The Adoption Gap

01

Why this, why now

In the space of eight months, the most quoted number in enterprise AI was born, moved billions in market value, got quietly deleted, and was publicly dismantled. Meanwhile the technology it described kept getting an order of magnitude cheaper every year. That collision is the story of this dossier.

Every boardroom now holds two beliefs at once. The first: AI is everywhere, our people use it daily, we have pilots running. The second: we cannot point to a line in the P&L that AI improved. Both beliefs are correct, and the distance between them has become the defining management question of this technology cycle. The loudest fight in enterprise AI, a 95 percent failure claim against a 29 percent traction claim, turns out to dissolve on close reading. What remains is harder and more useful: adoption is now table stakes, and profit belongs to a small group that reorganised work around the machine. This dossier adjudicates the numbers honestly, anatomises why pilots die, and ends with what we would actually do.

• Aug 2025 — Fortune headlines an MIT-branded claim that 95 percent of GenAI pilots return nothing. AI stocks wobble.
• Sep 2025 — After 29 days, the report disappears from MIT's website. The number stays in circulation.
• Nov 2025 — McKinsey's State of AI: 88 percent of firms use AI, 6 percent see real profit impact.
• Jan 2026 — Exponential View publishes its forensic teardown. MIT calls the report "unpublished, non-peer-reviewed work."
• Apr 2026 — a16z counters with hard data: 29 percent of the Fortune 500 are live, paying customers of AI startups.
• 2026 — Agents enter the enterprise. 62 percent of firms are experimenting; 21 percent have governance ready for them.

02

1. The number that moved a market

On August 18, 2025, Fortune reported that an MIT study had found 95 percent of corporate GenAI pilots failing to produce any return. Within days the number was in every deck, every board pack and every skeptic's back pocket.

The report behind the headline was called "The GenAI Divide: State of AI in Business 2025," produced under MIT's Project NANDA. Its opening line did the damage: despite 30 to 40 billion dollars of enterprise investment in generative AI, "95% of organizations are getting zero return." The methodology, buried further down, combined 52 executive interviews, 153 leader surveys and a review of about 300 public AI deployments, gathered between January and June 2025. The report's own diagnosis was organisational rather than technical: companies buy or build static tools that demo well, then discover the tools cannot remember context, adapt to feedback, or fit the way work actually flows. The authors called it the learning gap.

The market reaction was real. Coming days after a summer of record AI capital spending, the claim that almost none of it lands as profit spooked investors and knocked AI-linked stocks. The number then began to travel on borrowed authority, cited as "MIT research" in speeches, investment memos and procurement debates, usually without a footnote. By late 2025 it had become the standard opening move of every AI skeptic in every strategy meeting in the Western world.

03

2. The autopsy of "95 percent"

In January 2026, Azeem Azhar's Exponential View published the result of a months-long forensic trace of the number, including correspondence with the authors and with MIT itself. The verdict: viral, methodologically weak, and buried under its own caveats.

The trail is worth walking, because the details change what the number means. The report was a preliminary draft, never peer-reviewed. It lived on an MIT web domain for 29 days, from August 18 to September 16, 2025, and was then removed; the PDF that still circulates carries MIT branding but has no canonical home. When Exponential View pressed the institution, MIT Media Lab's faculty director described it as "a preliminary, non-peer-reviewed piece," and MIT's media relations office eventually stated, in writing: "It was unpublished, non-peer-reviewed work."

The statistics inside were shakier still. No confidence intervals were provided; on a core sample of around 52 interviews, the honest interval around "95 percent" spans roughly the high 80s to 100. The denominator shifts between pages: at one point the failures include organisations that never even attempted the tool category being measured, which counts not-boarding-a-flight as a plane crash. And the fieldwork window, January to June 2025, judged eighteen-month rollouts and three-month-old pilots by the same yardstick, in a market where enterprise AI spending was tripling per year. Rerun the arithmetic with the report's own numbers and a defensible denominator, pilots actually launched, and the success rate lands near 5 in 20, about 25 percent, with wide error bars. Azhar's conservative rewrite of the whole picture: by early 2025 perhaps one in seven organisations was already seeing measurable gains, possibly more. For a technology that entered the corporate bloodstream in late 2023, that is a diffusion success story wearing a failure headline.

"The '95 percent' figure should be treated for what it is: not reliable. It is viral, vibey, methodologically weak and it buries its caveats." — Exponential View, "How '95%' escaped into the world," January 2026

None of this makes the report worthless. Its qualitative core, the learning gap, matches what everyone in the field sees daily, and this dossier will lean on that insight repeatedly. The lesson is narrower: the headline number was a temperature check dressed up as a measurement, and it moved capital for half a year before anyone checked its pulse.

04

3. The counter-claim, and the honest adjudication

In April 2026, a16z compiled the opposite dataset: 29 percent of the Fortune 500 are live, paying customers of leading AI startups. Signed contracts, converted pilots, production usage. So who is right?

Both are, and that is the actual finding. The a16z report measured something deliberately hard: enterprises that signed a top-down contract with an AI startup, converted the pilot, and went live. By that bar, 29 percent of the Fortune 500 and roughly 19 percent of the Global 2000 were paying customers within about three years of ChatGPT's launch, a penetration speed enterprise software has never seen; historically, startups needed years before their first Fortune 500 logo. Demand concentrates where value is legible: coding leads by nearly an order of magnitude, with customer support and search behind it, and adoption is no longer confined to tech-forward sectors.

Put the two studies side by side and the contradiction evaporates. MIT NANDA asked, in effect: of the custom GenAI pilots running in early 2025, how many had already produced measurable P&L impact? Answer: few, on a flawed sample, in a window arguably too short for P&L impact to exist. a16z asked: how many of the world's largest enterprises are paying real money for AI products in production? Answer: a remarkable share, more every quarter. A pilot that has not yet reached the P&L and a signed production contract are different units, from different denominators, at different moments of a steep curve. The fight was never about the facts. It was about which question you wanted answered.

05

4. What "adoption" actually measures

McKinsey's State of AI survey, published November 2025 with nearly 2,000 respondents in 105 countries, is the cleanest single picture of the gap. Read it as a funnel, and read the funnel as a diagnosis.

Eighty-eight percent of organisations now use AI in at least one business function. About a third have begun scaling beyond pilots; only 7 percent have scaled fully across the enterprise. Thirty-nine percent report any EBIT effect at all, and most of those put it below 5 percent of earnings. The group McKinsey calls high performers, the ones attributing a material share of EBIT to AI and treating it as a growth engine rather than a cost tool, is 6 percent of the sample. That last number is the honest size of the winners' circle in late 2025.

The word "adoption" is doing heavy lifting in every survey that quotes the 88. Answering yes to "do you use AI somewhere" costs nothing: one team on Copilot qualifies. The funnel's lower stages measure progressively more expensive truths: redesigned workflows, retired processes, measurable earnings. The gap between the first bar and the last is not a technology gap. Individual usage keeps climbing on its own, with Gallup finding frequent AI use among US workers nearly doubling in two years, while the organisational numbers crawl. People adopt in weeks. Organisations adopt in years. The funnel is a picture of that mismatch.

One more McKinsey finding belongs here because it predicts the rest of this dossier: the high performers are three times more likely to have senior leaders who personally own the AI agenda, and they set growth ambitions for it rather than cost-cutting targets. The profile of the 6 percent is not "bought better software." It is "run by people who redesigned the work."

06

5. Zoom out: most of humanity has not even started

Inside the tech bubble, AI feels saturated and the race feels lost. One zoom-out corrects the vertigo: 84 percent of the world's population has never used AI at all.

The breakdown that circulated in February 2026, built on Microsoft and DataReportal data, puts roughly 6.8 billion people at zero AI usage, about 1.3 billion on free chatbots, around 25 million paying for an AI service, and a few million using AI in advanced ways, agents and code. Treat the exact figures as directional; the shape is what matters, and the shape is a pyramid with an extremely narrow top. The paying user base of the defining technology of the decade is roughly the population of Australia.

For a business audience this chart cuts two ways. It kills the excuse of lateness: the curve has barely left the ground, and the perception that "everyone is already doing this" is an artifact of who you follow online. And it sets up the more uncomfortable reading of section 9: even inside organisations, the real usage sits in the free-chatbot layer of the pyramid, broad, shallow and unmanaged. Depth, the top of the pyramid, is where the 6 percent live. Both readings say the same thing: the game is early, and the advantage goes to whoever builds depth first.

07

6. The anatomy of a dead pilot

Strip the numbers away and ask the mechanical question: what actually kills an AI pilot? The research keeps returning four causes, and none of them is model quality.

Cause one: the tool cannot learn

MIT NANDA's most defensible finding: organisations buy or build static tools that demo brilliantly and then never improve. No memory of context, no feedback loop, no fit with the daily workflow. The winners in the same sample deployed systems that learn from use, plug into real workflows and get better over time. A pilot without a learning loop has a shelf life of exactly one wow-moment.

Cause two: the organisation does not know its own process

Ethan Mollick's favourite framing borrows the garbage can model from organisational theory: companies are semi-chaotic containers where problems, solutions and decision-makers collide, and nobody fully knows how work actually happens. His sharpest anecdote comes from researcher Ruthanne Huising: teams assigned to map their own company's processes routinely discover that no one, at any level, knows the whole picture. Now hand that organisation an automation project that requires the process to be specified. The pilot dies in the mapping phase.

Cause three: congestion

Exponential View's May 2026 analysis names the subtler killer. AI speeds up individual output: developers ship 50 percent more, sales drafts proposals overnight. All of it then queues at the same human decision gates as before, review cycles, sign-offs, legal. Output accelerates, throughput does not, and the gains evaporate in the waiting room. Adding more AI to a blocked decision pipeline makes the blockage worse.

Cause four: the productivity paradox at the desk

An Upwork study found 47 percent of AI-using workers unsure how to achieve the gains their employer expects, and 77 percent saying the tools had added to their workload in at least one way. Tools handed down without redesigned expectations produce visible activity and invisible drag. The pilot "succeeds" as usage and fails as economics.

Four causes, one theme: the model was never the constraint. Every failure mode above is a property of the organisation, which is why buying a better model fixes none of them, and why the 95-percent-style headlines keep misleading. The failure rate measures organisational readiness while being quoted as a verdict on the technology.

08

7. What the six percent do differently

The winners' playbook is remarkably consistent across McKinsey's high performers, Mollick's field work and the case studies: redesign the work, own it at the top, and connect the tinkerers to the strategy.

Mollick's operational frame is Leadership, Lab and Crowd. Leadership means executives who paint a concrete picture of how work changes and put their own name on it, which McKinsey independently finds triples the odds of scaling. The Lab is a cross-functional team that turns promising hacks into tested, benchmarked workflows. The Crowd is the workforce already experimenting, roughly 40 percent of employees by Mollick's reading of the surveys, mostly in silence. Each leg fails alone: leadership without a lab is a memo, a lab without the crowd is a silo, a crowd without leadership stays underground. The 6 percent, in various vocabularies, run all three.

The best recent case anatomy comes from Wharton's Generative AI Labs, which interviewed twenty game studios, an industry where AI pressure arrived early and hard. The studios sorted onto a four-stage ladder: secure chatbots for everyone, then top-down workflow pilots, then the interesting stage where individuals use AI to cross team boundaries and context starts compounding in shared documents, and finally the AI-first studio, small generalist teams organised around outcomes. Two details deserve a highlight. Stage two, the classic pilot, is where progress stalls on tacit knowledge and resistance, exactly the garbage-can prediction. And only three of twenty studios reached stage four. All three were built around AI from their founding. Nobody in the sample rebuilt an existing organisation to stage four; the ones that got there started there.

That last fact is the uncomfortable one for every established company, and it sets the bar honestly: the question is whether an incumbent can do deliberately what those three studios did by birth. The Monday-morning section returns to it.

09

8. The agent complication

2026 is the year agents, AI that executes multi-step work rather than answering questions, arrived in the enterprise. They raise the ceiling on the prize and the price of the same old organisational gaps.

The numbers first. McKinsey found 62 percent of organisations at least experimenting with agents by late 2025, with 23 percent scaling them, concentrated in IT and R&D. Deloitte's 2026 State of AI survey adds the warning label: close to three quarters of companies plan to deploy agentic AI within two years, while 21 percent report having a mature governance model for autonomous systems. An agent is an employee-shaped piece of software with no manager. Most organisations are hiring thousands of them with no HR department.

The field reporting is more instructive than the surveys. Box CEO Aaron Levie, after a road tour of enterprise IT leaders in April 2026, described the shift from "let a thousand flowers bloom" chat deployments to targeted automation of specific workflows, and catalogued what actually occupies the adopters: change management above all; "tokenmaxxing," the very real budget fights over compute, with one company running a shark-tank for token allocations; decades of legacy systems that agents cannot reach; and a firm consensus that the use cases are new capacity rather than headcount removal, work the company could never prioritise before. One company now has a head of AI in every business unit reporting to a central team. And his meta-observation: agents made the work more technical, Skills, MCP servers, command lines, so diffusion needs engineers, and everyone is working more than ever.

"Most companies are not talking about replacing jobs due to agents. The major use-cases are things the company wasn't able to do before. More emphasis on ways to make money vs. cut costs." — Aaron Levie, field notes from enterprise visits, April 2026

Mollick adds a genuinely open strategic question in "The Bitter Lesson versus The Garbage Can": will agents need organisations to understand their own processes, or will they train on outcomes and route around the chaos? If the second, the garbage can stops being a moat for incumbents who at least knew their own mess. Watch this one; it decides how much of section 6 survives the decade. Workforce effects we deliberately park here: McKinsey records 30 percent of firms expecting reductions, and that question gets its own dossier in this series, on entry-level work.

10

9. Shadow AI: the crowd already adopted

While the official pilot waited for the steering committee, the workforce went ahead. The most reliable adoption in most companies is the one nobody approved.

The same MIT NANDA research that produced the 95 percent headline contained a finding that aged far better: workers at over 90 percent of companies were regularly using personal AI tools for work, while only about 40 percent of companies had bought an official LLM subscription. Netskope's cloud analytics, measured rather than surveyed, put 47 percent of workplace generative AI usage on personal accounts through 2025. And multiple surveys find a majority of AI-using employees, 59 percent in one, hiding their usage from their managers, fearing judgment, policy trouble, or being handed more work for the same pay.

Security teams read those numbers as an incident report, and they have a point: unmanaged accounts, unknown data flows, no audit trail. Strategically, though, the shadow economy is the best free asset an organisation has. It is a completed, self-funded pilot program. It shows exactly where AI already works, run by exactly the people who found it worth the career risk. The employees hiding their usage are the Crowd from section 7, driven underground by fuzzy policy and misaligned incentives: the observed reward for showing your AI workflow is often more work at the same pay, or an awkward conversation about whether your job is needed at all.

The winners flip the incentive. Amnesty for past use, explicit permission with clear data rules, and visible rewards for shared workflows: an #ai-wins channel, office hours run by the power users, recognition that treats a discovered workflow as a contribution rather than a confession. The cheapest AI strategy available to any company this year is to surface the adoption it already paid for.

11

10. Efficiency theatre and the decision loop

The subtlest failure mode is succeeding at the wrong metric: visible busyness, faster drafts, impressive usage dashboards, and a P&L that never moves. History has a name for the fix, and it comes from electricity.

Exponential View's framework, built on the classic study of factory electrification, gives the gap its cleanest structure. Factories first replaced steam engines with electric motors and saw little gain, running the new power through the old layout of shafts and belts. The transformation arrived only when engineers gave every machine its own motor and redesigned the floor around workflow. AI is repeating the pattern in three stages: the lightbulb stage, workers using chatbots for personal productivity; the group-drive stage, agents accelerating existing workflows; and the unit-drive stage, where the decision loop itself, observe, evaluate, act, is rebuilt around the machine and weeks compress into hours. Most firms are at stage one. Only 27 percent of executives say AI has met ROI expectations, which is what you would expect when new power runs through old shafts.

The cautionary tale for skipping the redesign and simply cutting is Klarna. In early 2024 the fintech announced its AI assistant was doing the work of 700 customer-service agents and froze hiring. By May 2025, CEO Sebastian Siemiatkowski conceded the company "went too far": cost had dominated the decision, quality dropped, complaints rose, and Klarna began hiring humans back into a hybrid model, AI handling routine volume, people handling complexity, emotion and judgment. The lesson is precise. Klarna automated the existing workflow to save money instead of redesigning the service around what each side does best, group drive mistaken for unit drive, and the P&L punished the difference. The redesigned hybrid, notably, kept most of the automation.

12

11. Europe and the Netherlands

Europe's version of the gap comes with better news than the continent's AI reputation suggests, and a familiar warning inside it.

The number that defined European AI pessimism was Draghi's: 13.5 percent of EU enterprises had adopted AI as of 2024, cited across the competitiveness debate as proof the continent was missing the wave. The 2025 Eurostat measurement moved: 20.0 percent of EU enterprises with ten or more employees now use AI, a jump of 6.5 points in one year, the fastest gain since measurement began. The Netherlands sits fifth at 33.2 percent, behind Denmark at 42.0, Finland at 37.8, Sweden at 35.0 and Belgium at 34.5. A third of Dutch enterprises formally use AI; among large European enterprises the figure is 55 percent, against 17 percent of small firms.

Now apply this dossier's funnel logic to those numbers, because "uses AI" is the 88-percent-style top of the funnel, measured more strictly. If the McKinsey ratios hold even loosely, the share of Dutch firms with real earnings impact is a low single-digit percentage, and the share of European firms is lower still. The European gap is therefore double: fewer firms in the funnel at all, and the same organisational drop-off inside it. The honest strategic read for a Dutch business mirrors the robotics dossier's conclusion about hardware: the models are American, increasingly the applications are too, and waiting for a European stack is a strategy for buying both later at retail. The counter-move is also the same: adoption speed and workflow depth are sovereignty you can build in-house, on any model, starting now. A Dutch firm at stage three of the electricity ladder beats an American competitor at stage one, whatever flag the model flies.

13

12. The gap is organisational. The clock is exponential.

Here is the fact that turns the adoption gap from an interesting survey result into a strategic emergency: while organisations stood still, the technology got two orders of magnitude cheaper.

The chart above is borrowed from this series' exponentials dossier, where the curve family is treated in full. The single line to retain: GPT-4-class output fell from roughly 20 dollars to roughly 7 cents per million tokens in three years, about a tenfold cost drop per year, while capability rose. Every organisational failure mode in section 6 was diagnosed against last year's price and last year's capability. The pilots that died in 2024 were attempted with tools ten to a hundred times worse per euro than what is available as you read this. Menlo Ventures measured enterprise AI spend growing 3.2x in 2025 alone; the buyers are not waiting for the debate to resolve.

This is also why the adoption gap does not close by itself, and why "we tried it and it didn't work" is the most expensive sentence in corporate AI. The firms in PwC's 2025 jobs barometer data tell the compounding story from the labour side: since 2022, revenue per employee in the industries most exposed to AI grew 27 percent, three times the 9 percent in the least exposed, and AI-skilled roles carry a 56 percent wage premium. The gap between the 6 percent and the rest is not static. It compounds at the rate of the underlying curve, which is the fastest cost-collapse in the history of general-purpose technology. Organisational change runs on calendar time; the technology runs on exponential time. Every quarter of hesitation is priced at the difference.

14

13. What we would do on Monday

Our read, and our practice: treat the gap as the opportunity, because the hard part is organisational and the organisational part is learnable. Here is the playbook we run with clients, mapped onto the evidence in this dossier.

First, see it. Nothing moves an organisation like watching its own work done differently. Skip the abstract awareness deck; demonstrate a live workflow from your own company, rebuilt, in front of the people who run it. This is the Leadership leg: a vivid, specific picture, owned by name at the top, which McKinsey's data says triples the odds of scaling. Deadlines help more than visions. One consultant we rate opened a workshop by replacing a client's "everyone tries one AI task by July" mandate with "by end of day," and got compliance by dinner.

Second, understand it. Before automating anything, find out what your organisation actually does, and accept that the org chart is fiction. Surface the shadow economy with amnesty and rewards: the 90-percent statistic from section 9 says your best pilot results already exist, unlabelled, on personal accounts. Stand up a small Lab, cross-functional, senior enough to retire a process, and let it benchmark the crowd's discoveries against real workflows. This is also where the honesty gates live: define what the machine may decide alone, what a human signs, and what stays human, before scale, not after the first incident.

Third, adopt it, by workflow, not by tool. Pick decision loops where congestion visibly bleeds money, quote-to-contract, applicant-to-interview, brief-to-published, and rebuild the loop end to end, including the approval gates, with a learning system rather than a static tool. Measure EBIT-adjacent outcomes, cycle time, cost per unit, revenue per employee, never "usage." Kill pilots that only produce activity. Expect the third one to work.

We run this on ourselves, and publish the results. Our own content operation went through exactly this arc, individual tools, then a redesigned pipeline with human approval gates, then autonomous loops with honest metrics, and the numbers are public in this series' pipeline dossier: 694,527 impressions in a half year, up 429 percent, at a marginal cost near zero, run by a team whose evenings got quieter rather than busier. Not because the tools were special. Because the workflow was redesigned around them. That is the whole trick, and it is available to any organisation willing to do the unglamorous middle part.

15

14. What we would watch

Our position: the adoption gap is real, the failure narrative built on it is not, and the next 24 months decide who compounds. Four questions will tell you most of what matters.

Watch the 6 percent. If McKinsey's high-performer share climbs toward 15 or 20 percent by 2027, the playbook is diffusing and the window for easy differentiation is closing. If it stays pinned at 6 while usage saturates, the organisational barrier is harder than argued here, and the AI-first insurgents from the Wharton sample become the main threat instead. Watch the agent governance ratio: three quarters of firms deploying agents against one fifth governing them is a gap that produces incidents, and the first spectacular one will set regulation and boardroom mood for years. Watch the EU catch-up rate: 13.5 to 20 percent in a year is a real slope; two more years of it and the European pessimism trade gets crowded. And watch your own metrics drawer: the moment a company measures AI in EBIT terms rather than usage terms, it has left efficiency theatre. That one you can check on yourself this week.

The last word goes to the number that started it all. The 95 percent was wrong in the details and right in the warning: most organisations are indeed not yet profiting from AI. Where the headline said "the technology fails," the evidence says "the organisation decides." That is a harder message and a far more hopeful one, because it puts the outcome back in the hands of the people reading dossiers like this on a Monday morning.

16

15. Verification and sources

This dossier draws on live web research and a personal archive of more than 15,000 sources. Load-bearing numbers were cross-checked against primary reports where available. The notes below flag confidence and the main caveats, in the spirit of showing our work.

Charts labelled "BFF" are our own, drawn from the sources named beneath them. The Exponential View chart in section 10 is reproduced in a BFF frame with attribution. Framed quotations are verbatim from the cited sources.