The survey finding
TechCrunch reports on a new survey with a thesis that will feel familiar to anyone who manages people using AI tools: the productivity is promised, the headache is delivered. Managers, the survey suggests, are not seeing the gains they were told to expect.
Worth being precise about what that does and does not mean. It does not mean AI tools don't work. It means the gains and the costs are landing on different people — and the people absorbing the costs are the ones holding the headcount budget.
Generation got cheap. Verification didn't.
The economics of knowledge work changed at exactly one point in the pipeline. Producing a first draft — of a report, a function, a legal summary, a marketing email — went from hours to seconds. Nothing downstream of that point got any faster.
The draft still has to be read, checked, corrected, and owned. That work does not parallelise, cannot be batched, and scales linearly with the number of drafts. Which means the more AI a team adopts, the more of its manager's week goes to review.
This is not a new observation in software — it is the oldest one there is. Code review was the bottleneck before anyone pointed a model at a codebase, and every "10x developer" claim in history has quietly assumed an infinite-capacity reviewer downstream. What is new is that the bottleneck has escaped engineering and moved into marketing, legal, finance, and support, where there is no culture of review and no tooling for it.
Three things managers are actually absorbing
Verification debt. Output that looks finished is more expensive to check than output that looks rough, because the reviewer's attention has to be actively recruited. Confident, well-formatted, subtly wrong work is the worst case, and it is precisely what models produce.
Unmeasurable baselines. You cannot claim a productivity gain without knowing what the task cost before. Most organisations never measured that, so every AI rollout is evaluated against a feeling.
Ownership ambiguity. When a person and a model co-author a document, accountability does not divide cleanly. Managers end up owning work they did not write, cannot fully reconstruct, and may not be able to defend in a meeting.
The "for now" is the important part
The headline hedges, correctly. The tools are getting better at the things that make review expensive — citing sources, flagging uncertainty, producing diffs instead of documents. The organisational side is what lags: nobody has decided how much review time a given AI-assisted task is allowed to consume, so it consumes all of it.
The practical fix is unglamorous. Budget review hours explicitly, the way you budget build time. Prefer AI on tasks where verification is cheap and automatic — code that compiles and passes tests, SQL that returns a sane row count, transformations with an obvious checksum. And measure rework rate, not output volume: the number that matters is how much of the generated work had to be redone, and by whom.
None of that requires new software. It requires admitting that the bottleneck moved, and that the person standing next to it is a person.
