How To Measure AI Productivity

A community team has tried an AI assistant for writing briefings. Drafting now takes thirty minutes instead of ninety. The team wants to expand the pilot. What should it check before calling this an improvement? In this lesson, How To Measure AI Productivity, we will follow the briefing all the way to an accepted result. We will count the work involved, investigate where effort changed, and use the comparison to decide what the team should test next.

Our community team and its figures are an imagined teaching case. The team needs ten short briefings based on approved public documents. Each briefing must contain the requested information, support factual claims with those sources, and pass the same review criteria. These accepted briefings are our useful output. That definition tells the team what to count. Generated words and unfinished drafts can be recorded separately to understand individual steps.

Now follow the people doing the work. Someone drafts, someone checks sources, and someone corrects errors. We will add their labor time to measure accepted briefings per total labor hour. We begin with recurring work for each batch, then account for the effort needed to prepare the redesigned process. State which workers and activities the record includes. A thirty-minute meeting involving two reviewers uses one labor hour, even though half an hour passes on the clock.

Before the assistant, drafting takes ninety minutes, source checking sixty, and correction thirty. Add those stages: one hundred and eighty minutes, or three labor hours. Ten accepted briefings divided by three hours gives about three point three three accepted briefings per labor hour. This is the team's baseline for the comparison. The first AI-assisted pilot produces the same ten accepted briefings. Drafting takes thirty minutes, checking one hundred, and correction seventy.

The total is two hundred minutes, or three hours and twenty minutes. Ten divided by that time gives three accepted briefings per labor hour. The first draft arrived sooner, while the complete process took more effort. Labor productivity fell ten percent under this measurement. That result gives the team a useful next question: why did checking and correction grow? The timing record shows where to look. The team should inspect rejected content, repeated checks and work handed to other people. Reviewers may be searching through poorly organized source versions

or reconstructing unsupported claims. These are explanations to investigate. Trace an actual correction back to the draft and source before selecting a change. Suppose the team prepares an approved source set and uses a consistent review template. For the next imagined batch, drafting takes thirty minutes, checking forty-five, and correction twenty-five. The same ten briefings pass the same acceptance criteria. Recurring effort totals one hundred minutes.

The rate is six accepted briefings per labor hour, eighty percent above the baseline. Source preparation and a common review process are parts of this redesigned workflow. Preparing that process also takes time. Assume it requires sixty additional labor minutes. Charge all of that setup effort to the first batch and its total becomes one hundred and sixty minutes. Ten briefings divided by two hours and forty minutes gives three point seven five accepted briefings per labor hour.

The first-batch and recurring rates describe different periods. Keeping both makes the transition effort visible. Training, integration and management effort belong in the record when they are part of the work being evaluated. Before applying the comparison, check that the work stayed comparable. Did both batches include similar tasks? Did reviewers use the same quality standard? Did the record count staff time, support and repeated work consistently?

State how waiting that occupies a worker is allocated; track elapsed service delay separately. A changed task mix or another process change could help explain a different rate. A real pilot needs records and a suitable comparison to investigate the assistant's contribution. The timing table helps the team design that investigation. The team also needs to track other resources. Labor hours, computing payments and energy

answer different questions and use different units. Record dollars per accepted briefing or energy per accepted briefing alongside the labor rate when those costs matter. An output-per-labor-hour improvement can accompany higher computing spending. Keep the quantities separate so the decision-maker can see that trade-off. Broader economic methods can combine inputs using specified calculations; our local task measure uses labor hours. Real research shows why defining the completed output matters.

The published 2025 study Generative AI at Work measured customer issues resolved per hour in one firm's support work, with results that varied across workers. It gives us a concrete outcome and setting to inspect. A September 2026 working-paper summary by economist Leon Musolff and his coauthors distinguishes coding activity from software released and used. We use that author summary's qualitative distinction here. Its particular software-development setting helps explain

why a team should follow an intermediate activity through to the result people need. Here is the first Learning Check. A vendor reports that its model now generates twice as many tokens per second. What else would you need before concluding that your team's labor productivity has improved? Pause and give a specific answer. You would need a defined useful output and acceptance standard,

then the total labor needed to reach that output, including preparation, review and correction under a stated setup allocation. Compare like tasks and periods. The token rate can explain one technical step, while the team measure follows the complete work. Now try a different comparison. One batch produces eight accepted reports in two labor hours. Another produces twelve accepted reports in four labor hours.

Both pass the same quality standard. Which batch has higher labor productivity, and why is the larger output count insufficient? Pause and calculate the two rates. The first batch produces four accepted reports per hour. The second produces three. The second batch has more total output, but its input doubled while output rose by half. Its labor productivity is twenty-five percent lower.

Keep both the total output and the rate when describing what changed. Return to the team deciding whether to expand its pilot. It can now explain three different records: the baseline, the first pilot with more review effort, and a redesigned process with recurring and setup effort stated separately. In a real decision, it should collect comparable accepted-output, quality and full-effort evidence before expanding. That gives the team a specific measurement task.

How any verified gain changes service capacity, roles or pay also depends on organizational choices. Lesson two follows tasks into jobs, and Lesson five examines skills, wages and who receives the value. For Practice, use the separate worksheet to compare six accepted notes in a baseline and an AI-assisted process. Add drafting, checking and correction, convert the totals into hours, and calculate the rates.

Write the conclusion your comparison supports and the evidence you would seek next. Complete your response before using the separate answer key. You can then apply the same method to a familiar task: define the accepted result, follow the whole effort and compare like work.