Named source: METR · study and 2026 follow-up
The case, explained simply
METR · when assistance slows work
Measure completed work instead of relying only on how productive AI feels.

- Comparable coding tasks
- AI allowed / not allowed
- Review and tests included
- Measured completion time
Explanatory workflow. Read the linked case or sector analysis for the evidence, conditions and limits behind each step.
What actually happened
METR’s early-2025 randomised study found experienced open-source developers took 19% longer with the tested AI tools. Its February 2026 update says later results were affected by selection and measurement problems, so they cannot reliably establish the current size of the productivity effect.
Why this matters for adoption
Reading suggestions, supplying context and fixing plausible but wrong code are real work. A developer may enjoy the assistance while the complete task takes longer. Equally, a different tool or task may help substantially. The adoption lesson is to measure end-to-end delivery and revisit the result as the technology changes, rather than treating a historical study as a permanent verdict.
THE CASE AT A GLANCE
Read a case like a decision-maker.
Do not stop at the headline. Follow the four things that make a case useful in a different organisation.
Which task actually changed?
What was measured, by whom and in what setting?
What must your team test before it relies on this?
THE LESSON BEHIND THE CASE
What has to be true for this to help?
People can feel faster while taking longer. For an experienced team, explaining the codebase to an assistant and reviewing its suggestions may consume more time than the original task. This does not establish that coding tools never help; it shows why local task measurements matter. Include review, debugging and later rework. Compare similar tasks and avoid judging success through lines of code, suggestion acceptance or enthusiasm alone.
Compare completed maintenance tasks of similar difficulty with and without assistance. Record time until review acceptance and defects found later, not just time until code first runs.
A PRACTICAL DECISION MAP
What your team needs to make this work
Have this ready
Comparable real tasks and measured completion time
Look for this result
Time to accepted work including later fixes
Pause if you see this
Assistance adds review and debugging time
An editorial guide for planning your own trial. Agree the required result and acceptable effort with the people who do the work.
What the evidence does not tell us
This is not evidence that AI always slows developers. The original result concerns a small, experienced population working in familiar repositories with tools available at that time.
What a Greek organisation can learn
- Choose one clearly defined task and the person responsible.
- Compare the complete work with the current process.
- Decide whether to extend, change or stop based on observed results.
How to test the idea in your organisation
Compare similar tasks and include review, testing and rework. Record tool versions and developer experience. Use actual completion time as well as staff feedback; neither tells the whole story alone.
Before spending money
Ask for setup, integration, training, review, recurring charges and exit costs separately. Public case evidence is not a price quotation for your organisation. Agree a limited trial before committing to a wider rollout.
Read the original evidence
METR · study and 2026 follow-up ↗The practical steps above are AIStrategy’s editorial suggestions. They are not reported actions or promised results from the organisation featured. Photography illustrates the subject unless its caption identifies the actual facility.