Two percent.
That is the share of workers whose observed activity shows AI genuinely integrated into how the work gets done โ not sampled, not used occasionally, but built into the flow of the job. The number comes from ActivTrak's Productivity Lab, which staged AI adoption across 120,620 workers by what those workers were seen doing, rather than by what they were licensed for or what they reported on a survey. Twenty-seven percent reached research assistance. Fourteen percent reached task execution. Two percent reached workflow integration (ActivTrak Productivity Lab, 2026).
Your AI adoption dashboard almost certainly reports a much larger figure. It is not lying. It is answering a different question โ one you stopped noticing you were asking.
Two Measurements, One Week, Two Realities
The timing is the tell. In the same week, Gallup reported organizational AI adoption jumping six points (Gallup, 2026), and ActivTrak published the 2% workflow-integration figure (ActivTrak Productivity Lab, 2026).
Both are credible. They are not in conflict, because they are not measuring the same thing. Gallup asks people and organisations what they are doing. ActivTrak observes what tools are actually open, in what sequence, at what frequency, over what stretch of a working day. Self-report captures intent, awareness, and access. Behavioural telemetry captures habit.
I want to be careful here, because this is a cross-method comparison rather than a head-to-head study, and treating it as a contradiction would be sloppy. Different samples, different instruments, different definitions of the word "adoption." The point is not that one dataset debunks the other. The point is that most operations teams are reporting the first kind of number to a board that is funding the second kind of outcome.
That gap is where AI budgets quietly go to die. Access is scaling fast. Work redesign is not. And every dashboard built on seats, logins, and monthly active users will show the first curve and stay silent about the second.
What the Three Stages Actually Distinguish
The staging model matters more than the headline percentage, because it gives an operator something to instrument.
Stage 1 โ Research assistance (27%). The worker asks the model questions. Drafting, summarising, looking things up. The job is unchanged; a reference step got faster.
Stage 2 โ Task execution (14%). The model performs a discrete unit of work: writes the first version, extracts the fields, produces the analysis. The task is now shared. The workflow around it is untouched.
Stage 3 โ Workflow integration (2%). AI is a load-bearing step in a process that has been re-sequenced around it. Handoffs moved. Approvals moved. The work is not the old work done faster; it is a different shape.
Only Stage 3 changes the cost structure
This is the operational crux. Stages 1 and 2 produce time savings distributed in fragments across individuals โ twenty minutes here, an afternoon there. Real, but unbankable. Nothing in the P&L moves because nothing in the process moved; the saved minutes get reabsorbed by the same queue that generated them.
Stage 3 is where cycle time, headcount plans, and unit economics respond, because the process itself has been rebuilt. That is the stage your business case was underwritten against. It is also the stage two percent of workers have reached.
The mechanism behind that gap is well documented. Deloitte's 3,235-leader study found the response to AI skewed heavily toward teaching people the tools rather than restructuring the work: 53% educate the workforce to raise AI fluency, while only 30% reimagine the organisation around new AI patterns and 33% redesign career paths (Deloitte, 2026). Everyone trains. Almost nobody redesigns. Then everyone measures training completion and calls it adoption.
Why Seat Counts Still Pass for AI Adoption
Login metrics survive for structural reasons, not stupid ones.
They are free โ the licence platform emits them with no instrumentation work. They move early, which makes them satisfying in the first two quarters of a rollout. They are unambiguous, so nobody argues in the steering meeting. And they rise monotonically, which is exactly what a sponsor under pressure wants a slide to do.
Changed-work metrics have none of those properties. They require someone to define the workflow, baseline it, and then admit when it has not moved. They are contestable. They can go flat for two quarters while the redesign is being fought over. Reporting them is a political act in a way that reporting seat counts is not.
So the incentive is clean and pointed the wrong way: the easiest number to produce is the one least connected to the outcome you promised. That is not a measurement failure. It is a governance failure wearing a measurement costume.
There is a second-order cost that rarely gets named. Once a seat-count metric is established as the definition of AI adoption, it starts steering behaviour. Budget flows toward whatever raises it โ more licences, broader rollout, another department onboarded โ because that is the number under management. The workflow redesign that would actually move the business case competes for funding against a metric that structurally cannot reward it. Teams do not choose distribution over redesign on the merits. They choose it because distribution is what the dashboard pays for.
The Fair Reading of Two Percent
Three caveats, because the number will be misused otherwise.
It is not evidence that AI does not work. It is evidence about diffusion, not efficacy. A 2% integration rate says the practice is rare, not that it is unproductive. The organisations inside that 2% are the interesting population, and nothing in this data suggests they are wasting their time.
The trajectory is real. The Stage 3 cohort grew from 1,739 workers in Q1 2026 to 2,369 in Q2 โ a 36% quarterly increase (ActivTrak Productivity Lab, 2026). Small base, steep slope. This is early-diffusion behaviour, not stagnation.
Behavioural telemetry has its own blind spots. It sees tool activity, not judgment. A worker who uses AI once to redesign a process and then runs the improved process without touching a model again is doing exactly what Stage 3 is supposed to describe, and will not necessarily register as such. Observed-behaviour data is a better proxy than self-report. It is still a proxy.
What survives all three caveats is the asymmetry between the two curves. Access is being provisioned at a pace that behaviour is nowhere near matching, and only one of those two things appears on the reports going to leadership.
Stopping at Stage 2 Has Its Own Bill
There is a further reason not to treat Stage 2 as a comfortable resting place, and it comes from the same instrument.
ActivTrak's behavioural comparison of roughly 10,584 employees, measured 180 days before and after AI adoption, found email activity up 104%, chat up 145%, business-tool usage up 94%, productive hours up 5% โ and average uninterrupted focus sessions down 9%, from 14 minutes 23 seconds to about 13 minutes (ActivTrak, 2026).
Read that as an operations statement. Work got faster and denser. It also got more fragmented, and the fragmentation lands precisely on the deep-focus capacity that process redesign requires. Stage 2 without Stage 3 buys you more transactions per hour and fewer uninterrupted stretches in which anyone can think about why the process has that many transactions in the first place.
Meanwhile the deployment curve keeps climbing. Microsoft's 2026 Work Trend Index reports active agents growing 15x year over year across its ecosystem, 18x in large enterprises, while only 26% of AI users report leadership aligned on AI strategy (Microsoft, 2026). Distribution is outrunning direction, and the dashboard is measuring distribution.
Rebuild the Dashboard Around Changed Work
Four changes, ordered by how cheap they are.
Retire seat count as a headline metric. Keep it as a cost input where it belongs โ next to the licence spend, not next to the outcomes. The moment it appears on an outcomes slide, it starts substituting for one.
Name one workflow and stage it honestly. Pick a single process with a real owner. Classify it: research assistance, task execution, or workflow integration. Most teams discover their flagship AI story is a Stage 1 with good anecdotes attached. That discovery is the deliverable.
Instrument process facts, not activity facts. Cycle time end to end, handoff count, rework rate, exception volume. These are the things that move only when the workflow moves, which is exactly why they are harder to game and worth more than any usage chart.
Gate seat expansion on a Stage 3 proof. Hold further licence growth until one named workflow demonstrably clears workflow integration. This is the only mechanism on the list with teeth, because it makes the measurement consequential rather than decorative.
Add one attention check alongside them: track uninterrupted focus-session length before and after each rollout. If AI adoption is fragmenting the workday, you want that in the same report as the throughput gains, not discovered a year later.
One Decision This Quarter
Take the workflow your AI investment is publicly justified by. Stage it against the three-stage model with someone who has no stake in the answer. If it is Stage 1 or Stage 2 โ and on this evidence it most likely is โ freeze further seat expansion and spend the next quarter moving that single workflow to Stage 3 instead.
That exercise is cheap, it takes about a week, and it produces the one artefact most AI programmes lack: an honest baseline. Everything after it โ the business case, the expansion plan, the argument about headcount โ gets more accurate for having a real number underneath it.
The question your AI adoption dashboard answers is how many people can use AI. The question your business case rests on is how much work is now done differently. Only one of those has ever shown up in the P&L, and it is not the one on the slide.
