What does it mean to measure AI adoption?

Define the construct with the USE frame

The USE frame has three parts: Unit, Scope, and Evidence. It forces a programme team to define the measurement before selecting a dashboard metric.

The U.S. Census Bureau Business Trends and Outlook Survey AI supplement provides a useful primary-source example of construct discipline. Its current-use question asks whether a business used AI in any business function during a specified recent period, defines AI, and provides response options including “do not know.” That measure concerns business-level recent use; it does not by itself show frequency, quality, competence, impact, or safe operation.

The frame should therefore produce a sentence that can be tested: “Among 240 employees eligible to use the approved assistant, what proportion completed at least one qualifying finance-analysis task with a retained work artifact during the four weeks ending 30 September?” Every word changes the result. Removing “eligible” distorts the denominator. Replacing “completed task” with “logged in” changes the construct. Removing the window makes trends uninterpretable.

• Unit: choose the entity being counted: employee, team, business, workflow, task, or transaction. Do not switch units in the denominator.
• Scope: name the AI capabilities, business functions, eligible population, and observation window. State whether embedded AI features count and whether experimental use is included.
• Evidence: set the minimum observable event, such as a verified workflow record, approved-tool telemetry joined to an eligible user, or a structured sample of completed work.

Sources: [1]

Keep four adoption states separate

Report four states as a sequence rather than a single maturity score. Access means the person or workflow can use the approved capability. Trial means at least one qualifying use occurred. Repeat means qualifying use happened in more than one defined period or task. Embedded means the use is part of a documented workflow with ownership, review, and an evidence trail.

These states answer different operational questions. Access reveals enablement coverage. Trial reveals initial behavior. Repeat reveals persistence. Embedded reveals whether the work practice has moved beyond isolated experimentation. None of them alone demonstrates quality, productivity, risk control, or business value.

The Census Bureau BTOS methodology explains the survey's panel design, weighting, response treatment, and publication approach. It is a reminder that a percentage is inseparable from sampling and measurement design. An internal adoption dashboard should likewise disclose source coverage, response or logging gaps, exclusions, question changes, and whether comparisons remain valid after a tool or definition changes.

Sources: [2]

Build a metric card before building a dashboard

For each adoption metric, create a card with nine fields: decision served, construct, numerator, denominator, window, data source, missingness rule, confounds, and interpretation limit. The card is the evidence artifact that keeps the chart tied to a decision.

For example, a “repeat task use” card might use a numerator of eligible employees with verified AI-assisted artifacts in at least two separate weeks, divided by employees eligible for the approved tool throughout the full four-week window. Terminated accounts, late joiners, service accounts, and employees on extended leave receive explicit treatments. A licence assigned without a task artifact counts as access but not repeat use. A copied AI disclosure without a retained deliverable fails the evidence rule.

Add at least one counter-metric. If repeat use rises, inspect exception rates, unsupported-claim findings, privacy escalations, rework, or review time. The NIST AI RMF Core says measurement should use appropriate quantitative, qualitative, or mixed methods, document uncertainty and limitations, and connect measurement to mapped risks and management actions. Adoption should therefore be read alongside the quality and risk evidence relevant to the workflow, not celebrated in isolation.

Sources: [3]

Worked example: a four-week operations cohort

This synthetic example starts with 120 operations employees. Ten lack access for the whole window, five join during week three, and five are service or test accounts. The defined eligible population is therefore 100 people. Tool logs show 72 active users, but only 58 have a qualifying work artifact. Of those, 41 have verified artifacts in at least two separate weeks, and 19 follow the documented review step for every sampled task.

The metric card reports access as 110 of 120 human employees, trial as 58 of 100 eligible employees, repeat as 41 of 100, and embedded use as 19 of 100 under the programme's evidence rule. It does not average those figures into an “adoption score.” A companion sample finds that 7 of 30 reviewed artifacts contain an unsupported factual claim requiring repair. The steering conclusion is not “adoption reached 72%” from active-user logs. It is: access is broad, evidence-backed trial is narrower, repeated use is narrower again, and the review control is not yet consistently embedded.

The team records two confounds. A mandatory month-end process created unusually high activity in week four, and one workflow generated multiple logs per task. It keeps person-level and task-level measures separate and schedules the next review using the same definitions.

Interpret the result without overclaiming

An adoption measure supports a bounded decision: where to investigate access, workflow fit, support, or control operation. It does not establish causation. Higher use may reflect mandatory policy, easier tasks, seasonality, tool changes, or logging coverage. Lower use may reflect inapplicability rather than resistance. Always pair the metric with qualitative evidence from the affected workflow and publish the definition beside the number.

Two Prune can provide task-bounded assessment evidence about how participants use AI during realistic work simulations, including framing, evidence navigation, orchestration, verification, judgment, and communication. That evidence may inform a broader capability programme, but it is not workforce telemetry, an enterprise adoption census, a productivity claim, a benchmark for all employees, or an employment decision.