AI has made every individual step of building a product faster: research, requirements, documentation, prototyping. It has not yet made the product faster to build. The gap between those two facts is where most AI transformation programmes fall through, paying an integration tax nobody put in the plan.
Amy Mitchell named this Invisible Product Work: acceleration didn’t remove the work of connecting decisions across teams. It made that work easier to postpone, because everything around it now looks like it’s moving fast.
The fourth job
Picture three teams shipping against the same product: design, backend, and growth. Each one now moves through its own decisions in a fraction of the time it used to take. A research synthesis that took a week takes an afternoon. A requirements doc that took three drafts takes one. A prototype that needed a sprint exists by lunchtime.
None of that speed touches the fourth job: someone still has to notice when design’s new flow assumes a data model backend hasn’t built, or when growth’s fastest-testing variant contradicts a decision product made two sprints ago. That job was never automated. It was never even sped up. What changed is how much of it there now is, because three teams making decisions three times faster produce a lot more decisions that need reconciling, not fewer.
That tax is the reconciliation work that used to get forced into the open by slow, sequential handoffs, and now accrues quietly in parallel because nothing about faster local work requires anyone to stop and check it against anything else.
I’ve seen this collision at close range. Two accelerated workstreams each shipped a decision that was correct on its own terms: one team restructured how data flowed into a report, another changed what that report was used to decide. Neither decision was wrong, and neither team could have known about the other’s change from anything in their own sprint. The collision surfaced in integration testing, three weeks before a deadline, at the exact point where undoing either decision cost more than building both correctly the first time would have.
Why it stays invisible until it’s expensive
A slow process fails loudly and early. A handoff that takes two weeks gives everyone two weeks to notice the assumptions don’t line up. Fast, parallel workstreams remove that forcing function. Each team’s own view looks coherent, because it is coherent, on its own terms. The incoherence only exists in the gaps between teams, and gaps don’t show up in anyone’s individual sprint review.
This isn’t the coordination theatre I wrote about in AI Won’t Save the Kingdoms We Built: performing collaboration through more messages and better-formatted updates while avoiding the actual decision. The integration tax accrues even when nobody is avoiding anything. Every team is deciding fast and deciding well, individually. The tax is simply the arithmetic of how many good individual decisions now exist per week, set against how few of them get checked against each other before they collide.
That piece was about leaders avoiding the hard organisational calls AI can’t make for them: who owns a contested decision, whose incentives get realigned. The integration tax shows up even when none of that is broken. A well-run org with clear ownership still generates it, purely from the arithmetic of throughput: more decisions per week, made faster, by more independently accelerated teams, with the same number of people available to notice where they collide.
The bill doesn’t arrive as a missed deadline. It arrives as a launch where three teams’ fast, individually reasonable decisions turn out not to fit together, and the fix costs more than doing it right the first time would have, because by then it’s wired into three different codebases and two go-to-market plans.
What actually has to change
Cross-functional teams of multiskilled people, building together continuously, are the strongest mitigation I’ve found: not a checkpoint, not a cadence, not design, engineering, and product running parallel streams and reconciling later. When the same room makes the design call and the data-model call in one conversation, most of the tax never accrues, because there was never a gap for it to hide in.
That doesn’t make the tax disappear; most organisations aren’t built that way end to end. I’ve written before about why that job lives specifically at the boundary between teams, not inside any one of them: the people positioned to catch a coming collision are the ones with enough range to recognise it from both sides. AI didn’t create the need for that range. It just raised how often everyone else now needs it.
Speed was never the part of this work that was hard. Noticing what your speed costs someone else, in time to do something about it, always was. That’s the job AI still can’t do for you.
Leave a Reply