The short version
- Choose the trial with the people doing the work.
- Count preparation, checking and corrections as well as drafting time.
- Protect paid learning time and agree how any gains will be used.
Start with the people doing the work
Stanford’s 2026 AI Index reports that 88% of surveyed organizations used AI in 2025. That is a survey adoption measure, not the share of all workers using AI or evidence that deployments paid off.[1]
Ask the team where time is being lost or mistakes are happening. Choose one small task, such as an internal FAQ from approved information. Agree on the useful result, not just a target to use more AI. If checking becomes the bottleneck, generating more drafts can create more work.
Agree on the checks before the trial
- Define success: one outcome, an acceptable error level and a baseline.
- Count preparation, generation, review, correction and downstream rework.
- Name who verifies the output and who is authorized to release it.
- Keep sensitive data in approved systems and define the stop or escalation rule.
- Check whether junior staff are learning to recognize mistakes, not only accepting suggestions.
Protect the ability to learn
Use reversible trials, involve the people doing the work and preserve time for instruction. A human checkpoint is not useful merely because a person clicks approve: the reviewer needs evidence, expertise, time and the authority to reject the output.
Agree in advance how gains will be used. Faster service, lower workload, better access and reduced staffing are different choices. A productivity study does not decide between them, and a tool purchase should not conceal that decision.
Read the productivity evidence both ways
Research has found benefits in some settings and costs in others. One customer-support study found 15% more issues resolved per hour. Three developer trials found more completed tasks with AI assistance. They measured different work, so neither gives your team a universal speedup to expect.[2][3]
In METR’s early-2025 experiment, 16 experienced open-source developers took 19% longer on AI-allowed tasks. Its February 2026 follow-up reported results shifting toward speedups, but judged the estimates unreliable because of selection and measurement problems. The old slowdown should not be presented as today’s settled effect.[4][5]
A two-week support-drafting pilot
- Before
- Counting drafts per hour and calling the increase ‘time saved.’
- Small trial
- Use approved, non-sensitive cases and compare similar work with the current process. Keep the reviewer and quality standard fixed.
- What to check
- Log total time, factual errors, repeat contacts and escalations. Stop if quality worsens. A small pilot informs a decision; it does not prove a causal effect.
Read the evidence.
The practical steps are our editorial recommendations. Numbered citations link research claims to their sources. Expand a source for its findings, methods and limits.
The 2026 AI Index Report: EconomyStanford Institute for Human-Centered AI · 2026
Finding. 88% of surveyed organizations reported AI use in 2025.
Limit. Not 88% of all businesses or workers. Adoption does not establish deployment depth, return on investment or job effects.
- Type
- Research synthesis · survey-based adoption
- Period covered
- 2025 adoption data
- Scope & method
- Organizational survey evidence assembled in the AI Index’s economy chapter.
- Where to look
- Economy chapter: organizational adoption.
Generative AI at WorkBrynjolfsson, Li & Raymond · Quarterly Journal of Economics · February 4, 2025
Finding. 15% more issues resolved per hour on average; larger gains among less-experienced and lower-skilled agents.
Limit. Not randomized. Results vary by experience; experienced agents had small quality declines. One setting cannot establish economy-wide gains.
- Type
- Peer-reviewed · field study
- Period covered
- Earlier-generation assistant; not a 2026 tool evaluation
- Scope & method
- Staggered rollout to 5,172 customer-support agents at one business-software firm.
- Where to look
- Abstract and heterogeneous effects; QJE 140(2), pages 889–942.
The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software DevelopersCui and colleagues · Management Science · February 27, 2026
Finding. Pooled estimate: 26.08% more completed tasks (standard error 10.3 percentage points).
Limit. Individual trials were noisy. Output is not the same as quality or time saved. Includes Microsoft-affiliated authors and company-run trials.
- Type
- Peer-reviewed · randomized field trials
- Period covered
- Earlier code-completion deployments; published in 2026
- Scope & method
- 4,867 developers across Microsoft, Accenture and a Fortune 100 company; randomized access to a coding assistant.
- Where to look
- Abstract; pooled results and online appendices.
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer ProductivityMETR · July 10, 2025
Finding. AI-allowed tasks took 19% longer; the reported confidence interval ran from 2% to 39% longer.
Limit. Small, specialized sample using early-2025 tools. Read alongside METR’s 2026 follow-ups, not as a current universal slowdown.
- Type
- Randomized study
- Period covered
- February–June 2025 tools and tasks
- Scope & method
- 16 experienced open-source developers; 246 tasks in repositories they knew well.
- Where to look
- Main results and linked study paper.
We are Changing our Developer Productivity Experiment DesignMETR · February 24, 2026
Finding. Raw estimates shifted toward speedups, but METR judged them unreliable as measures of the current effect.
Limit. Developers and tasks selected out of non-AI conditions; lower pay and concurrent agents complicated measurement.
- Type
- Study follow-up · methodological warning
- Period covered
- Experiment begun August 2025
- Scope & method
- Follow-up with returning and newly recruited open-source developers.
- Where to look
- Opening results and ‘Wider adoption of AI has made it more difficult to measure task-level productivity’.
Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker ProductivityMETR · May 11, 2026
Finding. Median perceived value multipliers ranged from 1.4× to 2× across differently worded questions.
Limit. Self-reports are not measured productivity; low response rates and inconsistent answers limit interpretation.
- Type
- Survey · self-reported
- Period covered
- Perceptions around March 2026
- Scope & method
- Convenience sample of 349 technical workers, including METR staff and professional networks.
- Where to look
- Methodology and ‘Productivity gains around March 2026’.
Adoption of Generative AI in the Workplace: Increasing and Shifting the Balance of Productivity and Communication ActivityYu and colleagues · arXiv; Microsoft-affiliated research · August 16, 2026
Finding. Among higher-use adopters, productivity-app actions rose 21.2% and communication-app actions 7.1%.
Limit. App actions are not business value or output quality. Adoption was not randomized; vendor-affiliated preprint, not established consensus.
- Type
- Preprint · observational
- Period covered
- January–September 2024 data
- Scope & method
- Microsoft 365 activity in 11 companies; difference-in-differences analysis of Copilot adoption.
- Where to look
- Abstract; Data and Methodology. Higher-use group: more than 100 uses over 20 weeks.
Reviewed September 5, 2026. Publication dates and periods studied can differ. Exposure is not job loss; projections are not outcomes. Reading time covers the main briefing; learning resources, source notes and linked research are additional.
Browse the research library