Prevent incompatible definitions of active, adopted and revenue from becoming a false trend line or market comparison.
Most impressive percentages are numerator stories with the denominator hidden. 'Active' can mean a login, a completed task or a paying account. 'Adopted' can mean experimentation, any business use or production use in a defined function. OECD's innovation guidance explicitly separates firms engaged in innovation activity from firms that have introduced an innovation during an observation period [1]. The U.S. Census Bureau likewise reports current use and expected future use as different questions, and changed its AI-use wording from producing goods or services to any business function [2]. A denominator audit makes those boundaries visible before a comparison is made.
Write the fraction in full
Express every rate as numerator divided by eligible population. Record unit of analysis, inclusion and exclusion rules, geography, observation window, weighting, missing-response treatment and version of the question or event definition. For revenue, add gross or net, recognized or booked, recurring or total, currency and consolidation perimeter. For active use, add the qualifying event and inactivity window. A rate without these fields is a label, not a comparable measure.
Then ask who had the opportunity to enter the numerator. Invitation-based pilots should not use all employees as the denominator unless every employee was eligible. A product with paid and free tiers should not call paying organizations a share of users unless accounts, people and organizations are reconciled. Survey respondents are not automatically the target population; AAPOR recommends disclosing population, sample construction, recruitment, question wording and weighting so readers can judge the inference [3].
Separate pipeline states
Build a state ladder appropriate to the decision: aware, intends, evaluating, piloting, enabled, used, repeated, embedded, renewed. Do not force these into one adoption field. OECD's distinction matters operationally: activity can be ongoing, completed, paused or abandoned without an implemented outcome [1]. A vendor announcement may establish availability; a contract may establish commitment; telemetry may establish use; repeat cohorts and renewals say more about persistence.
For money, use a separate ladder: pipeline, contracted value, bookings, billings, recognized revenue, cash collected and retained recurring revenue. State transitions can reverse. A multi-year contract is not annual revenue, gross merchandise value is not net revenue, and a cumulative payout is not a current-period run rate. When only one state is available, report that state plainly rather than estimating the missing transitions from a familiar conversion rate.
Reconcile changes before drawing a trend
When a series jumps, check definition, questionnaire, eligibility and weighting before explaining behavior. The Census AI series illustrates why: its 2026 analysis states both the reference dates and the revised wording, and separately reports current use and six-month expectation [2]. A break created by broader wording should be annotated as a measurement change. It may coexist with real growth, but the available figures do not isolate how much comes from each mechanism.
Hypothetical example: a dashboard reports adoption rising from 20% to 45%. The earlier denominator is 1,000 licensed employees and the numerator is weekly task completion; the later denominator is 600 invited employees and the numerator is any monthly launch. Recomputed on a common definition, 180 of 600 eligible employees completed a task in the later month: 30%, not 45%. That still indicates movement, but it changes capacity plans and the claim the team can defend.
- Freeze the numerator event and denominator eligibility rule before comparison.
- Keep intent, pilot, use, repetition and renewal as distinct states.
- Annotate questionnaire and population changes as breaks, not footnotes.
Take it into the meeting
- Write every headline rate as a fully defined fraction.
- Use state ladders for adoption and revenue instead of one overloaded label.
- Recompute on a common definition before attributing a change to behavior.
Sources & boundaries
Source statements are attributed; the decision process is Signal Atlas analysis. Examples marked hypothetical are teaching inputs, not observed outcomes.
- Organizations may not retain historical event definitions or eligibility snapshots.
- A harmonized denominator can reduce comparability with published external benchmarks.
- Recomputed examples cannot recover nonresponse or selection bias without additional data.
- Concepts and definitions for measuring business innovation: Oslo Manual 2018OECD and Eurostat · Source publication: not established · Retrieved 2026-09-19
Innovation activity and implemented innovation are distinct statuses. Observation-period length affects classification of firms as active or innovative.
- Large Firms With at Least 20 Employees Biggest AI UsersU.S. Census Bureau · Source publication: 2026-05-26 · Retrieved 2026-09-19
BTOS asks separately about recent use and expected use in the next six months. The survey wording changed from use in producing goods or services to use in any business function.
- Best Practices for Survey ResearchAmerican Association for Public Opinion Research · Source publication: not established · Retrieved 2026-09-19
Survey interpretation depends on disclosed population, sample, mode, recruitment, wording and weighting. Low response can omit types of respondents and bias estimates.