A pipeline has four stages. Three run at about 30% utilisation and one runs at 95%. The team adds capacity to all four. Throughput barely moves. Why?

The mechanism

A chain's throughput is set by its slowest stage, so capacity added anywhere else is capacity that idles. The three stages at 30% were never limiting anything; they were waiting. Making them faster makes them wait more.

Only the 95% stage was binding. Whatever fraction of the new capacity landed there is the only part that did anything — and if the team spread the investment evenly across four stages, roughly a quarter of the money bought the whole of the improvement.

Why the 95% stage behaves so differently

Utilisation does not affect waiting time linearly. As a stage approaches full utilisation, queueing time rises sharply: at 50% utilisation a unit of work waits about as long as it takes to process, and at 95% it waits roughly twenty times that. This is why the last few percent of load feel so different from the first fifty, and why a stage at 95% dominates end-to-end time even if its own processing step is short.

It also means the payoff is asymmetric in the other direction. Taking the binding stage from 95% to 80% utilisation cuts its queueing time by roughly four times — a far larger effect than the 15-point change suggests.

The decision rule

1. Measure where work waits, not where it is busy. Queue depth and time-in-stage, per stage.
2. Improve only the stage that binds.
3. Measure again, because the constraint moves. Fix the 95% stage and something that was at 60% is now the limit. This is not a failure of the fix; it is the expected result, and teams that do not expect it conclude the fix did not work.
4. Stop when the constraint sits outside your system — in a vendor, a human approval, or demand itself.

When adding capacity everywhere is right

When you genuinely cannot measure per-stage waiting, and the whole system is small enough that over-provisioning costs less than the investigation. Under a few hundred dollars a month, buy the headroom and move on. The discipline matters when capacity is expensive or when the stages are teams rather than machines, where adding people to a non-binding stage makes things slower rather than merely wasteful.

Common weak answers

• "They needed more capacity overall." They needed more in one place. Spreading it is the error.
• "The stages are coupled." Sometimes true, but not required: independent stages in series behave this way purely from the arithmetic of a chain.