AI Assistants Made Individual Output Cheap. Coordination Is Still Expensive.

The productivity story of the last two years has been individual: people produce more drafts, more code, more analysis, faster. That part is real. What has not changed is the organisation's capacity to review, decide, and integrate, and the result in many teams is a queue that got longer rather than a system that got faster.

The bottleneck moved and nobody moved with it

If writing a proposal took two days and reviewing it took two hours, halving the writing time barely changes the cycle. If writing now takes two hours and review still takes two hours, review is suddenly fifty percent of the process. Multiply that across a team and the constraint is unmistakably in the approval path, not the production path.

Teams that felt no speed-up from better tooling usually have this shape. They optimised the step that was never the constraint.

Where the new time actually goes

• Review load per reviewer rose sharply. More output means more to check, and checking is not something a tool can do on your behalf without becoming the thing being checked.
• Volume made quality signals noisier. When producing a document is nearly free, the existence of a document stops indicating that someone thought hard.
• Decisions became the queue. Work now waits on people with authority far more often than it waits on people with capacity.

Rebalancing moves that work

Push decision authority down until it sits with the person doing the work, with a clear list of what still needs escalation. Cap work in progress so the team finishes rather than accumulates. And make review a scheduled activity with owners rather than an interrupt that competes with everything else.

Then reduce the review burden itself: smaller units of work are faster to review and safer to reject, and a document with a two-line summary at the top gets read while a five-page one waits.

Generating more artefacts does not increase throughput if every artefact needs a human decision and the humans have not changed.

Measure the right interval

Track cycle time from idea to shipped, and the share of that time spent waiting versus working. If waiting dominates, adding production capacity, whether human or automated, makes the queue longer without making anything arrive sooner. That number is the one to watch when someone proposes another tool.