AI Productivity Measurement Evidence Record

AI Productivity Measurement Evidence Record

Evidence pattern

Workplace AI effects vary by task, worker, system, and measure.

Brynjolfsson, Li, and Raymond's customer-support field study examined 5,179 agents and reported an average productivity increase in issues resolved per hour, with larger effects among less-experienced workers and limited effects among the most experienced workers. That result belongs to the studied support workflow.

Dillon and colleagues' 66-firm field experiment reported reduced email time and less work outside regular hours among users of an integrated generative AI tool, while detecting no shift in the quantity or composition of tasks from individual access alone. The authors disclose relevant Microsoft employment and review.

Dell'Acqua and colleagues' published consulting experiment found that effects differed across tasks and could worsen performance outside the tested system's capability boundary.

These studies support measurement in context. They do not supply a universal ROI percentage.

Measurement unit

The unit should be accepted work, not generated text. Start time includes setup, source collection, prompting, waiting, review, correction, formatting, filing, support, and incident handling. Completion occurs when the result meets the same acceptance standard as the baseline.

Quality must cover consequential correctness, completeness, traceability, fit, and harmful failure. Average time cannot erase a rare severe error.

Comparison design

Use comparable tasks, stable definitions, representative cases, a preserved baseline, and the same outcome standard. Record task difficulty, worker experience, tool version, configuration, data boundary, reviewer, and source conditions.

Report sample size, missing cases, uncertainty, outliers, adoption rate, and changes during the test. A self-reported estimate of time saved is a perception measure, not a complete ROI calculation.

Financial boundary

Net value should include licenses, integration, administration, training, review, correction, support, compliance, incidents, vendor management, and opportunity cost. Avoid converting every minute into realized cash savings unless labor, capacity, demand, and accounting treatment support that claim.

The decision can be approve narrowly, modify and retest, pause, or stop.