Part 4: The People Next to the Agents

What changes for people who work next to AI agents: Which tasks move, the new roles of reviewer, operator and agent manager, what managers do differently, and what early evidence says about skills and junior jobs. Built on Stanford payroll data, a Harvard hiring study, Microsoft's 20,000-worker survey and a randomised trial on learning.

Tue, 27 Oct 2026 · 16 min

Made for you, by AI, from the sources below.

Updated in October 2026

Listen to it in AI Agents at Work

Transcript

The script the two voices were given, with the sound cues taken out. A voice may say a word a little differently, so the audio is the episode.

Mia: Welcome to Walking Points. Here is a puzzle from an AI lab. Its own engineers said they use their coding agent in about 60 percent of their work, and yet more than half of them said they could fully hand off only up to a fifth of it.

Adam: That sounds like a contradiction. If it's in most of the work, why can't they hand more of it over?

Mia: Because they still check it. Using an agent and trusting an agent are two different things, and the space between them is what this episode is about.

Adam: Okay, so what is the episode, exactly?

Mia: It's about the people next to the agents: tasks, new roles, managers, skills and junior jobs. Earlier episodes covered what agents do. This one covers what they do to the people around them, and it earns a whole episode because the evidence is young and a lot of loud claims rest on very little of it. We'll take three ideas in order. First, which tasks move and which stay. Then the new jobs forming around agents, and what managers do differently. And last, skills and junior jobs, which is where the data is thinnest and the arguments loudest.

Adam: Is there a way to keep the loud claims honest?

Mia: One rule for the whole episode: for every number, ask who was measured and for how long.

Adam: Fair. And where did we leave things?

Mia: Last Tuesday we covered when agents go wrong: the incidents, the causes, and why human oversight so often fails in practice.

Adam: So, tasks. What actually moves when an agent shows up?

Mia: The cleanest picture comes from a company studying its own staff. In August 2025 Anthropic surveyed 132 of its engineers and researchers, interviewed 53 of them, and looked at usage data from its coding agent. They estimated a large productivity gain over a year earlier. Debugging was the most common use, at 55 percent, then making sense of unfamiliar code, proposing refactors and building small internal tools.

Adam: That's the headline. What's underneath it?

Mia: Two details matter more. About 27 percent of the agent-assisted work was work that would not have been done at all otherwise, like dashboards and side projects that were never worth a person's afternoon. So part of what moves is brand new work. The second detail is the one you opened with: most engineers could fully hand off only between none and a fifth of their work, because they felt the need to check the output.

Adam: So what moves is the drafting, and what stays is the checking.

Mia: Exactly that. With a caveat: it's a self-reported survey of people at an AI lab, who are not typical workers.

Adam: Right, so is there anything from ordinary workplaces?

Mia: Microsoft's Work Trend Index this year. It surveyed 20,000 AI users in ten countries between 18 February and 20 April, and looked at more than 100,000 anonymised Copilot conversations. Only 17 percent of those conversations were about producing an output. Nearly half, 49 percent, were about analysis, problem solving and evaluation.

Adam: Hang on. Microsoft sells Copilot. Why believe a survey from the company that wants you to use it?

Mia: Fair challenge, and the level of any figure there deserves care, since it's self-reported and the seller ran it. But I'd point at the direction rather than the size. When asked which skills are gaining importance, half named quality control of AI output and 46 percent named critical thinking. And 86 percent said they treat AI output as a starting point, not a finished product. That matches the lab's engineers, from a very different crowd.

Adam: So the work shifts from making the first version toward judging it. Does that create new jobs, or just more checking for the same people?

Mia: Both, and the names are still unsettled. Anyone who tells you there is a stable job family yet is ahead of the evidence. But three functions keep showing up in company writing and analyst reports. The reviewer checks what the agent produced before it counts: the engineer who reads the code, the accountant who approves the reconciliation, the support lead who samples conversations.

Adam: That one sounds like what people already do.

Mia: Yes, most people become a reviewer quietly the moment an agent enters their team. And it's the role the last episode warned about, because reviewing a mostly correct machine is tiring, and attention fades.

Adam: It's like proofreading a text that's right ninety-nine times out of a hundred. By the time the error arrives, your eyes have already stopped reading.

Mia: That's a good way to put it, and it pushes the analogy somewhere useful: the fix for proofreading isn't more willpower, it's a structure. Time set aside, a checklist, someone else sampling. We'll come back to who sets that. The second function is the operator, who keeps the agents running: which agents are deployed, what each costs to run, where they stall, and when one should be retrained or retired. It looks more like running a service than managing a person.

Adam: And the third?

Mia: The agent manager, sometimes called an orchestrator. That person decides how work is divided between people and agents, keeps agents inside the rules, and coordinates them across systems. Alongside sits a governance function: defining what an agent may touch, auditing what it did, and keeping the evidence trail a regulator would ask for. The next part comes back to that.

Adam: Is that just a lot of job titles for a slide deck?

Mia: It might be, and I'd say so plainly. These descriptions come mostly from vendors, consultancies and analysts. Job postings and surveys will tell us in a year or two which titles are real. What does look solid is the question Microsoft says every organisation must answer: who reviews agent performance, who updates agent workflows, and how a local win spreads to other teams. If nobody can answer those, the agent has no owner, and an unowned agent is how the incidents from earlier in this series begin.

Adam: So for someone who leads a small team, the practical point is that an agent without a named person is a risk, even if it works well.

Mia: Right, and that leads to managers. The Microsoft data is most specific here, with the usual caution that it's correlation from a survey. It found a group it calls frontier professionals, 16 percent of AI users, who work in a recognisably different way. Among them, 53 percent said they pause before a task to decide whether a person or an AI should do it, against 33 percent of other users. And 43 percent said they deliberately work without AI sometimes to keep their skills sharp, against 30 percent.

Adam: Wait, so the people who work in that different way are also the ones who sometimes switch AI off on purpose?

Mia: Some of the time, yes. They treat it like a gym membership for judgment. The manager finding is sharper still. When managers visibly use AI themselves, workers reported a 17 point lift in the value they get from it, a 22 point lift in critical thinking about it, and a 30 point lift in trust in agent systems.

Adam: Mia, those are big jumps. Isn't that just good managers doing everything well, with AI as a coincidence?

Mia: That's the right objection, and the data can't rule it out. It's a survey, so it shows things travel together, not that one causes the other. What I'd add is that the rest of the picture leans the same way. Psychological safety to experiment tracked with about 1.4 times the likelihood of frequent agent use. And Microsoft's own summary is that culture, manager support and how people are developed account for roughly two thirds of AI's real impact, with individual mindset the remaining third.

Adam: So what does a manager actually do differently, in plain terms?

Mia: Three things. They decide the split of work between people and agents, rather than just assigning tasks. They set the standard for checking, including whose job it is and how much time it gets. And they model the behaviour, because people copy what their boss does with the tool, not what the policy says. If you lead a team, that means the agent is no longer a tool you hand out. It's part of how you divide the work, and your own use of it is the training your people actually receive.

Adam: And is that happening?

Mia: Not much yet. Only 26 percent said leadership was clearly aligned on an AI strategy, and only 13 percent said they would be rewarded for redesigning their work if the results fell short.

Adam: So people are asked to experiment with no cover if it goes wrong.

Mia: That's the reading. Adam, here's the part everyone argues about: junior jobs. I want to be exact about what each study can and cannot show. The best known is "Canaries in the Coal Mine?" from Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab. It uses payroll records from ADP, the largest US payroll provider. The latest update, from August with data through June, found no widespread, economy-wide job loss. But workers aged 22 to 25 in the most AI-exposed jobs, such as software development and customer service, now have employment 19 percent below where it would be had it tracked their peers in less exposed work.

Adam: Nineteen percent. Is that people losing jobs?

Mia: Mostly not. The mechanism is reduced hiring, not layoffs. And the gap was 15 percent in the July 2025 data, so it has widened. In absolute terms, employment for this young group in exposed jobs fell about 11 percent between November 2022 and this June, while the comparison group grew about 10 percent. Experienced workers in the same occupations show no comparable gap.

Adam: Why would experience protect you?

Mia: The update offers a mechanism. Young workers lost ground in jobs that rely on codified knowledge, the kind written in manuals and textbooks, which AI handles well. Experienced workers in jobs built on tacit knowledge, the kind learned by doing, kept growing. The earlier version also found the drop was in uses of AI that automate work, not in uses that augment it.

Adam: Does it prove AI is the cause?

Mia: The authors themselves say no. They call it not definitive evidence, estimates are sensitive to specification choices, the gaps shrink when education is accounted for, and some differential trends predate generative AI. The pattern does persist when they control for interest rates and remote work.

Adam: So one study, and a cautious one. Is there a second?

Mia: Yes, from different data. Seyed Mahdi Hosseini Maasoum and Guy Lichtinger of Harvard used résumé and job-posting data from Revelio Labs on about 62 million workers at roughly 285,000 US firms. They identified firms adopting generative AI by their postings for roles dedicated to implementing it. At those firms, junior employment fell about 7.7 percent within six quarters from early 2023, while senior employment kept rising. Again slower hiring, not firing: in wholesale and retail, about 40 percent fewer junior hires. They call it seniority-biased technological change.

Adam: And the catch?

Mia: It's observational, and spotting adopters by job postings is an imperfect proxy. So two independent datasets agree on the direction and disagree on the size. Neither proves AI is the cause, and neither shows the jobs are gone for good. What they show is that the entry door is narrowing in exposed fields, while the people already inside are fine.

Adam: Which raises an awkward question. If the first rung gets thinner, how do people learn the job?

Mia: That's the apprenticeship problem, and there's a randomised trial on it. Anthropic recruited 52 mostly junior software engineers, each with more than a year of Python experience, and asked them to complete two coding tasks with a Python library for asynchronous programming that was new to them, some with an AI assistant and some by hand. A quiz followed. The AI group averaged 50 percent and the hand-coding group 67 percent, a gap of nearly two letter grades.

Adam: Presumably the AI group at least finished faster?

Mia: About two minutes faster, but that difference was not statistically significant. And how people used the assistant mattered more than whether they did. The ones who scored well asked for explanations, asked conceptual questions, or generated code and then asked follow-up questions to understand it. The ones who scored badly handed the whole task over, or started independently and gradually delegated everything, or used the AI to debug without trying to understand.

Adam: So it's the difference between asking it to do the work and asking it why.

Mia: Yes. With caveats: small groups per pattern, the best-scoring patterns had two to seven people each, and the quiz measured understanding right after the task, not long-term retention.

Adam: Is there a bigger theory behind that?

Mia: An economist's framing. Enrique Ide's paper on automation and the intergenerational transmission of knowledge argues that when AI takes over knowledge-heavy tasks, it can break the apprenticeship link by which experienced people teach younger colleagues, trading immediate productivity for expertise built up over decades. The Anthropic engineers raised the same worry about mentorship. But that's theory plus early signals, not an observed outcome, and it's worth saying so.

Adam: Let me try to put the layers together. Agents take the drafting and leave the judging, but judging is learned by drafting for years.

Mia: That's the tension. Reviewing well requires having done the work. It's why the finding about people deliberately working without AI sometimes is more interesting than it first sounds.

Adam: So what does a curious professional do with all this, without anyone telling them what to buy or quit?

Mia: Three readings, each for a role. If you review agent work, the evidence says the review needs to be a real task with real time, not a rubber stamp, because attention fades on a mostly correct machine. If you manage people, the data points to deciding out loud who checks what and using the tools visibly yourself, since people copy what you do. If you're early in a career, the trial suggests that asking the AI why, not only for output, is what preserved learning. And hold the headline numbers loosely: the gap for young workers is real in the data we have, the cause is still argued, and the evidence is only a few years old.

Adam: Is there one thing to actually try?

Mia: One checkable thing for this week. Pick one piece of agent output that you approve, time how long the review really takes, and write down the name of the person who owns it. If the time is close to zero or nobody owns it, then that approval is a rubber stamp.

Adam: Simple enough to test by Friday.

Mia: That was part four of five. Part five is called Rules, responsibility and your first agent, and it is about the governance of what agents may touch.

Heard something wrong? Tell us