools, so the result describes that moment and that group.

The 2026 update complicates the story

In February 2026 METR published an update on a later study. The point estimates in the newer data lean towards a speedup, but the confidence intervals include zero, so they cannot exclude no effect or a slowdown. METR calls the new numbers an unreliable signal of the current effect.

Its explanation is concrete. Developers increasingly declined to take part because they do not want to work without AI, which likely biases the estimate of the speedup downwards. METR reports that 30% to 50% of developers said they withheld some tasks, pay for participation fell from $150 to $50 an hour, and time tracking becomes unreliable when developers run several agents at once.

Quoted alone, “AI makes developers 19% slower” is stale. Quoted alone, “AI makes developers faster” is not yet supported by METR’s own data. The honest position is that the question is open, and that the best-designed study in this area says so itself.


McKinsey 2026: Attribution Reported by the Respondents

McKinsey’s State of AI survey is the most transparent of the large corporate datasets, though the figures below come from secondary coverage rather than the primary report. According to Digital Strategy AI, the 2026 edition surveyed 1,719 respondents between 4 May and 8 June 2026, and HPCwire covered it on 26 August.

According to FM Magazine, around 37% reported some EBIT impact from AI, about the same as the year before. About 6% qualified as high performers. Digital Strategy AI puts that 6% on a base of 1,521 AI users (92 organisations), which would be a slightly smaller share of all 1,719 respondents.

What “some EBIT impact” means

These are answers to a question about attribution. A respondent who says AI contributed to EBIT has not shown a ledger. And a respondent who does not report impact has not shown zero impact. They have said that they cannot attribute any, which can mean nothing happened or that nobody measured. The survey is not designed to describe Swiss or European SMEs.

Two more large-firm surveys

Two other sources are worth a short note. An NBER working paper surveyed nearly 9,000 executives across seven countries and found that executives report AI has raised annual labour productivity growth by 0.2% over the last three years, while expecting about 0.6% a year over the next three. These are executive perceptions and expectations.

BCG’s survey of 1,250 executives sorted companies into 5% future-built, 35% scalers and 60% laggards using its own index of 41 capabilities. The split reflects BCG’s scoring, and BCG sells transformation services.


The Headlines at a Glance

StatisticPublisherPopulationWhat it does not show
95% of AI initiatives lack sustained P&L impactMIT NANDA, 202552 organisations interviewed, 153 leaders surveyed, 300+ public initiativesA failure rate for AI projects, or any measured financial return
Over 40% of agentic projects cancelled by end 2027Gartner, June 2025January 2025 poll of 3,412 webinar attendees on spending intentionsAnything that has already happened
55% of AI-native startups above $400k per employeeAWS, June 20263,400+ founders and leaders, 20 countriesThe method, or any result for established SMEs
Revenue per employee at AI companiesTrackers and newslettersFamous or funded companiesAn audited number, or an industry average
19% longer with AI, later revisedMETR, 2025 and 202616 developers in 2025How AI affects developers in general, or non-coding work
About 37% report some EBIT impactMcKinsey, 20261,719 respondents across mixed organisationsMeasured EBIT, or SME outcomes

None of these samples is a random sample of Swiss or European SMEs. The closest local data, such as the AXA and Sotomo survey of 336 Swiss SMEs discussed in AI-native vs AI-enabled, asks about adoption and does not measure revenue or EBIT.


What to Measure in Your Own Company

A headline tells you what happened to somebody else. Your own baseline tells you what happens to your money. The method fits in a few weeks and a spreadsheet.

  1. Pick one process. Choose something recurring where delay or error has a visible cost, such as quoting, invoice coding, support triage or order entry.
  2. Record the baseline before changing anything. Measure cycle time from start to finish, error rate (rework, corrections, complaints) and cost per case (staff time at a loaded hourly rate plus tool costs). Two to four weeks of data is enough to start.
  3. Run the change on a defined slice. Apply the AI tool to one team or one product line so you can compare it with the rest.
  4. Count the full cost. Include licences, integration work, the time staff spend reviewing outputs and the cost of mistakes the tool introduces.
  5. Measure the same three numbers again. Compare like with like, over a similar period, and write down what else changed.

A small illustrative example. Suppose supplier invoices take 12 minutes each to code by hand and 8% need correction. After a pilot, coding takes 5 minutes plus 2 minutes of review, and 5% need correction. Those numbers are invented, so use your own. The structure is what matters: the saving per case is visible, the quality change is visible, and the finance lead can check it.

For choosing which indicators to track, see AI agent KPIs, and for turning the results into a return calculation, AI agent ROI.


Why Your Own Number Beats Every Headline

The reason to measure locally goes deeper than distrust of statistics. Every figure above answers a question asked of a different group: executives who spoke to researchers, founders who answered a sponsor’s survey, 16 developers on their own codebases. Their definitions differ and their incentives differ. A board that averages them is averaging unlike things.

A baseline on one process has none of those problems. It uses your definition of success, your labour cost and your error tolerance. It can be wrong in your favour or against you, and either way it is a number you can audit and defend. For two studies that describe what organisations changed rather than how often AI succeeds, see what 51 successful AI deployments changed and how companies reorganise work around AI.

It also changes the conversation with vendors. A supplier who knows you will measure cycle time, error rate and cost per case on one process before signing a larger contract has to promise something checkable. The accurate summary of the statistics is that adoption is wide, measurement is thin, and the firms that learn the most will be those that count their own work.

Frequently asked questions

Is it true that 95% of AI projects fail?

No, that is not what the MIT NANDA report measured. Its 95% mixes pilots, integrated tools and organisations, and success means that executives or users remarked on marked and sustained productivity or P&L impact. The sample was 52 interviewed organisations and 153 leaders surveyed at conferences, and the report itself calls its figures directionally accurate rather than precise.

Does Gartner say 40% of AI agent projects fail?

Gartner forecasts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs and unclear business value. That is a prediction from June 2025 rather than a measured failure rate. The only data point in Gartner's release is a January 2025 poll of 3,412 webinar attendees about spending intentions, so treat the 40% as an analyst warning about risk.

Why do AI productivity studies contradict each other?

They measure different people, tasks and definitions. METR tested 16 experienced open-source developers on their own codebases, while other studies survey executives about perceived gains. Perception and measurement diverge, since METR found developers believed they were about 20% faster while taking 19% longer. Differences in tools, time period and sponsor interest also move results.

Can a Swiss SME use these statistics for budgeting?

Only as background. None of the major samples is a random sample of Swiss or European SMEs, and most describe large firms, startups, executives or developers. A defensible budget starts from your own baseline on one process, measuring cycle time, error rate and cost per case before and after. That produces a figure your finance lead can audit.

How should a company measure the return on an AI project?

Pick one process, record its cycle time, error rate and cost per case for a few weeks before the change, then measure the same three numbers after. Compare against the full cost of the tool, the integration work and the staff time spent on review. Perception surveys and licence usage numbers are weak proxies for that comparison.

Insights

Put these ideas to work

A 30-minute call is enough to find out whether an AI agent fits your workflow — and what it would return.