How to Measure AI Adoption Beyond Licenses and Active Users

Start with two people on the same sales team. Both show up as “active” in every license report.

Each level of the framework sees these two differently. Access says both can use AI. Adoption says both do, and that A does more of it. Only proficiency separates them, and only impact tells you that B’s change is worth spreading to the rest of the team.

Each level answers a different question and each is a prerequisite for the next. Reporting only the first two is how organizations end up with a high adoption rate and no idea whether anything improved.

Access is the share of the workforce that could use an approved AI tool today: a license, a seat on an enterprise plan, or an SSO entitlement to a sanctioned product. It is the easiest level to measure and the easiest to mistake for progress.

Metrics

• Employees with AI access
• Licensed employees
• Access by department
• Approved AI tools available

Where the data comes from. The identity provider and each vendor’s license admin console, joined to the HR system for department and role.

Watch out for. Access is a prerequisite for adoption, not evidence of it. A fully licensed workforce with 30 percent usage is a rollout problem, not an adoption success.

Adoption is whether people with access use the tools, how often, and whether they keep coming back. This is the level most dashboards stop at, and it is where the headline number lives.

Metrics

• Weekly active AI users
• Monthly active AI users
• Percentage of licensed employees using AI
• Usage frequency
• Usage by department
• Usage by tool
• Retention over time

The core formula is simple: AI adoption rate = active AI users ÷ employees with AI access. If 600 of 1,000 licensed employees used an approved tool in the last 30 days, that is a 60 percent monthly adoption rate. Measured weekly, the same organization might sit at 41 percent. Both numbers are true; the weekly one is the one to manage.

Where the data comes from. Vendor admin reports (the ChatGPT Enterprise, Copilot and Gemini consoles each expose active-user counts), browser or network telemetry for tools outside the approved list, and identity-provider sign-in logs as a floor.

Watch out for. Define “active” as at least one real interaction, not a login. Report weekly as well as monthly; the gap between the two tells you how many people are occasional visitors. And report by department and by tool, because a company-wide rate hides the sales team at 72 percent and finance at 18.

Proficiency is where most measurement programs stop short, because it cannot be read off a license report. It has four observable parts.

Workflow complexity

The same tool, two very different requests.

Complexity cuts both ways. In a field experiment with 758 consultants, people using GPT-4 on tasks within the model’s capability completed 12.2 percent more tasks, 25.1 percent faster, at 40 percent higher rated quality. On a task chosen to sit outside that capability, the people using AI were less likely to reach the right answer than those working without it.³ Knowing what people use AI for matters because the same tool helps or harms depending on the task.

Use-case diversity

Is AI applied to one kind of task, or across the work a role does? The categories worth tracking are research, analysis, decision support, customer preparation, planning, document creation and repetitive operational workflows. A team whose usage sits entirely in “document creation” has found one use, not adopted a capability.

Organizational context

Generic prompts get generic output. Proficient use brings the organization’s own information into the work: CRM data, internal documentation, customer information, company policies, analytics and the approved tools that hold them. The share of usage that carries this kind of context is one of the cleanest proficiency signals available.

Repeatable workflows

The clearest sign of proficiency is a task a role does the same way, with AI, every time. In sales, preparing for a discovery call. In product, analyzing customer interviews. In customer success, preparing a quarterly business review. In finance, explaining the variance between forecast and actuals. Count them per team, and count the ones with an owner.

The behaviors underneath

Published frameworks converge on the same set of behaviors: deciding what to delegate, describing the task with the right context, judging the output critically, and taking responsibility for the result. Anthropic’s AI Fluency framework, developed with Rick Dakan and Joseph Feller, names them delegation, description, discernment and diligence.⁴ The judging part is not automatic: a Microsoft Research survey of knowledge workers found that the more confidence people placed in the AI, the less critical-thinking effort they reported, while confidence in their own skills pushed the other way.⁵ Verification is a behavior to measure, not to assume.

Where the data comes from. Sampled conversation review under an explicit privacy policy, aggregate behavior analytics from tools that provide them, workflow inventories kept by team leads, and self-assessment as the weakest input.

Watch out for. Never measure proficiency by reading individuals’ conversations without consent. Every proficiency metric on the scorecard below can be produced in aggregate.

Usage measures whether AI is being used. Impact measures whether work is getting better, in units the business already tracks.

Attribution is the hard part. Three rules make it tractable. Measure against a baseline taken before the rollout. Compare adopters with non-adopters in the same role over the same period, rather than the whole company against last year. And separate modeled estimates from measured results, labeling each.

Expect the gains to be uneven. In the best-known field study, giving customer support agents an AI assistant raised issues resolved per hour by 14 percent on average, with a 34 percent gain for novice and low-skilled agents and little change for experienced ones.⁶ That distribution, large for some people and negligible for others, is why impact has to be measured per workflow and per population, not as one company-wide number.

Do not substitute self-reported time savings for measured ones. In a 2025 randomized trial, experienced open-source developers took 19 percent longer on tasks when allowed to use AI tools, despite having forecast that the tools would make them faster.⁷ People are poor judges of their own speed with AI, in both directions.

Where the data comes from. Time studies for named workflows, and the CRM, ticketing, content and finance systems that already record throughput and quality.

Watch out for. An impact number without a named workflow behind it is a modeled estimate. Report it as one.

Treat this scorecard as your AI adoption KPIs: one metric per question, reported on the same cadence, grouped by level so a strong number at one level cannot hide a weak one below it. Every row can be produced without reading anyone’s conversations.

Ten rows is enough. If a metric cannot be tied to one of the four questions, it belongs in an appendix, not on the scorecard.

Here is one sales organization reported three ways. Each step adds a population, a use case, and finally a change in the work itself. The figures are illustrative.

The goal is not more AI activity. The goal is better work, and a report that cannot show the work is not a report on adoption.

1. Counting logins as usage. Opening a tool is not an interaction. Every vendor console can distinguish the two; make sure your definition of “active” does.
2. Optimizing for the adoption rate. It is the easiest AI metric to move, which is exactly the problem. Make the tool the default browser tab, run a launch campaign, add a reminder to the weekly all-hands, and weekly active users climb without anyone changing how they work. An organization can reach a high weekly active rate while most of that usage is quick lookups, one-line questions and drafts that get rewritten by hand. Before celebrating the rate, be able to say what work AI is being used for, how sophisticated those workflows are, whether effective practices are spreading, where people are struggling, and which workflows create value.
3. Trusting self-reported time saved. Survey answers about productivity are beliefs, not measurements, and the developer trial above shows how far the two can diverge.⁷ Use self-reports to decide where to run a time study, never as the result.
4. Measuring one vendor’s dashboard. Most employees who use AI at work brought at least one tool themselves.¹ A Copilot admin report describes Copilot, not adoption. Measure across every tool in use, including the ones nobody rolled out.
5. Skipping the baseline. Without 30 days of pre-rollout data on access, usage and the tasks people were already doing, every later number is a level, not a change.
6. Measuring individuals instead of aggregates. Ranking people by prompt count damages trust and rewards volume. Report by team, role and workflow, and keep individual conversations private by policy, not by promise.

Measurement only pays off if it feeds a loop: see what work people do today, find where AI could improve it, learn who is already using AI well, spread what they do, and check whether the organization improved. Four steps make that loop run.

1. Baseline before the rollout

Take 30 days of data before anything changes: who has access, who uses which tools and how often, and a short survey of the tasks each role spends the most time on. For the sales team in the examples above, the baseline would be 41 percent weekly active and a median of 45 minutes to prepare each discovery call.

2. Instrument each level

Wire the scorecard to its sources: the identity provider and license consoles for access, vendor admin reports and telemetry for adoption, sampled review or aggregate analytics for proficiency, and the systems of record for impact. Agree the privacy policy for the proficiency sample before collecting a single conversation.

3. Report on a cadence

A weekly adoption pulse for the rollout team. A monthly adoption and proficiency report for department heads, by team and by tool. A quarterly impact review per named workflow for leadership, with the baseline alongside the current figure. Each audience gets the level it can act on.

4. Intervene, then re-baseline

Every intervention, whether a documented workflow, training, or a change to which tools are approved, gets its own before-and-after. Once that sales team documented the call-preparation workflow and trained on it, weekly active rose from 41 to 72 percent and median preparation time fell from 45 minutes to 20 in one quarter. The next quarter’s baseline is 72 and 20, not 41 and 45.