Seventy-one percent of employees say AI tools help them feel more productive. Among the heaviest users, that climbs to 93 percent. Ask the same people whether their workload is reasonable for their role and the answer barely moves: 72 percent of power users against 69 percent of non-users โ a difference of three points (Culture Amp, 2026).
Those two findings come from the same survey, the same employees, the same twelve months. One curve rises steeply with usage intensity. The other is flat.
That is not a measurement error. It is a management artefact, and it means the capacity your AI rollout created is being reabsorbed inside the same roles rather than redeployed into anything you chose.
Two Curves From One Dataset
Culture Amp's first AI at Work benchmark draws on 342 surveys run by 292 organizations on its platform, representing roughly 112,000 employees and 343,000 individual responses collected in the twelve months to July 2026. Respondents span 16 industries and all major regions; individual contributors are about two-thirds of the sample. Employees were standardised into four usage tiers โ non-users, light users (once or twice a week), daily users, and power users (multiple times a day) (Culture Amp, 2026).
Within those tiers, almost every experience measure improves with usage. Risk literacy rises from 67 percent among non-users to 95 percent among power users. Motivation to go beyond what the role requires elsewhere runs 72 percent among the heaviest users against 57 percent among non-users. Agreement that AI has helped them learn and grow averages 66 percent globally, dragged down by non-users at 27 percent.
One item refuses to follow. On "Generally, my workload feels reasonable for my role," favourability ranges from 69 percent among non-users to 71โ72 percent among everyone who uses AI at all. Culture Amp's own reading is blunt: initial adoption slightly improves workload perception, and increased frequency of use yields no additional relief (Culture Amp, 2026).
Felt productivity runs at 71 percent across the sample and 93 percent among the heaviest users. Felt workload relief, across those same tiers, is worth three points.
Why Your Adoption Dashboard Cannot See Flat Workload
Most mid-market AI programmes are instrumented for the wrong variable. Licence counts, weekly active users, prompts per seat, tool-level time-saved estimates โ every one of those rises when adoption rises. None of them changes when the freed hour is quietly consumed by more of the same work.
The reabsorption is invisible by construction. An employee who finishes the Thursday report in forty minutes instead of ninety does not report a lighter week; they report a week with more in it. Culture Amp makes this point against itself, and fairly: plates adjust as people get more efficient, and that is normal. Nobody expects a quilter with a quilting machine to produce the same number of quilts per year as one working by hand.
The problem is not that the plate refilled. It is that nobody decided what went on it.
That is where the Culture Amp finding stops being a curiosity and starts being a budget question. If your business case assumed recovered hours would show up as relieved capacity, headcount avoidance, or faster cycle time, and your own engagement data says perceived workload is flat across every usage tier, then the business case is unverified โ not disproven, unverified. You have no instrument pointed at the thing you promised.
Where the Recovered Hours Actually Go
Two other 2026 datasets map the destinations, and neither is strategic.
BCG's fourth annual AI at Work survey โ 11,749 workers across 14 markets โ found that 42 percent of regular frontline users report saving at least a full workday per week through AI, while 66 percent say they get limited or no guidance on what to do with that time, and more than half do not redirect it into strategic work. In the same sample, 47 percent now report spending more time managing and directing AI than doing the work themselves, and 72 percent say AI has already considerably changed the skills their role expects (BCG, 2026).
Read those two numbers together. The hour is real. The decision about the hour does not exist. In its absence, the hour defaults to the nearest available demand โ which, increasingly, is the AI itself.
BambooHR's survey of 1,608 US desk workers put a stopwatch on that defaulting. Workers average 87 minutes a day inside AI tools; 42 percent of that time goes to troubleshooting errors and iterating on prompts against 35 percent on work that actually advances their own workload (BambooHR, 2026). More AI time goes to fixing AI than to producing with it.
So the capacity is not vanishing into idleness. It is being spent โ on supervision, correction, and the new coordination surface the tools created. That work is genuine. It is simply work nobody budgeted, named, or assigned to a role.
The Guidance Vacuum Is the Mechanism
The same benchmark that found flat workload also found what is causing it, and it is not employee resistance.
Eighty-five percent of employees say their organization actively encourages exploration and experimentation with AI. Only 58 percent say their leaders have clearly explained how AI connects to where the company is headed โ the lowest-scoring item in the entire benchmark, and a 27-point gap from the encouragement item. Closer to the work, only 60 percent say their manager shares examples of how AI can support their goals, and support from systems and processes scores 64 percent (Culture Amp, 2026).
Amy Lavoie, Culture Amp's VP of People Science, draws the conclusion from the productivity figure rather than the sentiment ones: most executive teams would read seven in ten employees feeling more productive as a return on investment, but with workload scores steady it is an unanswered question. The organizations that see genuine return, she argues, "won't be the ones with the highest adoption rates. They'll be the ones that decided, explicitly, what they were going to stop doing with the time it gave back" (Culture Amp, 2026).
This is a workforce that has done its part. The missing artefact sits above them: a written decision about what stops.
A Second Signal in the Same Release
One adjacent finding is worth holding carefully. Employee awareness of internal career opportunities fell 10 percentage points since July 2025 โ the largest single-year movement in Culture Amp's benchmark data โ while satisfaction with career development stayed at 66 percent, where it has sat since 2021. Employees are not less satisfied with the development they receive; they are less clear about what comes next.
The caveat matters: Culture Amp states those career figures come from its wider engagement benchmark, a larger population than the 112,000-employee AI sample. Do not attribute them to the AI cohort. Treat them as the surrounding weather, not the same reading.
The Honest Counter
Three limits, stated plainly.
This is perception, not output. Every figure here is employees reporting how they feel. Culture Amp says so explicitly โ each benchmark result is a swath of employees telling you how they feel about something, not a measure of output or hours. A flat workload score is consistent with real capacity gains that people simply absorbed without noticing.
The sample is self-selecting. Organizations that run an AI survey are, by definition, further along than those that do not. North America accounts for 58 percent of companies using the template, technology leads by industry, and organizations of 1,000-plus make up about a third of users. Culture Amp's framing is correct: read this as a portrait of AI-engaged organizations, not of every employer. (The press release body also cites "123 organizations" in one paragraph against 292 in its About the Data block; the methodology statement is the one to quote.)
The alternative explanation is timing, not failure. NBER working paper 35684 โ Babina, He and Jiang โ ties early AI productivity gains to the accumulation of organization capital, the durable firm-specific knowledge built through learning-by-doing, rather than to tools or AI headcount as such (NBER, 2026). The paper body is gated, so no effect size should be attached to it. But the direction is a fair counter: a year of flat workload may be the cost of building process knowledge that pays later. That reading is only available to firms that are actually codifying something. It is not available to a firm with 400 individually optimised workflows and no record of any of them.
What to Decide Before the Quarter Closes
Four moves. None of them is a new licence.
- Write the stop-doing list, per team, with names on it. Not "reduce manual work" โ three specific recurring outputs that no longer get produced after this quarter, and who confirmed they are no longer needed. If nothing on the list is uncomfortable, the list is decorative.
- Name the destination for the recovered hours before you claim them. An hour released without a named destination is an hour absorbed. Pick one per role: a development objective, a cross-team project, an internal move, or deliberate slack. "Strategic work" is not a destination.
- Add one workload question to your next engagement pulse and segment it by AI usage tier. You can reproduce Culture Amp's finding inside your own company in one survey cycle. If your heaviest users report the same workload as your non-users, you have measured reabsorption directly โ in your own data, with your own roles attached.
- Instrument the role, not the tool. The stop-doing decision needs a view of what each seat actually contributes, which competencies the recovered hours could build, and which internal move the freed capacity makes possible. That is role-level data, not adoption data. Scovai's performance and succession layer measures competency per cycle and maps internal mobility and succession readiness from it, if you want a worked reference for what that instrumentation looks like.
The benchmark's real finding is not that AI failed to reduce workload. It is that 112,000 employees were handed a productivity gain and nobody above them specified what it was for.
Your heaviest AI users already found the time. The only question left for this quarter is whether you decide where it goes, or whether their inbox decides for you.
