The pitch hasn't changed in three years. Agents absorb the routine, people ascend to creative and strategic work. It's an appealing story, and I understand why everybody repeats it: it promises productivity without a headcount fight, and it sounds like a promotion for the people whose jobs are being rearranged.
I've been watching what actually happens when agents land in a workflow, and the reclaimed hours are not going to strategy. They're going into specifying what the agent should do precisely enough that it can act without wrecking anything, into keeping the conditions stable enough for it to keep operating, and into deciding, output by output, what gets accepted into the real workflow and what gets thrown out. That work is detailed and consequential, and it doesn't resemble strategy in any way an org chart would recognize.
Autonomous freight shows the shape of it plainly. Einride runs trucks that drive themselves on permitted routes. The human job isn't driving. It's defining the operating envelope — which routes, which weather, which vehicle states qualify — and then watching whether reality stays inside it. When conditions drift, a person steps in. An investigation by Senator Markey's office found that all seven autonomous-vehicle companies it surveyed used remote operators for exactly this kind of guidance. None of them disclosed how often.
Then there's the Danish hospital where researchers documented what happened after automatic speech recognition took over clinical note-taking. The medical secretaries who used to check dictation were removed from the loop. The checking wasn't removed. It moved to the clinicians, who now catch errors, fix formatting and verify accuracy on top of the job they were already doing. The researchers call this articulation work: the ongoing, mostly invisible labor of making an automated system function in practice. The clinicians reported fixing the same errors again and again, with no sign the system learned anything from being corrected.
Articulation work is the closest label I've found for what I'm describing, and it comes out of sociology and human-computer interaction, not out of anyone's job architecture. It hasn't crossed into how organizations design roles, set pay, or build career ladders. The work exists, but no corresponding job does.
And the judgment it takes doesn't map cleanly onto anything we already know how to hire for. Specifying objectives means anticipating situations you've never encountered, on behalf of an actor that won't stop and ask you when a case is ambiguous. Accepting or rejecting output means holding a quality bar the system has no idea it's supposed to meet. Monitoring may be the strangest part: you're watching a system whose reasoning you can't inspect, so when the outputs start drifting you're diagnosing a process you can't interrogate. That's a new kind of debugging, and most people doing it are inventing their technique as they go.
Sevda Polat argued in this publication that automation removes the predictable middle of a workload rather than shrinking it evenly, leaving a denser and more expensive exception layer behind. That's part of it. But a lot of what I'm pointing at isn't exceptional at all. It's the unglamorous maintenance of the normal path: updating specifications when conditions shift, noticing drift before it becomes a problem, calling whether a result is good enough to ship.
I've spent twenty years watching this industry rename operations — DevOps, SRE, platform engineering, anything to avoid saying "ops," as if keeping production alive is somehow beneath us. Something similar is forming around agent adoption. The human work that makes autonomy possible is being created much faster than anyone is naming it, valuing it, or building a career path around it. And because the story everyone's been told says these hours belong to strategy, the people doing the work often don't recognize it as a skill they're developing. They think they're just not being strategic enough.
That last part is what worries me. The pattern is early and I'd rather trace it than pronounce on it. But it's shown up in every setting where I've watched agents go into production, and until organizations can see this work well enough to staff and pay for it, they're going to keep resourcing a future that isn't showing up.
- Workers want more agency: Stanford's WORKBank study found that across 104 occupations, workers preferred more human control than AI experts considered technically necessary on nearly half of surveyed tasks.
- Oversight happens in episodes: A qualitative study of professionals at two German firms found that oversight of AI-generated work occurs during drafting, refining, and review rather than at one final checkpoint.
- More output, flat returns: McKinsey's 2026 global survey found enterprise AI scaling rose to 44%, but the share of organizations attributing positive EBIT impact to AI held essentially flat from 2025.
- Human-AI teams underperform expectations: A peer-reviewed meta-analysis of 106 experiments found that human-AI combinations performed better than humans alone on average but worse than the better of the human or AI working solo.