Companies who blindly jump on the hype train, and deploy AI solely through machines, fail to gain value from their investment. Others integrate at the root of a skilled, up to date workforce, and see increasing returns.

95%

Forty billion dollars in corporate AI spending has produced little measurable profit. The exception are the employees that knew how to use what they were given.

Ninety-five percent of company AI projects deliver no measurable profit. Thirty to forty billion dollars spent, almost nothing to show for it. The conclusion wrote itself: the bubble is popping.

It isn't, and the study that was quoted so numerously doesn't say that it is. Most people read the chart rather than the report, then passed the chart along. Which is a pity, because the report contains something far more interesting than its famous number.

Where the number comes from

The study came out of MIT in July 2025, from a research group working on AI systems. If you plan to quote it in a meeting, it helps to know what you're standing on.

It's thin. The report calls its own results preliminary, no independent researchers checked the work before it went out, and the original link on MIT's website no longer works. What circulates now are copies.

The sample is the real problem. The report says it drew on 52 interviews and 153 surveyed managers, while Fortune, writing about the very same report, says 150 interviews and 350 employees. When the sample size changes depending on who's describing it, something has gone wrong somewhere. A professor at Wharton asked the authors publicly to either publish their raw data or withdraw the paper, and one tech publication went through the report's own charts without finding the ninety-five percent figure in any of them. A year on, the data still hasn't been released and the paper still hasn't been withdrawn.

There's one more thing, and it rarely gets mentioned. The research group that wrote the report builds AI systems for companies, and the report concludes that companies need better AI systems. Nobody has to have acted in bad faith for that to matter, it's the kind of coincidence you'd spot instantly in a sales brochure, and it deserves the same attention here.

So the number isn't a measurement. It's a sketch of a real problem, and the good news is that the problem doesn't depend on it.

The problem is real anyway

S&P Global asked over a thousand companies in North America and Europe how their AI projects were going. In 2025, forty-two percent had abandoned most of them, up from seventeen percent a year earlier. On average, companies scrapped nearly half of their test projects before anything went live, and forty-six percent could not name a single business goal where AI had made a strong positive difference.

McKinsey surveyed almost two thousand companies worldwide and found the same thing from a different angle. Nearly nine in ten use AI somewhere in the business, but only four in ten can point to any effect on profit at all, and for most of those it's under five percent. Two-thirds haven't started rolling it out properly. And of everything McKinsey measured, the change most closely tied to actual profit was redesigning how work gets done: something about one company in five has tried.

Then Deloitte, in the largest survey of the three: more than three thousand two hundred business and technology leaders across twenty-four countries. Thirty-four percent of companies are using AI to genuinely change what they make or how they work. Thirty percent are rebuilding a few important processes around it. The remaining thirty-seven percent are using it on the surface, with nothing underneath actually different. Eighty-four percent have not redesigned a single job around what AI can now do.

Four surveys, four different methods, nobody with the same stake in the answer, same direction. And the numbers have not improved. The forty-two percent abandonment figure was still being cited as current in mid-2026, with nothing published since to replace it.

The part nobody quotes

Now the section of the MIT report that never made it onto anyone's slide.

While the official projects stalled, employees at more than nine out of ten of those companies were already using AI at work. Their own accounts, their own tools, nobody's permission. Only about forty percent of the companies had actually paid for it. The report calls this the shadow AI economy, and its example is a lawyer whose firm had bought expensive contract-review software. She kept using ChatGPT instead, because the thing her firm bought gave her rigid summaries she couldn't adjust.

The authors are clear about the cause. It isn't that the models are bad. It isn't regulation, and it isn't a shortage of engineers. People and organisations simply hadn't learned how to use what they had. More than half of AI budgets went into sales and marketing, where projects get noticed, while the steadier savings sat in office work, where nobody was looking.

Read this way, the report isn't a verdict on the technology but on how companies buy things. Where they failed, they had bought software. Where something actually worked, a person had taken a tool and pointed it at a job they already understood.

Access is not the same as use

Deloitte's survey contains the cleanest illustration of that anywhere in the research. Over a single year, companies expanded approved access to AI tools by fifty percent. An enormous procurement effort, executed successfully.

Among the workers who received access, fewer than sixty percent use it in their daily work. That share is essentially unchanged from the year before.

So the licences arrived and the behaviour didn't. Which is why Deloitte names the skills gap as the single biggest barrier to getting AI into a business: and why, when companies did adjust their talent strategy, the most common response was education rather than redesigning roles or workflows. They can see the problem. Most are answering it with training, which raises an obvious question about what kind of training actually works. More on that below.

The vanishing workday

In June 2026, BCG asked nearly twelve thousand employees across fourteen countries what AI had done to their week. Three-quarters of office workers now use it regularly, and four in ten of those save at least a full day every week.

Two-thirds get no guidance on what to do with that day.

More than half never put it toward anything more useful than what they were already doing, so the time is real, it genuinely gets saved, and then it disappears before anyone can point to it in a set of accounts. That's the whole puzzle, and it was never about whether people use the tools. In the same survey, having a clear plan for AI deployment through employee training raised the reported impact by twenty-five percentage points. Buying better tools raised it by five.

BCG has a rule of thumb from working with hundreds of companies: about ten percent of the value of an AI project comes from the AI itself, twenty percent from the technology around it, and seventy percent from people, training and process. Most budgets are built the other way round.

What we can and can't conclude

This is where these arguments usually overreach, so lets be careful.

BCG found a small group, roughly five percent of companies, actually making money from AI, and they do behave differently. They plan to train more than half their staff, where the companies seeing little value gains plan to train only a fifth. These 5% are four times more likely to run a real training program and to protect working time for it. Nearly nine in ten of their managers use AI visibly in their own work, against a quarter elsewhere. Their shareholders did about four times better over three years.

That last number is a correlation, not proof. It's just as possible that companies already doing well can afford to protect learning time in the first place, the data come from managers describing themselves, and the report was published by a consulting firm that sells exactly this kind of change.

Here's the version that holds up: across four independent surveys, the companies getting results differ from the rest far more in how they organise work and train people than in what technology they bought. Nobody has found a company that skipped that part and gained significantly from AI anyway.

What this means for the workforce

The story usually told to employees is that AI arrives and the junior job disappears. The evidence points somewhere more specific.

Companies aren't short of software, or models, or money: they have plenty. What they don't have enough of are people who can take a tool and a process and connect them, so the productive impact ends up somewhere a manager can see. The shadow AI economy already proves employees will adopt these tools with no help at all. What it doesn't produce is structure: a process someone wrote down, an answer someone checked, a saving someone measured, and the judgement to say that at this particular step, the model shouldn't be involved. Precise training with AI seems to be inevitable, and presents a great career opportunity for eager learners.

In most companies, nobody is doing that. Which makes the person who can explain where AI went into a process, how they set it up, how they checked the result and what it saved rare, and incrementaly valuable. Not just as a skill on a CV, but as the answer to a problem the company has already spent real money failing to solve. Now the technology that was said to replace jobs needs even more skilled workforce to be implementable.

How ironic...

The link has been copied!