How do you distinguish activation, retention and abandonment of AI use?

Activation marks a defined first useful use. Retention describes whether the same people use the practice again over a stated interval. Abandonment means they have stopped that practice, which requires more evidence than an empty usage report.

For an adoption lead reviewing a rollout, the important question is what happened between the first attempt and the next opportunity. Someone who lost access, changed tools or had no suitable task needs a different response from someone who tried the workflow and decided it was not worth repeating.

In a public Reddit discussion, one user described teams switching from GitHub Copilot to Codex and Kiro. They also said that running out of credits meant waiting until the next month. The account is a personal report, with the employer and team size unspecified. It illustrates two ways a product's activity can fall while the explanation remains hidden in the dashboard.

Before interpreting a decline, define whether you are following one tool, one task-specific practice or any approved AI use. Keep that scope throughout the analysis. A tool switch can end use of one product while the underlying practice continues.

Define the first useful use

Choose an activation event that matters to the workflow. An account invitation establishes access. A first prompt establishes an attempt. Neither necessarily establishes that the person completed useful work.

For a wholesaler's sales team, you might define activation as completing a first AI-assisted customer reorder brief that the account manager checks and accepts. The brief brings together what a shop bought last month, items it is still waiting for and questions to discuss before its next order.

Record the qualifying event and its date. If your telemetry only records prompts, label that measure first activity and keep accepted work as a separate observation. Do not claim a quality check that your data cannot show.

Agree four things before examining the numbers:

• Practice. The task and approved tools included in the analysis.
• Starting event. What qualifies someone to enter the group being followed.
• Repeat event. What counts as another use, including any required output check.
• Observation interval. When another opportunity should occur and when follow-up is complete.

The group sharing a starting event or period is your cohort. Once someone enters it, retain their place in the record even if their circumstances change. You can show an additional rate for people who remained eligible, but readers should still be able to account for the original group.

The AI adoption metric dictionary covers the broader definitions and denominators. A cohort analysis adds the sequence needed to distinguish starting, continuing and stopping.

Give people a comparable opportunity to return

A person who started yesterday has had less time to return than someone who started two months ago. Comparing them within the same calendar-week total can confuse the rollout's growth with its persistence.

Choose a follow-up interval relative to each person's starting event. For example, compare repeat use during weeks five through eight after activation. Include only people whose full interval has elapsed in that comparison, and report how many are still awaiting follow-up. They have not failed to return; the measurement is unfinished.

Microsoft's Copilot adoption report documentation, updated August 2026, illustrates why these choices matter. It offers different activity settings and usage windows, and a trend view measured in weeks since first use. A report's settings help define the result. Its frequency labels do not establish a universal standard for useful AI adoption.

Let the task set the cadence. A wholesaler's monthly customer reorder brief may offer one meaningful opportunity in a period when a retailer's daily delivery enquiries offer many. The Cabinet Office's People Factor guidance, published June 2025, used weekly login data aggregated monthly for its Assist rollout while recognizing that relevant use varies by role and task. That context belongs beside the measure.

Write down the interval before reviewing results. If the work changes, revise the rule for subsequent comparisons and explain the break. Moving the window until the rate looks reassuring makes the result harder to use.

Account for everyone in the starting cohort

Two calculations answer different questions:

• 40% of the starting cohort repeated: 24 divided by 60. This is the qualifying repeat use observed across the original group in this interval.
• 60% of people with confirmed access and opportunity repeated: 24 divided by 40. Show this narrower rate alongside the original cohort and the other 20 people's circumstances.

The second rate adds context; it does not replace the first. Neither means that everyone outside its numerator abandoned the practice.

The groups must be mutually exclusive for these calculations to work. Here, the four groups without an observed repeat account for 36 people. Someone whose status cannot be established stays in the unknown group; you cannot assume they had no opportunity or chose to stop.

The 16 people with an opportunity deserve closer attention. Some may have used another approved method. Some may have tried AI again without producing an accepted reorder brief. Others may have preferred their previous process. The table identifies where to ask, but does not supply those answers.

Nor does it establish whether the 24 returning people saved time. Checking and correction effort still matter. Use the guide to measuring AI value when connecting repeated practice to better work or business results.

Ask what happened on the next task

Start the follow-up with the work, rather than asking someone to justify an empty chart. The Cabinet Office guidance recommends feedback from users and non-users, including interviews and focus groups. Hearing only from people still attending the program's sessions would leave important experiences unexamined.

Use a short conversation with the person doing the task:

1. Did the task come around again? Establish the actual opportunity and whether AI remained appropriate for it.
2. Could you use the approved workflow? Check access, limits and changes in responsibilities.
3. How did you complete the work? Distinguish a different tool, the previous method, an unsuccessful attempt and an unfinished task.
4. What made that choice worthwhile? Ask about the result, checking effort and any obstacle to repeating it.
5. What do you expect to do next time? Record whether the practice is continuing, paused, being replaced or deliberately stopped.

Keep an unanswered question unresolved. If only a few people respond, report how they were selected and how many remain unaccounted for. Their reasons can suggest a useful investigation without establishing why everyone else stopped appearing in the data.

Collect only the detail needed to support the workflow, agree who can see it, and report the program's findings at an appropriate group level. A usage review should help people explain friction without becoming an individual performance ranking.

Separate an inactivity rule from a decision to stop

An inactivity rule is useful for deciding when to check in. For a sales adviser who normally prepares monthly customer reorder briefs, you might investigate after two expected briefs pass without a qualifying AI-assisted repeat. That is an operational choice to test against the task's cadence, not a universal abandonment threshold.

Use a stronger label only when the evidence supports it. Someone saying they have stopped using the practice because checking takes longer than the previous method is a reported discontinuation. A removed license is an access change. A quiet account with no follow-up is observed inactivity with an unknown reason.

Retain the date and basis of the classification. A person can stop one practice and later return after their task, tool or needs change. Do not overwrite that history by treating every return as a brand-new activation.

If people return after support, record the return and examine the work they completed. A new deadline, a product change or a different task may also explain it. The sequence alone does not show that your intervention caused the improvement.

Once discontinuation is established, use the rollout recovery guide to reconstruct the failed task and choose a specific correction. Keep that decision separate from the measurement that first revealed the decline.

Questions about AI activation and retention

Does one use count as activation?

Yes, if that use meets the activation rule you set for the AI practice. For a wholesaler's sales team, a useful rule could be the first AI-assisted customer reorder brief accepted by its account manager after checking the order history, outstanding deliveries and questions for the next order. Sending a prompt alone would establish first activity, but would not show that the person produced acceptable work.

Write down the qualifying event before counting people. If your data only shows logins or prompts, report that limited measure explicitly. The section on defining the first useful use sets out the practice, starting event, repeat event and observation interval to agree.

How long should someone be inactive before we investigate?

Base the check-in on when the person's task should recur, rather than applying the same number of quiet days to everyone. A wholesaler's monthly customer reorder brief and a retailer's daily delivery enquiries offer different opportunities to use AI. First confirm that the person had another suitable task and could access the approved workflow.

Choose and document an inactivity trigger before reviewing the results. Crossing it should prompt a conversation, not an automatic abandonment label. Use the five follow-up questions to distinguish missing opportunity, access problems, a different method and a deliberate decision to stop.

Is switching AI tools abandonment?

A switch can end use of one product while the underlying AI-assisted practice continues. A wholesaler's sales adviser might stop drafting customer reorder briefs in one approved tool and prepare the same checked briefs in another. Whether that counts as abandonment depends on whether your measure follows the original product or the task-specific practice.

State that scope beside the result and record an established replacement separately. Do not add overlapping user counts from both products as though they represent different people. The AI adoption metric dictionary explains how to keep the population, event and denominator clear.

Should new users be included in a retention comparison?

Include people only after they have completed the follow-up interval used for that comparison. If you measure repeat use during weeks five through eight after first useful use, someone who started last week is not yet eligible for that result. They have not failed to return; you have not finished observing them.

Show recent starters as awaiting follow-up, and report their activation separately if useful. Compare groups at the same age so rollout growth does not get confused with continued use. See how to give people a comparable opportunity to return for the interval and task-cadence choices.

Does higher retention mean the rollout is working?

Higher retention shows more repeat use under your chosen definition and observation window. It does not, by itself, show that the work is better or that the total effort has fallen. People may keep using a workflow that still requires substantial checking and correction, while stopping an unsuitable practice may be a sensible decision.

Review accepted output and total effort alongside repeat use, then decide whether to continue, change or stop the practice. The guide to measuring AI value explains how to connect activity with work quality and business results. Keep both the repeat-use finding and the outcome evidence visible when reporting progress.