AI at Work · Source published 10 July 2025
Research signal
METR’s July 2025 study found that experienced open-source developers in its specific experiment took 19% longer with the tested AI tools, despite expecting a speed benefit.
Source: METR — experienced developer productivity study · 10 July 2025.
Event Systematic perspective
The result is a reminder that perceived ease and measured completion time are different outcomes. It is not a universal verdict on AI coding or on later tools. Task complexity, familiarity, review effort and the available system all affect a useful evaluation.
An employer can apply the lesson without copying the experiment. Select comparable tasks, define what counts as finished and include review and correction time. Record the tool version and the user’s experience. Ask participants about ease and confidence separately from the timing measure. Both can matter, but they should not be combined into an unsupported savings claim.
Put it into practice
- Measure accepted output rather than the amount of generated material.
- Include downstream checking and integration in the time comparison.
- Repeat the evaluation when the task or tool changes materially, and retain the original baseline.
Explore employer services, career development and the AI Academy. Discuss your workforce priorities with Event Systematic →
