Counting AI users tells you who was already ahead
A usage rate says how many people touched a tool. It cannot say who they were, what they checked or what changed in the work. This essay lets you handle the numbers yourself and ends with an audit for your next AI usage report.
Counting AI users tells you who was already ahead. That sounds like a provocation until you read the survey behind the kind of number most organisations now put on a dashboard.
What a usage rate counts
In May 2026 the University of Konstanz surveyed 1,105 employees in Germany. The share using AI in their work rose from 35 to 38 percent in a year.[1] It is a clean number, and it is the kind a dashboard produces by default: one numerator, one denominator, one trend line. It answers a narrow question. How many people touched an AI tool in the period?
It does not say whether the tool suited the task, whether the output was any good or whether anyone's work changed. The essay on how you know an AI tutor taught anyone anything made the same point about learning: usage is where evidence starts, not where it ends.
There is also the question of what ahead means. The people who adopt first are rarely a random slice of the workforce. They may hold the roles, the confidence and the spare attention that a new tool rewards, which lets an early adoption curve read as progress while it mostly describes a starting position. A count taken at one moment cannot separate the two.
Foundations · What does a usage rate measure?
A usage rate is the number of people who did something, divided by the number of people you counted. It records presence, not quality. Whether someone counts as a user depends on a definition: opened a tool, used it weekly, used it on a task. Change the definition and the rate moves without anyone behaving differently. Ask for both the numerator and the denominator.
The same release adds a detail that a licence count would miss. Only 55 percent of AI users said the tool they use most often had been officially introduced by their employer.[1] The other 45 percent did not say so. Multiply the two published figures and about 21 percent of all employees use an officially introduced tool as their main one, against 38 percent who use AI at all. That is arithmetic on two reported numbers, not a separate finding, but it shows the gap between what an instrument sees and what people do.
Figure 1 above puts the two instruments side by side: a volume metric as a funnel and a judgement metric as a loop.
Konstanz found 56 percent of highly educated employees using AI against 21 percent of the rest. Drag the share of highly educated staff to see the overall rate those two published rates would produce for your organisation. The 49 percent mark is arithmetic only: the release gives the group rates, not the split of the sample.
share*56/100 + (100-share)*21/100. A first approximation to show proportion, not a forecast.From 35 to 38 percent is three points. Of every hundred employees, 65 were not using AI a year earlier, so three newcomers reach about 4.6 percent of them, roughly one in 22. Growth in users, 3 of 35, is the larger one in twelve. Move the two rates to see the converted share.
(after-before)/(100-before)*100. A first approximation to show proportion, not a forecast.A single survey of 1,105 people at 38 percent carries a margin of about 2.9 points either way, assuming random sampling. Change the sample size or the rate to see how much of a three point move could be noise.
1.96*sqrt((p/100)*(1-p/100)/n)*100. A first approximation to show proportion, not a forecast.Zagami reports rho of minus 0.357 between achievement and active AI engagement. Rho squared gives the share of rank variation the two measures have in common, about 13 percent. Rule of thumb bands: under 0.1 negligible, under 0.3 small, under 0.5 moderate, above that large. Move the marker to compare.
rho*rho*100. A first approximation to show proportion, not a forecast.The gap inside the average
The Konstanz team also split the figure by education. Among employees with a high level of education, 56 percent used AI at work. Among those with lower attainment, 21 percent did.[1] The headline is an average of two very different groups, and an average moves when the mix changes, not only when behaviour does.
That is a selection effect in plain sight. If adoption starts among the people best placed to use new tools, a rising count first reports who was already ahead. Set the share of highly educated staff in Figure 2. The output is the rate a dashboard would report if the two published rates applied to your people, with the formula blended = share times 56 plus (1 minus share) times 21, in percent. It assumes two education groups with the published rates, the same in every organisation.
Two organisations with identical behaviour inside each group, one with 20 percent highly educated staff and one with 60 percent, would report 28 and 42 percent. The dashboard would call the second one fourteen points more adopted. It is mostly a headcount of who works there. A 38 percent headline would imply about 49 percent highly educated staff under this arithmetic, which is a reading of the two group rates, not a figure the release gives.
None of this makes the 56 percent figure wrong. It makes it a different kind of statement. A rate for a group describes the group, and a rate for a whole organisation describes the group plus its composition. Before you compare your number with a benchmark, ask what mix of staff produced each one. A benchmark from a workforce unlike yours compares workforces.
Foundations · What is a selection effect?
A selection effect appears when the people in a group differ from everyone else before anything is measured. Early AI users are not a random draw from the workforce. If they differ in education, role or confidence, then comparing them with non users compares the people as much as the tool.
Three points, read slowly
From 35 to 38 percent looks like movement. Read slowly, it is three points. Of every hundred employees, 35 used AI a year earlier and 65 did not. Three more are counted now, so about 4.6 percent of the earlier non users, roughly one in twenty two, have been reached. That assumes the change is made of newcomers alone. People who stopped, and plain sampling noise, would hide inside the same three points. Figure 3 lets you move the two rates.
The other reading is that the number of users grew by about one in twelve, 3 of 35. Both are true. The first tells you how much of the remaining workforce has been reached, and it is the smaller. For scale, a single survey of 1,105 people carries a sampling margin of roughly three points either way at this level, assuming random sampling; Figure 4 shows the margin for other sample sizes and rates. The release itself calls the rise slight.[1]
The stranger direction
Konstanz measures workers. Jason Zagami's survey covers 484 students at one Australian university, so what follows speaks about students and should not be carried over to employees.[2] Higher achievers reported less active engagement with generative AI. The rank correlation between achievement and active AI engagement was minus 0.357, computed on the 469 respondents with usable grade data.[2] Open responses describe students checking AI outputs against their own judgement and keeping them subordinate to it.[2]
Read that against a usage count. Someone who uses AI for one narrow step, checks every output and moves on looks like a light user. That is also what a careful organisation would want. A volume metric ranks this person as a laggard. The open responses fit that reading, but a correlation cannot establish it.
For an organisation the lesson concerns instruments, not students. If careful use can look like light use in one population, the same distortion is possible in another, and the only way to find out is to ask what people checked. The first reading of a low usage score should be a question, not a verdict.
Time based measures fail in the same way. Time spent was the first number I stopped trusting in my own research, because it rewarded the slowest route.
Foundations · What is a rank correlation?
Spearman's rho ranks every person on two measures and asks whether the two rankings move together. It runs from minus 1 (opposite order) through 0 (no consistent order) to plus 1 (identical order). It ignores the size of the gaps between people, and it says nothing about cause. Figure 5 shows how much rank variation a given rho implies the two measures share.
A funnel and a loop
Put the two readings side by side, as Figure 1 does. A volume metric is a funnel: everyone enters, some cross a threshold, one number comes out. A judgement metric is a loop. Someone decides whether AI suits the task, uses it selectively, checks the output and keeps, changes or discards the result. The funnel records one crossing. Whatever happens inside the loop, where the quality is made, never reaches the count.
Judgement can be sampled without surveillance. Ask a team to bring five recent outputs and to say, for each one, where AI entered, what was checked and what changed before it went out. A short review like that answers the second, third and fourth audit questions in an afternoon. It will not produce a dashboard number, which is the point.
What this means in practice
Usage is where measurement starts, not where it ends. None of this argues against counting. A count is cheap, it trends, and it tells you whether a rollout reached anyone. It argues against presenting that count as the quality of adoption. Split it by who, ask where in the work AI entered, sample what people checked and name one change in the work. A dashboard that answers only the first of those is a headcount with a trend line.
The same caution applies when a number becomes a target: The Day Your Target Starts to Bind shows what happens when a figure starts driving behaviour instead of describing it. And once the work is handed to an agent, the checking question moves with it, as The Job Moved from Prompting to Delegating explains. The diagnostic that opened this page is itself a small learner model, the distinction drawn in Adaptive Learning Is Not Personalisation.
Which of the four questions could your dashboard answer today?
Exercise · Try it at work
- Self report. Both surveys ask people what they do. Reported use is not logged use, and reported checking is not observed checking. - Cross sectional. Each survey is a snapshot. Order and cause cannot be recovered from one moment in time. - One country, one university. Konstanz covers employees in Germany. Zagami covers students at a single Australian university, so the student finding says nothing directly about workers. - Education as a proxy. Education stands in for skills, role, task mix and access. The 56 against 21 percent gap does not say which of them matters. - No causal claim. Nothing here shows that education causes adoption, or that AI use changes achievement. - Publication and coverage bias. The Konstanz figures come from a news release, which foregrounds headline numbers. Surveys that find little tend to be published and covered less often. - Absence of evidence. Neither source measures the quality of work. That is absence of evidence about quality, not proof that heavy users produce worse work or that careful users produce better work.
- A usage rate counts who touched a tool. It does not show who they were, what they checked or what changed.
- In the Konstanz survey, 56 percent of highly educated employees used AI against 21 percent of the rest, so a headline average partly reports the mix of staff.
- The rise from 35 to 38 percent is about one in twenty two of the earlier non users, near the size of one survey's sampling margin.
- Among the students Zagami surveyed, higher achievers reported less active AI engagement (rho minus 0.357) and described checking outputs. A volume metric would rank them last.
The follow-up quiz arrives in three days.
Spaced retrieval is the best evidenced way to keep what you just read. Leave an email and you will get three recall questions on this essay across three days (what a usage rate counts and cannot show; why a headline rate can rise without anyone changing how they work; what a rank correlation of minus 0.357 supports and does not), plus the next essay when it publishes.
References
Primary and peer reviewed sources only. Links go to publisher pages.
- University of Konstanz (2026). The use of AI at work: slow, informal and unequal. EurekAlert! news release 1144563, 18 September 2026. eurekalert.org/news-releases/1144563
- Zagami, J. (2026). Guarded adoption of generative AI in higher education: high-achieving students, successful-student identity, and epistemic agency in a single-university mixed-methods survey. International Journal of Educational Technology in Higher Education, 23, 47. link.springer.com/article/10.1186/s41239-026-00625-6