Directors and above save nine hours a week with AI. Individual contributors save four. That 2.25x spread comes from SHRM's Navigating AI in the Workplace: 2026 report, fielded across 4,065 workers who actually use AI at work, and the difference across levels is statistically significant (SHRM, 2026). The aggregate headline โ about six hours saved per week, nearly a full workday โ is real. It is also an average that no single tier of your org actually experiences.
This matters because of who runs the pilot. AI rollouts in 50โ500 FTE companies almost always start with the people closest to the decision: the ops leader, the department heads, a couple of senior managers. They measure their own hours saved, extrapolate, and build a business case for the whole headcount. The SHRM data says that extrapolation overstates front-line gains by roughly a factor of two โ and that AI time savings by job level run in the opposite direction from the "AI lifts the floor" assumption most rollouts are budgeted on.
The Dividend Compounds at the Top
Here is the full gradient from the SHRM sample: directors and above save 9 hours per week, managers 7, individual contributors 4 (SHRM, 2026).
The intuitive story about AI has been the opposite. Junior people benefit most; the tool substitutes for experience; the floor rises faster than the ceiling. That story has real empirical support in narrow task studies, which is why it spread. But at the level of an actual week of actual work, it inverts. The people with the most discretion over how they spend their time are the people extracting the most time back.
The gradient is not a rounding artifact, and it does not disappear when you net out AI's costs. The same report finds directors spend roughly 6 hours a week correcting AI output โ the highest rework burden in the sample โ against about 3 hours for individual contributors. Net of rework, directors still clear roughly +3 hours a week and ICs roughly +1. The dividend shrinks at every level. It stays regressive at every level.
Why the Gap Is Structural, Not Motivational
The reflex explanation is that senior people are simply better at using AI, so the fix is training. The data points somewhere less flattering and more actionable.
Start with exposure. AI touches 63% of directors' work, 50% of managers' work, and just 34% of individual contributors' work (SHRM, 2026). A tool cannot save you time on work it does not touch. Director-level work โ drafting, summarizing, synthesizing, restructuring documents, preparing analysis โ is disproportionately the kind of unstructured language work that current models are good at. Front-line work in a mid-market operation is disproportionately process-bound: it runs inside a system of record, follows a defined sequence, and is gated by handoffs. Those steps do not get faster because someone has a chat window open beside them.
Second, discretion. A director who saves an hour can reallocate it themselves. An IC who saves an hour inside a queue-driven process usually cannot โ the queue sets the pace, and the saved time either evaporates or reappears as waiting.
Third โ and this is the part most rollouts under-fund โ the front-line saving requires workflow redesign that the senior saving does not. ActivTrak's behavioral telemetry across 120,620 workers is blunt on how rare that redesign is: 27% of workers have reached research-assistance use, 14% task execution, and only 2% have reached genuine workflow integration (ActivTrak, 2026). That measurement is derived from observed activity rather than licenses, logins, or self-report โ which is exactly why it disagrees with the survey picture. Gallup reported organizational AI adoption jumping six points in the same week ActivTrak published the 2% figure (Gallup, 2026). Access is scaling fast. Redesign is not.
Individual contributors are the tier whose gains depend most on the redesign that has not happened. That is the mechanism. Not motivation, not prompt skill.
Run the AI ROI Math on Your Own Business Case
Take a 200-FTE company: 15 people at director level and above, 35 managers, 150 individual contributors. Model it the common way โ pilot with the senior group, observe roughly 9 hours saved, apply that across the org โ and you project about 1,800 hours a week.
Apply the SHRM gradient instead: 15 ร 9 + 35 ร 7 + 150 ร 4 = 980 hours. Then net out rework and you are closer to 300 hours of genuinely reclaimed capacity. The naive model is not 10% optimistic. It is off by roughly 2x before rework and considerably more after it โ and the error is concentrated precisely where 75% of your headcount sits.
Two consequences follow for how you budget.
The first is that per-seat license economics stop working the way the deck says. If the value per seat at the front line is less than half the value per seat at the top, then a uniform rollout priced uniformly is quietly subsidizing your lowest-return seats with your highest-return ones. That can still be the right call โ but it should be a decision, not an accident of procurement.
The second is that the payback window on front-line seats is not a licensing question at all. It is a redesign question, and redesign is staff time you have to schedule. Buying 150 seats and booking 150 seats' worth of savings assumes a workflow change that nobody has been assigned to make.
The Retention Risk Sits in the Same Table
There is a second gradient in the SHRM data, and it runs parallel to the first.
Before AI was adopted, 74% of directors were informed it was coming. Among managers, 55%. Among individual contributors, 33% (SHRM, 2026). Trust tracks the same shape: 80% of directors trust senior leadership on AI, versus 65% of managers and 47% of ICs. And 26% of individual contributors report that their trust in leadership declines as leaders rely more heavily on AI โ against 9% of directors.
Now stack that against expectations. Half of all workers say AI has raised performance expectations on them, and 70% of directors report raised expectations in their own patch. Pace of work is up for 75% of directors and 59% of managers, while 50% of ICs report no change in pace at all.
Read the two gradients together and the front-line position is specific: you were told least, you trust least, you save least, and the expectations on you went up anyway. That is not a communications problem to be solved with a town hall. It is a fairness ledger, and people keep it accurately.
BCG's fourth annual AI at Work survey, covering 11,749 employees across 14 markets, shows what happens when nobody governs the freed time: 66% of AI users report no meaningful guidance on how to reallocate the hours AI frees, and 47% now spend more time managing and directing AI than doing the work themselves (BCG, 2026). Ungoverned time does not become capacity. It becomes supervision โ and supervision is a real attentional cost that shows up on the ledger as "busier," not "freer."
The Counter-Argument, and Why It Only Buys You Time
The strongest objection: this is an early-adoption curve, not a structural feature. Senior people got the tools first and had the latitude to experiment. Give it eighteen months and the front line catches up.
Partly true, and worth pricing. But two things cut against waiting it out. First, the exposure gap is a property of the work, not the calendar โ 34% of front-line work being AI-touchable does not rise because time passes. It rises because someone redesigns the process. Second, the 2% workflow-integration figure tells you that the redesign is not happening organically anywhere, at any tier (ActivTrak, 2026). Waiting converts a gap you could have engineered into a gap you inherit โ while the expectations gradient keeps climbing on the people least served by it.
What to Change This Quarter
Re-baseline the ROI model by tier
Rebuild the business case with three separate lines โ director, manager, IC โ instead of one blended average. Use your own observed numbers if you have them, the SHRM gradient if you do not. If your projected savings drop by half, the model was wrong before, not now.
Pick one front-line workflow and take it to workflow integration
Not a seat expansion. One named process, one owner, one redesign pass, measured on whether the AI step is genuinely inside the workflow rather than beside it. Given that only 2% of workers reach that stage, one process that gets there is worth more than a hundred seats that do not.
Close the 33% communication gap before the next deployment
The cheapest line item on this list is telling individual contributors what is coming and why, before it arrives, at the same rate you tell directors. It costs nothing, and the trust gradient it addresses is the one that eventually shows up in your attrition numbers.
One Decision
Pull up your AI business case and find the tier it was measured on. AI time savings by job level are the single input most likely to be wrong in it. If the hours saved came from you and your direct reports, you are holding the 9-hour number and forecasting it across a population that will deliver four โ while raising expectations on the tier that got told last and trusts you least.
Halve the projection, or fund the redesign that makes it true. Those are the only two honest options, and one of them is free.
