The Two Adoptions: Why Enterprise AI Is Succeeding and Failing at the Same Time

Here are two facts about AI in the enterprise, both well sourced, both from the last twelve months.

Fact one: generative AI is the fastest-adopted workplace technology ever measured. Economists at the St. Louis Fed found that roughly 40% of American adults were using it within two years of ChatGPT's launch.¹ The personal computer needed about three years to reach half that. The internet needed about two years to reach half that. Nothing in the history of workplace technology has moved this fast.

Fact two: enterprise AI returns are flat. McKinsey's latest State of AI survey, published this week, found that only 37% of organizations attribute any earnings impact to AI at all, unchanged from a year ago, and that the share of true high performers, companies attributing 5% or more of EBIT to AI, has been stuck at 6% for two straight years.² This while investment climbed the whole time.

The standard move is to call this a paradox, or a bubble, or proof that the technology is oversold. A famous MIT-affiliated study put a number on the gloom last year: 95% of enterprise generative AI pilots showed no measurable P&L impact.³ That number went viral, briefly rattled the stock market, and became the one thing everyone knows about enterprise AI.

I think the paradox dissolves the moment you stop treating "enterprise AI adoption" as one thing. It is not one thing. It is two completely different adoptions that happen to share a name. One is the adoption of AI by individuals who work at enterprises. The other is the adoption of AI by the enterprise itself: its processes, its measurements, its org chart. The first is one of the great technology success stories of our time. The second has barely started. Almost every confusing statistic, every contradictory survey, and every failed pilot postmortem becomes legible once you hold these two adoptions apart.

The individual adoption curve looks like a consumer internet product, because that is what it is. People did not wait for procurement. The same MIT study that produced the 95% headline also produced a far more interesting number that almost nobody quoted: while only 40% of the companies studied had bought an official LLM subscription, employees at 90% of them were using AI tools regularly anyway, mostly through personal accounts.³ The researchers called it the shadow AI economy. The shadow economy was more than twice the size of the official one, inside the same buildings, during the same fiscal quarters that produced the 95% failure number.

The firm adoption curve looks like every enterprise technology curve before it, which is to say slow, shallow, and concentrated. The Census Bureau's business survey, the only large representative sample of actual American firms, puts AI use at roughly 20% of businesses in mid-2026, around 37% for large firms.⁸ And the adoption that exists is thin: a Census working paper this spring found that a majority of adopting firms use AI in three or fewer business functions, and two thirds restrict workers to a handful of tasks.¹⁰ Corporate card data from Ramp makes the same point in dollars. The median business on their platform spends about $12 per employee on AI. The top 1% spends over $7,400.¹⁴ Firm adoption is not a wave. It is a power law with a very long, very flat tail.

So when a headline says AI adoption is exploding, it is describing the first curve. When a headline says AI is failing in the enterprise, it is describing the second. Both headlines are correct. They are just not about the same event.

The enterprise is not adopting AI. The people inside it are. Those are different events, and only one of them shows up on a dashboard.

What makes this framing more than wordplay is that the gap between the two adoptions, the wedge, is now directly measurable, and at least five independent research programs have measured it without coordinating with each other.

The survey wedge. In McKinsey's 2026 survey, 80% of respondents who use AI say it has improved their individual productivity.² The same survey shows firm-level earnings impact flat. The gap between "it helps me" and "it helps the company" is 74 points wide, inside a single questionnaire.

The tax-record wedge. The cleanest study yet comes from Denmark, where economists Anders Humlum and Emilie Vestergaard tracked 25,000 workers across 7,000 workplaces and matched their AI usage to actual tax records. Workers reported real time savings, averaging about 3% of work hours. But across all eleven occupations studied, AI chatbots produced no significant effect on earnings or recorded hours, and the researchers estimate only 3% to 7% of the productivity gains passed through to workers' pay.⁴ The time savings were real. They just evaporated somewhere between the individual and the payroll system.

The time-leak wedge. BCG's 2026 workplace survey of roughly 12,000 employees found where the evaporation happens. Among frontline AI users, 42% now save eight or more hours a week, a stunning number. But two thirds of them receive little or no guidance about what to do with the freed time, and more than half do not reinvest it in higher-value work.⁵ The hours are saved by the individual and then absorbed as invisible slack. Nobody redesigned the job around the fact that the job now takes one day less.

The seat wedge. Microsoft has sold roughly 30M paid Copilot seats, a big number that also represents well under a tenth of the Microsoft 365 commercial base after three years of the most aggressive enterprise software push in history.¹⁷ Independent analyst estimates put weekly active usage at a fraction of licensed seats. Enterprises bought the official adoption. Their employees, in many cases, kept using whatever they were already using. A seat license is a record of a procurement decision, not of a behavior change.

The telemetry wedge. Network security firms can see which accounts workplace AI traffic actually flows through, and Netskope's longitudinal data is the single best series we have. In early 2024, 82% of workplace generative AI use went through personal accounts. A year later, 72%. In the 2026 report, 47%.⁶ Even now, after three years of official rollouts, nearly half of observed workplace AI usage still routes around the enterprise entirely. And the human layer of this is stranger still: in a 48,000-person global study by KPMG and the University of Melbourne, 57% of employees who use AI said they hide it and present the output as their own.⁷ Slack's workforce research found nearly half of desk workers uncomfortable telling their manager they use AI at all.¹⁸ The most successful workplace technology in history is partly a secret.

Five methods. Surveys, tax records, time-use studies, license data, network telemetry. One shape: enormous individual adoption, thin firm capture, and a wedge in between where the gains go to die, or at least to hide.

Here is the uncomfortable question the wedge forces: if individuals are saving all this time and the firm is not getting it, who is?

The individuals are. This is the part of the story that I think is genuinely new, and genuinely under-discussed. In previous enterprise technology waves, the company bought the tool, the company deployed the tool, and the company captured whatever surplus the tool created. With AI, the tool arrived through the employee's personal browser tab, fluent in the employee's own working style, at a price the employee could pay out of pocket. So the surplus lands with the employee first, in the currency employees actually value: time, slack, reduced drudgery, the ability to look more competent with less effort.

Once you see it this way, the hiding behavior stops looking irrational and starts looking like straightforward economics. An employee who discloses that AI does 30% of their job is volunteering to have that 30% repriced. The KPMG study's 57% who pass off AI work as their own are not being deceptive so much as protecting their share of a surplus that no one has negotiated over yet. The Danish data quantifies the current split: workers are keeping roughly 93 to 97% of the gains, in the form of time the firm cannot see.⁴

This also explains why the famous failure statistics coexist so peacefully with obvious, visible usefulness. A pilot program is an attempt by the firm to capture value in the firm's own instrumentation: P&L, tickets closed, cycle time. Most pilots fail to register there, and the 95% study, whatever its methodological problems, is measuring something real about that. But the individual adoption never needed the pilot. It was already producing value the instrumentation cannot see, because the instrumentation was built to measure the firm, not the person.

If the individual adoption is real, why can't firms just collect the gains? Because collecting them requires the firm to change shape, and changing shape is the slowest thing large organizations do. We know this because it has happened before, in almost exactly this sequence.

When factories first electrified, they did it wrong for decades. Paul David's classic study of the dynamo tells the story: twenty years after Edison's first central power stations, electricity drove less than 5% of factory machinery. Factories that did adopt electricity mostly used the big electric motor as a drop-in replacement for the steam engine, spinning the same central line shafts and leather belts that had defined factory architecture for a century. The gains were marginal, because the technology was new but the factory was old.

The productivity revolution only arrived when engineers stopped replacing the steam engine and started replacing the factory. Small motors on each machine, "unit drive," meant the building no longer had to be organized around a shaft. Factories went single-story, laid out around the flow of materials instead of the flow of power. By 1919, half of factory mechanical drive was electrified, and across the 1920s American manufacturing productivity grew at roughly 5% a year, more than triple its rate in the fifteen years before the first world war.¹² Forty years separated the light bulb from the payoff, and the missing ingredient was never the technology. It was the rebuild.

The modern version of this finding has numbers attached. Erik Brynjolfsson and his collaborators established two decades ago that every dollar of computer capital in a firm was associated with roughly $9 more in intangible organizational capital: redesigned processes, training, new decision rights.¹³ The technology was the tenth of the iceberg above the water. Nobody has budgeted the $9 for AI. The industry poured on the order of $700B into AI infrastructure this year, and I am aware of no company that reports, or even estimates, its complementary investment in reorganizing work around what the models can do. A Federal Reserve note this July said it plainly: the evidence so far is consistent with a buildout phase, and the micro-level gains are not yet visible in aggregate productivity.⁹

The most direct evidence comes from new Census microdata on American manufacturers, which found that AI adopters initially experience a measurable drop in total factor productivity before any gains arrive, and that the drop is worst at older establishments, where existing structured management practices decay before new ones form.¹¹ The J-curve that Brynjolfsson theorized is now visible in the data: things get worse before they get better, and they get worse longest at the firms with the most accumulated process to unlearn. Which is exactly what a rebuild looks like from the inside.

McKinsey's survey, for all its flat top-line numbers, contains the confirming detail: among the 6% of high performers, 73% fundamentally redesigned workflows around AI. Among everyone else, 25%.² Workflow redesign is the single strongest statistical differentiator between the companies where AI shows up in earnings and the companies where it shows up only in employee browser tabs. The firms that treated AI as a drop-in steam-engine replacement got steam-engine results.

You do not have to take my word for the two-adoptions framing, because the companies selling AI are already pricing it in. Watch what they do, not what they say.

Start with pricing. Salesforce has cycled through three different pricing models for its agent platform in well under two years: per conversation, then consumption credits, then unmetered per-user deals.¹⁹ Zendesk now bills per ticket its AI actually resolves rather than per agent seat.¹⁶ Intercom charges 99 cents per resolution. The per-seat model, the foundational unit of enterprise software economics for thirty years, is being abandoned in real time. A seat is a bet that a human will sit in it and use the tool. Vendors have seen enough seat-usage data to stop making that bet, and outcome pricing is what it looks like when the seller concedes that only realized work, not licensed capacity, will get paid for. The pricing model is a confession.

Then look at where the AI labs are hiring. I wrote in The Deployment Epoch about OpenAI and Anthropic building forward-deployed engineering and services arms, and the trend has only steepened since: multi-million-dollar deployment engagements, dedicated services subsidiaries, armies of consultants trained on specific models. Accenture booked nearly $6B of generative AI work last fiscal year.²⁰ None of this would exist if enterprise adoption were a software problem. You do not need forward-deployed engineers to install software. You need them to do the organizational surgery that the software cannot do to its host: the workflow redesign, the process mapping, the intangible $9. The entire services boom is the market pricing the wedge.

Even the cautionary tales fit the frame. Gartner expects a large share of agentic AI projects to be cancelled within the next year and a half, for costs and unclear business value.¹⁵ Of course they will be. An agent dropped into an unredesigned workflow is a line-shaft factory with a very large electric motor bolted where the steam engine used to be. The motor is not the problem.

If the two-adoptions framing is right, it implies things about what happens next. These are directional calls, not dated ones, but they are falsifiable in shape: each one names the signal that would prove it wrong.

• The wedge closes from below, not above. Firm adoption will not be achieved by transformation programs pushing AI down the org chart. It will be achieved by organizations reshaping themselves around people who are already fluent, the way companies eventually stopped having an "internet strategy" because everyone under a certain seniority simply worked that way. Hiring, promotion, and team design become the real adoption mechanisms. If, five years from now, the companies with the best AI results got there mainly through top-down enterprise programs, this call was wrong.
• The seat dies as the unit of account. Per-seat AI pricing and seat-count adoption metrics will go the way of website hits: an artifact of measuring the wrong thing because it was the easy thing. Spending shifts toward consumption and outcomes, and "how many licenses" stops appearing in earnings calls, replaced by resolution rates, cycle times, and cost per completed task. The early signs are already everywhere in vendor pricing.
• The J-curve resolves unevenly, and the split shows up in margins before it shows up in GDP. Young firms and AI-native challengers, with less accumulated process to unlearn, will cross to the steep side of the curve first, and the gap will appear as diverging margins within industries well before national productivity statistics move. Electrification's payoff arrived only after the new architecture passed the halfway mark; expect the aggregate AI numbers to stay murky until workflow redesign, not tool deployment, crosses a similar threshold. If aggregate productivity surges while adoption remains a thin layer over unredesigned processes, I will have been too pessimistic about the sequencing.
• The saved time becomes the negotiation. Right now employees hold the surplus, invisibly, and firms are starting to notice. The next phase of enterprise AI is not technical but political: who owns the recovered hours. Companies that treat disclosed AI gains as capacity to confiscate will push usage back underground and re-widen the wedge; companies that let people convert some of the surplus into visibly better work, and share the upside, will be the ones whose shadow adoption converts into firm adoption. The KPMG hiding number, 57% today, is the metric to watch. It falls where trust is real, and it will stay stubbornly high everywhere else.

Step back far enough and the story stops being about AI at all. A general-purpose technology arrived, individuals grabbed it at historic speed, and institutions are doing what institutions do: lagging, measuring the lag badly, and mistaking their own reflection for the technology's failure. The dynamo went through this. The computer went through this. Robert Solow's famous line, that you could see the computer age everywhere but in the productivity statistics, was true when he said it, and then, about eight years later, it wasn't.

So the next time a study announces that enterprise AI is failing, it is worth asking which adoption it measured. The individual adoption is not failing. It is arguably the most successful technology diffusion ever recorded, and it happened without permission, without training budgets, and half in secret. The firm adoption is not failing either. It has barely been attempted, because attempting it means rebuilding the factory, and almost everyone is still bolting the motor to the line shaft.

The gap between those two adoptions is where the next decade of enterprise value sits. It will not be closed by a better model, and it will not be closed by another pilot. It will be closed the way it was closed the last two times: slowly, structurally, and by the organizations willing to change shape. Watch the org charts, not the model releases.