Most measurement of AI at work counts the wrong thing. Seats bought. Licences activated. Logins per week. Those numbers go up and everyone declares transformation.
ActivTrak’s Productivity Lab tried counting something harder: whether the work itself changed. It tracked behavioural data from 120,620 workers across 1,009 organisations, quarter by quarter, from Q4 2025 through Q2 2026, and sorted every AI user into one of three stages based on what they actually did with the tool.
Twenty-seven per cent reached Stage 1, consulting AI for questions and information without handing it any work. Fourteen per cent reached Stage 2, using it to draft, generate and complete routine tasks with a human validating everything. Two per cent reached Stage 3, where AI is embedded across multiple steps of a workflow and the employee sets the intent and approves the output.
Adoption is sticky. Progression is not.
The stickiness number is genuinely impressive. More than 82% of AI users kept using AI from one quarter to the next — this is not a tool people try and abandon. Retention held at the deeper levels too: roughly 80% of Stage 2 users and 70% of Stage 3 users stayed at their level quarter over quarter.
What almost nobody did was move up. The Stage 3 population grew from 1,739 users in Q1 to 2,369 in Q2 2026 — a 36% increase that still leaves it at 2% of the workforce studied. Habits formed fast and then set.
The productivity gap is smaller than the discourse suggests. AI users averaged 6 hours 34 minutes of productive time daily in Q2 2026 against 6 hours 17 minutes for non-users. Seventeen minutes. Real, but not the step change anyone is budgeting for.
The report argues with its own CEO
Here is the part worth sitting with. The July release read the deep-integration data optimistically: Stage 3 users hold 69% healthy utilisation while logging a shorter average workday span — 6 hours 46 minutes against 7 hours 5 minutes for Stage 2 — which it framed as “work efficiency rather than overwork.”
Writing in Fortune on 16 August, ActivTrak chief executive Heidi Farris read the same dataset the other way. Work-health metrics rise as people move from little AI use to regular task-level use, peaking at 75% healthy utilisation — and then fall roughly five points once AI is embedded in workflows, to a level she describes as statistically indistinguishable from people who barely use AI at all. Her conclusion: the optimum for most employees sits in the middle, not at the top.
Our take: the shape of this curve matters more than the headline. If work health peaks at Stage 2 and gives back its gains at Stage 3, then “deeper AI integration” is not a ladder you climb until you win — it is a trade you make. Neither source publishes a statistical test behind “indistinguishable,” and one is a vendor measuring its own customers with its own definitions. Treat the direction as a hypothesis worth testing on your own team, not as settled science. But if your plan for the next two quarters is “get everyone to Stage 3,” the company selling you the measurement tool is quietly telling you that is the wrong target.
What to watch
- Whether Stage 3 growth compounds or stalls. 1,739 to 2,369 in one quarter is a fast rate on a tiny base. Two more quarters tell you whether workflow integration is spreading or capped.
- Whether the 17-minute gap widens. ActivTrak says the margin “stays relatively fixed across periods.” A fixed gap is the single most deflationary fact in the report.
- Independent replication. Every figure here comes from one vendor’s telemetry on its own installed base, and the release explicitly warns against comparing these numbers to its own prior publications. Until someone outside the category measures stage progression, this is one instrument’s reading.
- Your own Stage 2 population. If the middle is where the gains are, the useful internal question is not how many people have access — it is how many have moved from asking AI things to giving AI work.
The uncomfortable summary: four out of five people who pick up AI keep using it, and roughly one in fifty rebuilds anything around it. Access was the easy part, and most organisations have finished it. Redesigning the work is the actual project, and by this measurement almost nobody has started.