The cheapest thing in your AI programme is the intelligence

In the summer of 2026 the strangest procurement pattern in enterprise software became impossible to ignore. The firms with the best models started paying billions to supply people. The sellers of capability have concluded that capability does not transfer itself.

In the summer of 2026 the strangest procurement pattern in enterprise software became impossible to ignore. The firms with the best models started paying billions to supply people. Microsoft assigned $2.5 billion and 6,000 staff to an implementation unit. Amazon committed $1 billion to embedding engineers inside customer operations, a move OpenAI and Anthropic had already made. The sellers of capability have concluded that capability does not transfer itself.

This essay is about the second P in the Process layer. Permission, which we covered on Monday, decides what the system is allowed to do. Practice decides whether anyone can actually do it. It is the layer where a tool becomes a routine, where a routine becomes standard work, and where standard work becomes the quiet compounding advantage that shows up in the numbers two quarters later, which is exactly where the Ramp and Revelio data, drawn from 21,599 firms, says it shows up, 6 to 12 months after adoption and not a day before.

The factory analogy is not decoration here, it is the whole argument. Toyota in 1985 did not out-buy General Motors, it out-practiced it. GM installed the same robots and got chaos, Toyota drilled standard work until improvement itself became a habit. The AI winners of the next decade will be built the same way, and this week the market, from Ford's recalled grey beards to Ibec's 64 percent reskilling figure, put an unusually clear price on believing otherwise.

What practice actually means

Most organisations believe they are already doing this. They point to the training budget, the lunchtime webinar, the e-learning module with the completion certificate. None of that is practice. Training is exposure. Practice is repetition inside the real workflow, with real stakes, observed by someone who can correct you. The distinction matters because exposure decays in about a fortnight while practised routines compound, and because the budget line for the first is easy to approve and the calendar time for the second is not.

An operational definition helps. A team is practising when three things are true. The tool is used on live work, not sandbox exercises. The use follows a written routine, short enough to fit on a card, that says when to reach for the system, what to give it, and what to check before the output moves on. And someone senior enough to change the routine watches often enough to improve it. Remove any one of the three and you have a pilot, which is to say a place where enthusiasm goes to be admired.

This is also where the curiosity argument lands. Ibec's report calls curiosity the capability we need to manufacture, and it is right, but curiosity in an operational setting is not a personality trait, it is a by-product of routine. People get curious about a tool once it has saved them an hour, twice, on work they recognise as theirs. The sequence runs from routine to result to curiosity, not the other way round. Organisations that wait for enthusiasm before building routine wait a long time.

The lag is the tell

The Ramp and Revelio study rewards a slow read. The headline is comfortable, firms that adopt AI grow headcount 10.2 percent over the following two years, with entry level roles up 12 percent at the heaviest adopters. The detail is more instructive. The gains are concentrated almost entirely in high-intensity adopters, classified by AI spend per employee in the first three months, and they do not appear until 6 to 12 months after adoption.

Consider what an organisation is doing in those silent months. It is not waiting for the model to improve. The model was as capable in month one as in month nine. It is rewriting job routines, discovering which handoffs break, learning what the system cannot be trusted with, and drilling the corrected version until it stops being remarkable. The lag between licence and payoff is a practice curve wearing a procurement disguise, and firms that cancel in month five, which is precisely when the finance review lands, pay for the curve without collecting the compounding.

Ford's grey beards tell the same story from the other side. The AI fell short, and the recalled engineers were not brought back for their capability with the new tools, they were brought back for their practised judgement about the old work. Judgement is what practice deposits. It cannot be procured on a purchase order, which is why Microsoft and Amazon are now selling people alongside the software.

Three routines that survive a regulated workflow

Frameworks are cheap, so here are the three routines that hold up when the work is regulated, audited and consequential. They are unglamorous by design.

The first is the daily rep. One task, done with the system every working day, by everyone in the role. Not the hardest task, the most frequent one. The point is volume, because volume is what surfaces the edge cases and builds the operator's feel for where the system is strong and where it lies. A rep that happens weekly is a demonstration. Daily is where feel comes from.

The second is the coached correction. Every output that a person amends before it moves on is a lesson, and almost every organisation throws the lesson away. The routine is simple, corrections are captured, the most frequent correction each week is discussed for ten minutes, and either the prompt, the routine card or the system configuration changes as a result. This is where the expert's judgement, the thing Ford had to rehire, gets transferred into the standard rather than retiring with its owner.

The third is the kata, the improvement loop itself practised as a routine. Once a week the team asks two questions, what did the system handle this week that it could not handle last month, and what are we still doing manually out of habit rather than necessity. The answers move work across the line item by item. Toyota's deepest insight was never the andon cord, it was that improvement is itself a skill that atrophies without reps.

Four minutes is the wrong number

A case study, offered as a composite of patterns we have seen rather than a single named client. A quality team at a pharmaceutical manufacturer drafts deviation reports, the structured write-ups required when production strays from its specification. The historical average was around 40 minutes per draft. With a well-configured assistant working from the batch record and the event log, the draft took four. The team celebrated, briefly.

Within six weeks the reviewers had started skimming. The drafts read fluently, arrived quickly and were right far more often than not, and that is exactly the combination that teaches a reviewer to stop reading. When a draft misattributed a temperature excursion to the wrong hold step, it passed two reviews before the QA lead caught it at final sign-off. Nothing left the building, but the near miss reframed the whole exercise.

The fix was not to abandon the tool, it was to redesign the routine. The draft now arrives with its three least certain claims flagged by the system itself, and the reviewer's first duty is to check those three against source. Four minutes became nine. Nine minutes against the original 40 is still a 78 percent reduction, and the catch rate at first review went back above where the manual process had it. The lesson generalises to any regulated workflow. The fastest configuration is rarely the best one, and the practised question is never how fast the machine can go, it is where the human's attention now earns its keep.

The Wednesday question

Practice is the least fundable-sounding word in the AI vocabulary, which is why it is systematically underbought, and why the biggest vendors in the world just spent $3.5 billion between them buying it for their customers. The capability in your organisation's tenancy this morning is, give or take, the same capability available to your fiercest competitor. The routines wrapped around it are not.

So the question from the operational floor stands. Who on your team used AI in the flow of work yesterday, what routine did they follow, and who coached the rep? If the honest answer is nobody, nothing written down and no one, the good news is that the fix costs less than the licences you have already bought.