Article · 24 August 2026
Ninety per cent of workers already use AI every day. Ninety-five per cent of organisations get nothing measurable back. Here are the foundations that turn six hundred private experiments into something that connects.
In July 2025, MIT's Project NANDA published a study called The GenAI Divide. It found that despite thirty to forty billion dollars of enterprise investment in generative AI, ninety-five per cent of organisations were getting zero measurable return, and only five per cent of integrated pilots were extracting real value. The sample was modest, fifty-two structured interviews, a survey of a hundred and fifty-three leaders and analysis of three hundred public deployments, so treat the precise figure with some care. The pattern it describes is harder to argue with.
Buried in the same study is the number that actually explains it. Around forty per cent of the organisations surveyed had an official large language model subscription. Ninety per cent of the workers surveyed said they used personal AI tools for work tasks every day.
That is not an adoption problem. That is the opposite of an adoption problem. The technology is already everywhere in the building. It simply is not connected to anything.
I wrote recently about AI landing on managers who have no capacity left to absorb it, and about Britain spending forty years with plenty of electricity and no electricity system. Several people asked the obvious follow-up. Fine, so what would connecting it actually look like?
The answer sits in the same history, and it turns on one detail that is easy to miss.
The Central Electricity Board, created in February 1927, did not generate a single unit of electricity. It did not own the stations. It selected which existing stations would generate for the board, it built and owned the transmission lines between them, and it set the standard everyone had to meet. Control without ownership. Wires and a standard, nothing else.
That is the inverse of what most organisations build when they announce an AI strategy. They create a central function that owns the tools, holds the budget, approves the use cases and becomes, within about six months, a queue. Sixty-two per cent of the chief HR officers in Gallup's recent survey said they were creating centres of excellence or internal AI champions. The intention is right. The design is usually upside down.
The board that built the National Grid did not own a single power station. It owned the wires and the standard.
Here is what the foundations look like if you build it the other way round.
Six hundred undertakings could not share power because they were not running on the same basis. Different voltages, different frequencies, some places still on direct current. No amount of goodwill or cable was going to fix that. Standardising frequency was the board's first and least glamorous job.
The organisational equivalent is a shared vocabulary, and almost nobody does it. Two departments cannot pool anything if an AI assisted output means one thing in finance and another in operations, if data classifications differ between teams, if checked by a human means a careful read in one place and a glance in another. This is dull work. It is also the work that determines whether anything else is possible.
The board did not raise all six hundred undertakings to a common standard first. It identified the most efficient stations, designated them as selected stations to generate for the grid, and let the rest go quiet. By 1935, twenty-eight power stations were supplying the majority of the country.
So find the three or four places in your organisation where AI use is already genuinely good. Not where it is most enthusiastic. Where the output is reliable and somebody can explain why. Name those as your generating stations and wire outward from them. This is the reverse of the usual approach, which is a training day for everyone and a proper pilot for nobody.
A connecting function should not approve, own or police what departments do with AI. It should carry the traffic between them. The moment it becomes an approvals queue, people route around it, and you are back to six hundred undertakings with an additional committee slowing everybody down.
This is a genuinely uncomfortable design for most senior teams, because it means giving up control of the thing while taking responsibility for the connections between the people doing it. The board was accused of exactly that at the time. It worked because the standard was mandatory and the traffic was free.
This is where these projects usually die. A shared prompt library is a generating station with no wires attached. It stores. It does not transmit. Nobody browses it after week three.
The test of whether the thing is alive is whether something actively pushes a discovery made in finance towards the operations team who never knew they had the same problem. Routing is a job, not a repository. If the design assumes people will come and look, the design has already failed.
Selected stations were paid for the power they fed into the grid. That is the mechanism that made a station willing to supply the country rather than hoard for its own borough. The board even accepted very low returns, and losses in the early years, in order to drive consumption up.
Only eight per cent of the organisations in Gallup's survey tie AI adoption to any performance incentive at all, and only twenty-six per cent embed AI skills into job expectations. So the department that spent three weeks working something out has every reason to keep the advantage and no reason at all to write it up. If contribution is not credited, your portal fills with the low value material people do not mind giving away.
The grid needed fault reporting far more than it needed success stories. Engineers had to know where the system failed, and under what conditions, or the next fault took the region down with it.
The single most useful thing one department can tell another is where the output looked completely convincing and was wrong. That information only moves if it is structurally separated from performance review. If admitting a failure can cost somebody, you will get a portal of triumphs, and a portal of triumphs is worse than nothing, because it teaches everybody that the whole thing is theatre.
The Central Electricity Board had a chairman, Andrew Duncan, and by 1934 it employed 1,248 people. It was not a steering group that met on the second Tuesday of the month.
If interconnection is added to the plate of managers who already have no room, it will not happen. That is the argument of the previous piece arriving here from a different direction. Somebody has to be given this as an actual job, with the hours and the authority that go with it, or it becomes another well intentioned initiative that people nod at and quietly ignore.
The grid did not switch on all at once. It began operating in 1933 as a set of interconnected regional grids, and only ran as a single national system in 1938. As each section was finished, that region rationalised its generation and standardised its frequency, so people felt the benefit without waiting for the whole project to complete. By early 1935, nine years after the board was founded, nearly ninety per cent of the grid was operational.
Connect two adjacent departments who share a real problem. Let them feel it. Then extend. Do not wait to launch an enterprise wide platform, because by the time it arrives the tools will have changed underneath it.
There is a lovely footnote to this, and it is the best argument you have for trusting the people on the floor. In October 1937, a group of engineers synchronised the regional grids overnight as an unauthorised experiment, and let power flow across the whole country for a single night just to see what would happen. It worked. The official national system followed. The interconnection was proved by the people doing the work, before the institution got round to sanctioning it.
The grid's key measure was not capacity. It was load factor, the proportion of the built capability actually in use. It had been stuck between 28 and 30 per cent through the 1920s, and reached 38.8 per cent by 1940.
Almost every organisation is currently measuring installed capacity. Licences issued, people trained, tools deployed, attendance at the AI lunch and learn. None of that tells you anything. The number worth tracking is what proportion of what one department learned is now in use somewhere else in the business. I have yet to meet an organisation measuring that, and it is the only figure that would tell you whether you have built a grid or six hundred undertakings with a logo on top.
Every one of these foundations is a decision about behaviour, incentive and permission rather than about software. Standardising a vocabulary. Choosing where to concentrate. Giving up control while keeping responsibility. Crediting contribution. Making failure safe to report. Being honest about what is actually in use rather than what has been bought.
None of that arrives in a procurement cycle. All of it has to be led, by people who have been given the room to do it.
The technology in your organisation is already good enough. It has been for a while. What is missing is the same thing Britain was missing in 1924, which is not power but the connections between the people who have it.
I write about the behaviour that decides whether change actually lands in an organisation. My first book, Rebuilding the Village, is out in early 2027 and looks at what happens to people when the structures around them stop holding.
Whether it's for you, your team or your organisation, the first conversation is confidential, informal and without obligation.
Get in touch