The value leaks at the top

Forty-two percent of people using AI at work now save at least a full day a week. Two-thirds of them are never told what to do with that time. The second number is the real story of the past seven days.

The thread this week

Five different studies landed on my desk this week. Different institutions, different methods, different countries. They were all describing the same hole.

PwC asked 1,217 executives across 25 sectors how much value they actually get from AI. Three quarters of it is going to one in five firms, and that top fifth is 2.6 times more likely to have changed how the business runs rather than how much it trains. BCG put a ratio on the same idea: 10% of an AI initiative's value comes from the algorithm, 20% from data and technology, 70% from people and process. Adrienne Down Coulson, COO of Rakuten International, named the blockage in Fortune. On the ground AI already works, she wrote, and then the work reaches a leadership team that refuses to operate any differently than it did five years ago.

BCG's AI at Work report, 11,749 workers across 14 markets, priced the leak. A clear strategy plus redesigned workflows adds about 25 points of measurable business impact. Better tools on their own move it about 5. And an NBER paper that surveyed roughly 750 corporate executives found the cleanest version of the problem: the productivity gains people feel are bigger than the gains anyone can put on the books.

Read those together and they stop being five findings. They become one. The productivity is individual and real. The value is organisational and mostly missing. What sits between the two is design, and almost nobody is funding it.

Going deeper

Here is the number that should end the "we need more training" conversation. MIT's NANDA initiative looked at 300 public AI deployments, interviewed 150 leaders, surveyed 350 staff, and found that 95% of enterprise generative-AI pilots produce no measurable return. Thirty to forty billion dollars spent, and the P&L does not move. Their reading is blunt: the wall is not model quality, it is the gap between a capable tool and an organisation that has not changed its work to use it.

Klarna is the example I keep using because it cuts the other way too. In 2024 the company replaced about 700 customer-service agents with an OpenAI-trained assistant and made it a headline. By 2025 it was hiring humans back, after satisfaction dropped and the bot kept failing on anything that needed judgement. CEO Sebastian Siemiatkowski admitted they had focused too much on efficiency and cost, and the result was lower quality. Klarna did redesign the work. They just redesigned it for substitution, swapping a person for a bot inside the same old process. Capability building asks a harder question than "who can we remove." It asks what the work is now for.

The honest objection to all of this is harder than it looks. I keep calling the individual productivity "real." What if it is not? METR ran a randomised trial with experienced open-source developers and found they were 19% slower when allowed to use AI, while believing they were 20% faster. If the felt gain can point the opposite way from the measured one at a single desk, maybe there is no leaked value to catch at all.

I think that finding sharpens the argument instead of breaking it. If perception and measurement can diverge that violently for one coder, then trusting the felt gain at the scale of a company is reckless. The response is the same either way: redesign the work, measure the output, and stop rewarding the feeling of speed. METR studied solo work on codebases the developers already knew well, which is the narrow case. The organisational argument holds because it is about restructuring decisions and roles, with a number attached.

What this means for you

If you run L&D, HR or transformation, two moves are worth more than your next licence renewal.

First, stop funding the saved hour and start funding the catch. Pick two or three workflows where people already save time, and put one named person in charge of redesigning the roles and handoffs around that time, with a result you can check in 90 days. The saved hour is free. The redesign that converts it is the job.

Second, aim part of the budget upward. The Rakuten COO is describing a constraint that lives in the executive team, in decision rights and in who can see what their people already do. A workshop for the frontline does not touch it. Your hardest work this year is getting the people at the top to work differently, and that is a brief almost no L&D function has written for itself yet.

The provocation

Everyone you work with can name a tool that made them faster. Almost none of them can name the workflow that got rebuilt to keep the gain. This week, find one process where the saved time is leaking, and put someone's name against fixing it before Friday. If you cannot name that person by then, you have your answer about where your AI value is going. So who owns the redesign in your organisation, and what happens if the honest answer is no one?