To measure AI adoption, define who could use a tool, what counts as use, and the reporting period. Then measure participation, frequency, repeat use, and use in a specific task. Assess the quality and time required for that task separately. A usage report can show adoption without establishing whether the work improved.
You can start with a spreadsheet. The important part is making the definitions consistent enough to repeat the calculation.
Define who and what you are measuring
Before collecting numbers, complete this short specification for one tool and one team:
| Field | Decision to record |
|---|---|
| Population | Which employees have access and a relevant use case? |
| Qualifying activity | What recorded action counts as using AI? |
| Reporting period | What are the start date, end date, and time zone? |
| Coverage | Are you measuring managed accounts, survey responses, or both? |
| Known changes | Did access, team membership, training, or the tool itself change? |
Choose an event that reflects actual use. Opening a website or receiving a license does not establish that someone used an AI feature. For example, Microsoft's Copilot usage report distinguishes intentional feature activity from merely opening a pane. Its definitions are specific to that report; check the definitions for each tool you include. Microsoft's usage-report documentation
Also decide what the denominator means. Active users divided by enabled users measures use among people with access. Active users divided by everyone in the company answers a different question. Label the population beside the percentage so readers do not have to guess.
Five adoption metrics you can calculate
The measures below are a starter set for this guide, not a universal industry score. Choose the ones that answer your rollout questions. If a formula has a zero denominator, report it as not applicable and explain why; do not display a misleading zero percent.
1. Active-user rate: how broadly is the tool used?
Formula: unique qualifying users during the period ÷ eligible users during that period × 100.
Count people once, even if they generate many requests. If you combine several tools, deduplicate users before reporting an organization-wide total. Otherwise, one person using three tools becomes three people in your report.
Use this measure to see how participation changes within a defined population. Keep a record of roster changes: a percentage can change because access expanded, even if the number of active users stayed the same.
2. Weekly active users: does participation persist?
Count unique qualifying users in each complete week. Put the weekly counts beside the full-period active-user rate.
Do not add the weekly counts to calculate monthly users; the same person may appear in several weeks. The weekly series helps you distinguish a rollout spike from participation that continues. It does not tell you whether the same people return—that requires a cohort comparison.
3. Active-day distribution: how frequently do people use it?
For each active user, count distinct days with qualifying activity during the reporting period. Summarize the median and spread, or group people into clearly defined frequency bands.
A frequency measure needs its own interpretation. An employee who uses AI once for a monthly task may have found a useful application. Someone who uses it daily may still spend considerable time correcting the output.
Survey measures can add context, but keep them distinct from recorded usage. Gallup's workplace research separates reported use by frequency. A survey response about someone's work is a different measurement from activity captured in a company-managed account. Do not merge the two into an unlabeled adoption percentage. Gallup's AI indicator and methodology
4. Next-period retention: are earlier users returning?
Formula: people active in both periods ÷ first-period active users who remain eligible in the second period × 100.
This requires matching users across periods. Two aggregate totals cannot reveal retention.
Document how you handle people who leave the team or lose access. Keep an employee whose access was removed out of the eligible returning cohort, rather than automatically interpreting that absence as abandonment.
5. Task-use rate: is AI entering the work you selected?
Choose one defined task, such as preparing an internal summary.
Formula: eligible instances of that task with recorded AI assistance ÷ all eligible instances of the same task × 100.
Record assistance consistently. A task log or review sample may be necessary if the tool does not provide this information. Do not infer the task from time spent on a website, and do not combine unrelated task types into a percentage that conceals their differences.
This metric tells you how often AI participates in the selected work. It does not establish that the assistance was useful.
Worked example: adoption rises, but some users stop
The following figures are fictional. Assume the same 80 eligible employees throughout two reporting periods, with complete coverage under the same usage definition.
| Measure | Period one | Period two |
|---|---|---|
| Eligible employees | 80 | 80 |
| Active employees | 40 | 48 |
| Active-user rate | 50% | 60% |
The rate increased by 10 percentage points: 60% minus 50%. That is different from saying it increased by 10% relative to its starting value.
Now match the users. Suppose 30 of the original 40 active employees also used the tool in period two:
- Returning-user retention is 30 ÷ 40 × 100 = 75%.
- Ten previously active employees did not return.
- The other 18 active employees in period two were new or reactivated users: 48 minus 30.
The larger active-user total therefore hides some turnover in participation. Follow up with the earlier users to understand whether their tasks, needs, access, or experience changed. The calculation cannot identify the reason.
For the selected summary-writing task, suppose 36 of 120 eligible summaries had recorded AI assistance. Task-use rate is 36 ÷ 120 × 100 = 30%. That provides a basis for choosing work to review; it does not show whether the summaries were better or faster to produce.
Check whether the task improved
Keep adoption and task outcomes in separate parts of the report. For a chosen task, record:
- The task definition: what work is included and what counts as complete.
- Total effort: preparation, drafting, review, and correction—not just initial generation.
- Quality criteria: the same accuracy, completeness, and usefulness checks for both sets of work.
- The sample: how many examples you reviewed, how you selected them, and the period covered.
- Other changes: differences in task difficulty, staffing, instructions, or systems that may explain a result.
A before-and-after comparison can reveal a difference worth investigating. It cannot by itself isolate AI as the cause. If the sample is small or unrepresentative, say so before extending the conclusion to the whole team.
Build a report someone can act on
A useful first report can fit in one sheet. Include the tool, reporting period, population definition, eligible and active counts, weekly counts, frequency summary, retention cohort, task-use rate, coverage notes, and next action.
Use reports already available in your accounts where possible; export availability varies by service and plan. Supplement a usage count with a short conversation or survey about task relevance. Keep self-reported answers labeled separately.
Finish the report with a decision. For example: investigate why earlier users stopped, review a sample of AI-assisted summaries, or clarify access for the next group. Repeat the same definitions next period before deciding whether participation or task outcomes changed.