On matched tasks, AI agents cut completion time from 269 minutes to 36 — an estimated 87% less time and 94% less cost than a human working with search alone (Yang et al., arXiv 2606.07489, 2026). That is the number every headline ran. It is also the number that will mislead a Head of Operations into the wrong move this quarter. Read on its own, a 94% cost reduction sounds like a headcount line about to fall. The same study, read past the abstract, says the opposite: the human work does not vanish. It moves. And where it moves — to verification, orchestration, and across job boundaries that never existed on anyone's org chart — is precisely the place a narrow role definition cannot follow.
This is the finding worth your attention, and it is buried under the cost figure. AI agents do not subtract a layer of people. They relocate the residual human work upward and sideways. If your roles are scoped for the old work, the relocation has nowhere to land — and the value the agent created leaks back out as friction.
The Number Everyone Quoted, and the One They Skipped
The study draws on production data from Perplexity's Search and Computer products across a 90-day window from late February to late May 2026, co-authored by a Harvard Business School researcher and Perplexity's own team (MarkTechPost, 2026). The headline contrast is real: an agent session performs roughly 26 minutes of autonomous work, versus 33 seconds for a conventional search. Compressed into a single matched-task comparison, that is the 269-to-36-minute collapse.
Here is the part that did not trend. The study measured what happened to the human alongside the agent, and two signals stand out. First, per-query dissatisfaction fell by roughly 55% on the agent product — users were not just faster, they were getting outputs they trusted enough to build on. Second, and more consequential for how you staff a team, the follow-up work shifted upward. Once the agent handled execution, the human's remaining queries concentrated on verification and extension — checking the agent's output and pushing it further — rather than on doing the task from scratch.
That is a different claim than "AI makes people faster." It says the content of the human job changed shape. The minutes the agent gave back were not returned as idle capacity to be cut. They were reinvested into a higher-order activity the worker was barely doing before: governing the machine's output and extending it into work that sits one level up.
Where the Residual Work Goes: Verification and Scope
Two relocations matter, and operations owns both.
The first is vertical. When an agent executes, the human stops being the executor and becomes the verifier and orchestrator. The job moves up the value stack — from producing the draft to judging whether the draft is right, from running the analysis to deciding which analysis to run and what to do with it. This is skilled work, and it is not the same skill that the role was hired for. A team selected to execute is not automatically a team that can verify and direct.
The second is horizontal. The study found that agent users began attempting tasks that crossed occupational boundaries — work that bundled interdependent subtasks from different roles, required higher-order cognition, and simply did not appear in pre-agent usage at all. The agent did not just speed up the existing job; it expanded the scope of what one person would attempt, pulling in work that used to require a second specialist or a handoff to another function.
Put those together and the operational picture inverts. The agent shrinks the task. It enlarges the role. The person at the desk is now expected to verify machine output and to operate across a wider band of work than their job description ever named. If the role is still scoped to the narrow, pre-agent task, two things break: the verification work goes undone (because no one is accountable for it), and the cross-boundary work stalls at the old silo walls (because the org chart still says it belongs to someone else).
Price the Conflict of Interest Honestly
A rigorous read has to flag the obvious: Perplexity co-authored a study that flatters Perplexity's product, and at least one outlet called that out directly (PPC Land, 2026). The efficiency magnitudes — 87%, 94% — come from a vendor with a commercial stake in those numbers being large, and they deserve the skepticism any vendor-authored benchmark gets. Treat the precise figures as directional, not gospel.
But notice which part of the finding the conflict actually threatens. A vendor has every incentive to inflate the cost-savings number. It has no particular incentive to surface the inconvenient one — that its tool relocates human work into verification and cross-role scope, which is a complication for buyers, not a selling point. The relocation finding cuts against the clean "AI replaces work" narrative that sells agents. That it appears anyway makes it more credible, not less. You can discount the 94% and still take the structural claim seriously: the residual human work moves up and out, regardless of the exact efficiency multiple.
Why Job Architecture, Not Tool Access, Is the Binding Constraint
If the work relocates but your roles don't, the relocation has nowhere to go. This is why the binding constraint on AI value is not how many seats you license or how well your people prompt — it is whether your job architecture can absorb the work the agent pushes upward and sideways.
The broader evidence already says most organizations are stuck on exactly this. PwC's 2026 AI Performance Study of 1,217 executives found that 74% of AI's measured economic value is captured by just 20% of companies — and the distinguishing trait of that leading fifth is not better tools but that they are twice as likely to redesign workflows around AI rather than bolt AI onto existing ones (PwC, 2026). In PwC's framing, the technology delivers roughly 20% of an initiative's value; the other 80% comes from redesigning the work. Deloitte's State of AI in the Enterprise 2026 puts a number on how few have done it: the most common response to AI was employee education, not role or workflow redesign, leaving the large majority of organizations with AI tools layered onto unchanged jobs (Deloitte, 2026).
The pattern is consistent across all three sources. Tool access is no longer the scarce input. Redesigned roles are. The companies winning are the ones rebuilding the job around what the agent changed; the companies waiting are the ones who bought the tool and kept the job the same.
The Mid-Market Exposure
This lands hardest on the 100-to-500-FTE company, and structurally so. Large enterprises carry slack: redundant specialists, an org-design function, the headroom to let work cross boundaries because someone, somewhere, owns the seam. The mid-market runs lean. Roles are tightly scoped because there is no bench, and occupational silos are rigid because every person is load-bearing in exactly one lane.
That is the worst possible starting posture for work that wants to relocate. When an agent compresses a 200-FTE operation's execution work and pushes the residual up into verification and across into adjacent roles, there is no slack role to catch it and no org-design function to redraw the boundary. The verification work falls through the cracks, the cross-boundary tasks die at the silo wall, and the efficiency the agent produced converts into unowned work rather than captured value. The mid-market is the segment most likely to buy the agent on the strength of the 94% figure and least equipped, organizationally, to collect on it.
The Redesign Move This Quarter
The high-leverage move is not another tool evaluation. It is to redesign one role family around the work the agent actually relocates — and to do it deliberately, before the relocation happens by accident and lands nowhere.
Pick one role family where agents are already live. Map what the human now does after the agent runs. You will find two clusters: verifying and correcting agent output, and reaching into work that used to belong to an adjacent role. That cluster is the new job. Write it down as the role, not as overflow.
Make verification an owned responsibility, not a gap. If three in ten agent outputs ship without a human check — the order of magnitude other 2026 workforce data keeps surfacing — the verifier role is your error-risk control, and right now in most teams nobody holds it. Name the owner, give them the time the agent freed, and measure the catch rate.
Hire and move people for judgment and systems-thinking, not task throughput. The relocated job rewards the ability to evaluate machine output and operate across boundaries — capacities a résumé of past task execution barely predicts. This is where objective psychometric signal beats proxy: select for the traits the new role demands, rather than for fluency in the old task the agent just absorbed.
This is the through-line in how we think about talent and operations intelligence at Scovai: when the work changes shape, the decision about who does it should rest on signal that is objective and traceable, not on a job description written for the work the machine just took.
The Decision This Quarter
Here is the one decision to make before the quarter closes. Take your most mature AI-agent deployment and answer a single question: have you redesigned any role around the work the agent relocated, or are your people still carrying job descriptions written for the task the agent now does in 36 minutes? If it is the latter, you are not capturing the agent's value — you are watching it leak out as verification nobody owns and cross-role work nobody is allowed to do. The agent already shrank the task. Your only real lever left is whether you redesign the job before the relocated work falls on the floor.
