Last month a team switched on a new AI tool. The dashboards turned green: adoption at ninety percent, turnaround time cut in half, a ready-made success story for the management report. When I spoke to the same team three weeks later, the picture had shifted — people were staying later, asking more questions, redoing some of the work twice. The dashboards were still green. The team wasn’t.
This contradiction sits at the centre of a mistake I keep running into: pricing the tool by its sticker and never pricing the learning at all. AI subscriptions get a little cheaper every month; some are close to free. But the old reflex a team has to unlearn, the shortcut it has to forget, the new species of error it has to recognise — none of that appears on the invoice. Accounting sees one number, the licence fee. The team’s number is buried inside working hours, with no line item of its own.
A cheap tool ends up expensive — because the cheapness stays in the price, and the cost hides in the time.
Redoing a piece of work twice tends to follow the same sequence. First the tool gives the right answer in the wrong format, and someone fixes it by hand. Then it gives the right format but skips an assumption nobody thought to state in the old process, the error reaches the customer, and it comes back. Eventually the team learns not the tool itself but the new checks that have to wrap around it — a lesson that lives in no training video, only in the mistake itself. Companies call this stretch “transition pain” and wave it off; the pain is exactly the cost that should have been measured.
When several tools land at once, this bill compounds rather than adds. Each one brings its own shortcut, its own exception, its own list of “don’t do it this way.” The team climbs three learning curves at once, and each one stacks on the others. Management sees three separate green dashboards; the person at the desk sees one red week.
MIT economist Erik Brynjolfsson’s productivity J-curve names exactly this spot: when a general-purpose technology arrives, productivity first dips before it rises — because the technology itself is cheap, but the complementary investments that make it useful (redesigned processes, new skills, new decision rights) are expensive and slow. The J’s first leg is the one everyone misreads; what looks like failure there is only the early instalment on the bill.
The real mistake here isn’t managerial, it’s accounting. When a company buys new machinery, it books depreciation on the balance sheet, spreads it over three to five years, and nobody is surprised. When a new AI tool arrives, the adaptation period gets treated not as a cost line but as “transition friction” — something to be embarrassed about and rushed past. But adaptation is depreciation; it’s simply never been booked. And this invisible line item lands hardest on the middle manager: the tool was leadership’s decision, the learning curve is their overtime, and the delay becomes their record.
When I bring in a tool now, I ask one question: who pays for its learning curve — the budget, or the overtime? If the answer is always the second, the saving on paper is an illusion; only whose pocket it comes from has changed. A company that sees this sets aside an actual adaptation budget: training hours, the first month’s lower output, the overlap cost of running the old and new process side by side — each its own line, not swept under “start-up friction.” Real efficiency doesn’t come from the tool’s speed; it comes from an organisation that sees, and covers, the cost of the team catching up to that speed.
Someone gets billed. The trouble is, it’s usually the person whose name never appears on the invoice.