What AI studies found

In a randomised task experiment with 99 accountants, AI assistance increased classification accuracy by about 18 percentage points. It did not produce a statistically significant change in average completion time.

The experiment tested one structured transaction-classification task. It did not establish workplace savings after review and operating costs.

I’d use synthetic transactions to compare correct coding, review time and later corrections with manual coding.

In simulated HR writing tasks with 91 final-year business students, the study reported 54% less completion time and 36% higher quality ratings from HR practitioners.

Students received ChatGPT at different points in the experiment. The study did not measure results for HR staff at work.

I’d compare the time to draft and review the work, its quality and corrections with current templates. Hiring decisions stay with people.

Three models scored 123 past bids for 31 Ohio IT projects. Scores were more consistent across repeated runs on minimum technical requirements than on what distinguished suppliers. Scores on minimum requirements also aligned with human reviews.

The study rescored historical bids. It did not measure review time or procurement outcomes.

I’d use public or synthetic bids to compare missed requirements, review time and score consistency. People still judge value and select suppliers.

Across three company trials with 4,867 developers, the authors reported a pooled 26% increase in completed tasks among developers using GitHub Copilot.

Results varied across the trials. The study did not establish code quality or business value. A separate 2025 trial of 16 experienced open-source developers found tasks took 19% longer with AI. METR’s follow-up suggested newer tools could help, but selection and time tracking made that estimate unreliable.

I’d compare accepted code, defects, rework and review time on similar tasks.