The automatable middle

AI-assisted development can help with repetitive implementation work. The amount of useful work completed depends on the task, the developer, the surrounding system and the checking required.

Measure the complete task

Count review, correction and integration effort as well as generation time. A quicker first draft can still take longer to reach an accepted result. Conversely, useful work may expand beyond what the team previously attempted.

Keep research in context

METR's February 2026 update explains why selection effects make its newer developer experiment an unreliable estimate of current productivity benefit. Its May survey reports perceived speed and value, with limitations from self-report and selection. Neither establishes a universal rate of improvement for enterprise delivery.

Sources: METR experimental update and METR survey.

Look for the next constraint

As implementation changes, unclear requirements, review capacity and business decisions may become more visible. This is a practical question to investigate in the workflow, not a fixed law about how every team's bottleneck moves.

AI value starts with the workflow explains how I approach the measurement.