AI wrote it in 12 minutes. How long until it was usable?

Count briefing, waiting, review and correction as well as generation. Keep faster completion separate from less effort or better output.

Our view: measure time until a task meets the same acceptance criteria, rather than time until AI produces a draft. A pleasant workflow and a fast workflow can both be valuable, but they answer different questions.

Count the whole sequential example

In our fictional task, working without AI takes 40 minutes. With AI, briefing takes 5, generation 12, review 18 and correction 8: 43 minutes in total. The draft arrived quickly, but accepted work took 3 minutes longer. These are invented timings, not measurements of a service.

Decide what counts as finished

Use the same requirements for accuracy, format and necessary checks in both workflows. Record failures and abandoned attempts too. Otherwise, a quick but unusable draft can appear to beat a completed task.

Why a simple productivity headline is hard

In its February 24, 2026 update, METR says selection effects and time reporting problems make its later developer experiment an unreliable signal of current productivity impact. It also notes overlapping agent work can complicate time measurement. That does not establish the effect on your own work.

Measure the benefit you actually need

For parallel work, separate elapsed time from active human minutes instead of adding overlapping intervals. Record quality and effort separately. A few tasks can reveal where to investigate; they cannot establish a universal AI productivity rate.

Published October 14, 2026. Editorial analysis based on a dated research update and an original arithmetic example. We ran no developer productivity experiment or AI service timing test.

Sources: METR · Productivity study update.