The AI Buildout Nobody Is Counting
The next wave of AI adoption will not be funded by technology budgets. It will be paid for out of payroll.
There are two AI buildouts happening. Only one of them has a line item.
The first is the one everybody counts: datacentres, accelerators, power contracts—the capital expenditure of a handful of companies large enough to spend it. It is measured quarterly and discussed constantly.
The second has no name and no budget line, because it is not being paid for out of a technology budget at all. It is being paid for out of payroll.
Work was staffed because there was nothing else to staff it with
Inside every company, there is a body of work that is repeatable and structured: chasing a lead until it replies, reading an inbox and deciding what matters, pulling the same numbers into the same report every Monday, checking that a document says what the last one said.
None of this was staffed because a person was the ideal instrument for it. It was staffed because a person was the only instrument that could hold the context, exercise the small judgements the task required, and be accountable for the result.
Software could do parts of it—the deterministic parts. That is why the last thirty years of automation took the pieces that could be specified exactly and left the rest with a human.
The rest is most of it. That residue is where the payroll went.
What changed is narrower than it sounds
The interesting shift of the last two years is not simply that models became clever. It is that they became able to work unattended: to take a goal, use tools, and act over time without somebody sitting in front of them supplying context turn by turn.
That is a much smaller claim than “AI can do knowledge work”—and a much more consequential one.
It means the residue is addressable for the first time. Not because judgement suddenly became perfect, but because the system exercising it no longer needs a person sitting beside it.
When the cheaper option is also the better one, and the cost line in question is the largest one a company has, what follows is not merely a trend. It is an accounting event.
And yet almost nothing has moved
Most companies have bought AI and their work has barely changed. This is usually explained as a capability gap. It is not one.
Ask why a specific job at a specific company is still done by a person, and the answer is almost never that the model cannot do it.
The answer is that nobody has installed it.
A system that is going to take over a job—really take it over, rather than draft something a human then finishes—needs three things that no model ships with:
Context. Not “knowledge”, which is the model’s problem, but the specifics: how this firm quotes, which client is difficult, what last quarter’s exception was and why it was allowed. This lives in a company’s own data and its people’s heads. Getting it into a system is manual, particular work.
Permission. The authority to actually do the thing: write to the CRM, send the email, publish the page—with a boundary around what it may do unsupervised and a gate in front of whatever is irreversible.
Evidence. A record that lets somebody check what happened without redoing it. A company will not hand over a recurring job to a process it cannot audit, and it should not.
Every one of these is specific to one company. None can be bought off the shelf. That is why buying a licence to a capable model changes so little. The capability was never the scarce part.
The transfer happens one job at a time
This is the unglamorous centre of it. There is no platform purchase that moves a company’s repeatable work into systems. There is only a sequence:
- Pick one job.
- Encode it.
- Give it permissions and a gate.
- Prove it runs.
- Move to the next.
It is slow. It does not demo well. It resembles implementation more than it resembles software, and every technology cycle produces a wave of people who find that beneath them and try to sell the tool instead.
But the sequence is the product.
A job that has been encoded, permissioned and proven does not go back to a person. It becomes a fixed piece of that company’s operating machinery—and it costs a fraction of what the salaried version cost, forever.
Why the companies best placed to lead this cannot
The obvious objection is that the incumbents will take it. They have the distribution, the data and the relationships.
They also have a pricing model denominated in the exact thing this transfer removes. Almost every large software company charges per seat. Their revenue is a function of how many people a customer employs.
They can sell tools that make each of those people faster—that is their whole current AI motion, and it is a good one—but they cannot lead a transition whose end state is fewer seats.
This is not a prediction about execution. It is arithmetic about incentives.
The unit that comes out the other side
What replaces the salary is not a seat. It is a licence attached to a job.
That is a different commercial object. It is not priced against a person’s attention but against the cost of the work itself, which means it can be worth ten times a seat and still be obviously cheap.
It grows by taking over more jobs rather than by adding more people to the same tool. And it sits, in the customer’s own accounts, on a different line than the one it replaced: not headcount, but infrastructure.
Multiply that by every small company that has been quietly deciding not to make its next hire, and you get a buildout comparable to the one being counted—funded from a budget nobody currently classifies as technology spend.
What we are doing about it
We build Orchestra, the installation layer for exactly this. We take the operating routines inside a small company, encode them as systems with permissions and evidence, and run them afterwards.
Fifteen firms operate this way today. Every one of them paid, in advance, to be installed. We take that as the most useful fact we have, because it says the transfer is worth money to the people making it before it has proven anything.
The models are not the constraint. They have not been for a while.
The constraint is that somebody has to do this company by company, job by job—and that is a business, not a bottleneck.