The Missed Reps

There is a growing argument in the business press, and it sounds reasonable on first hearing. Large language models are now articulate, patient, consistent, and knowledgeable in ways that human agents often are not. Customers in certain contexts are starting to prefer them. If that trend continues, AI will set a standard for service that raises the bar for actual humans. The mediocre agent, the distracted analyst, the tired associate on their fourth call of the afternoon, may find themselves measured against a machine that does not get tired, does not go off-script, and does not have a bad morning.

The argument is sincere, and the observations behind it are accurate. But it measures the wrong thing. It measures the quality of a single interaction at a single moment. It does not ask what those interactions were producing before the AI took them over. The clearest pattern I see is AI removing the reps that made humans good at the job in the first place, rather than raising the bar for them. The missed reps are the cost nobody is measuring, and they are larger than the efficiency gains that triggered them.

The bar argument#

Jonathan Peachey put the question most sharply in an October 2025 essay. As LLMs get better at seeming human, will they raise the bar for actual humans? Will we start judging one another against the calm, articulate, knowledgeable and perfectly patient standard set by machines? The observation is consistent with the research. A September 2025 meta-analysis in the Journal of Marketing, drawing on 327 experimental studies involving almost 282,000 participants, found that in many service contexts artificial agents were received almost as positively as human ones, and in some contexts more positively. The human advantage, when present, concentrated in emotionally charged or highly ambiguous interactions.

So the provocation has teeth. Against the AI's calm, patient, knowledgeable baseline, the variable human is at risk of looking worse, not because humans have got worse, but because the bar has moved to include things, like infinite patience and perfect consistency, that humans were never trying to offer. But the argument assumes the only thing a human interaction produces is the interaction itself. And that is not how humans are made.

The missed reps#

Every senior became senior by doing work that, in hindsight, mostly did not need to be done by them. The junior lawyer who took the notes. The first-year associate who drafted the memo that was rewritten three times. The graduate analyst whose first model contained an error the vice-president caught and explained. The contact-centre agent who handled a furious caller at 4pm on a Friday and figured out what to do when the script ran out.

None of those reps were efficient. That was not the point. The point was that the human doing the work was absorbing something through it: context, tone, the shape of the client's worry underneath the question they had asked, the feel of when an answer was complete. Call these the missed reps. They are what AI collects when a firm re-routes the interactions that used to build its people. The tier-one call handled by a bot is also the call that would have taught a new agent what a customer sounds like when they are about to escalate. The reps are not being outsourced. They are being retired.

The missed reps do not show up in this quarter's results. They show up in the bench strength of the firm five and ten years out, in the promotion panel that realises a candidate who looks excellent on output has never sat with a client through a difficult meeting, in the leadership pipeline that suddenly has a shape nobody noticed it was acquiring.

What the evidence suggests#

Three strands deserve to be read together. The Klarna case is the most publicly documented. Between 2022 and 2024 the fintech eliminated around 700 customer service positions, replacing them with an AI assistant; by early 2024 it reported the assistant was doing the work of 700 full-time agents. By mid-2025 Klarna was rehiring, with the CEO acknowledging the company had overestimated AI's capabilities and underappreciated the human aspects of service, as satisfaction dropped on complex cases. The reps cannot be recovered retrospectively.

The second strand is Steve Hasker of Thomson Reuters, writing in November 2025, describing how entry-level legal work has been progressively automated and how generative AI has accelerated it. As the White & Case partner Nandan Nelivigi put it, much of the process now happens on a personal screen rather than in a conference room where juniors could observe how others work, so a different approach to conveying basic skills is needed. Thomson Reuters' survey of nearly 2,300 knowledge workers found 81 percent had used AI to start or edit work, while only 22 percent of employers had a clear AI strategy.

The third strand is the 2025 study by Cui and Demirer of nearly 5,000 developers, which found AI gave junior developers productivity gains two to three times larger than seniors. This is widely read as evidence that AI closes the junior-senior gap. But the study measured productivity, not the formation of judgement. My reading is that the gap closes partly because AI is doing the work that used to distinguish the two. A junior who produces senior-level output because the AI is pulling them up has not yet built the judgement that made the senior worth pulling up from. When that junior becomes the senior in charge of the AI, the absence of that underlying capability will matter. Read together, the strands point the same way: where AI is deployed most aggressively in exactly the tasks juniors grew through, short-term output improves and long-term capability erodes.

The deliberate response#

The answer is not to be sentimental about manual processes or to hold back from AI adoption. It is to treat the apprenticeship question as a first-order strategic decision. Every AI deployment that removes a category of interaction from humans is also removing a learning opportunity, and leaders should assume by default that they must replace that learning value deliberately or accept that their senior bench will hollow out.

The firms doing this well are counting the deliberate practice a junior gets in a year, not just AI usage. They are redesigning what juniors do rather than simply giving them less: the best response is not to have graduates review AI output, which teaches surface-error recognition but not judgement, but to give them a smaller number of harder interactions with more senior coaching around each one. And they are protecting the reps the AI cannot replace, the hard client meeting, the genuine complaint, the negotiation that does not follow a script, even where a cheaper AI solution exists. None of this is complicated. It is uncomfortable, because it requires leaders to hold a short-term cost to protect a long-term capability whose absence will not be visible for years.

What we are actually measuring#

Will AI raise the bar for humans? In a single interaction, measured on a single dimension at a single moment, yes. But the service interaction is not the only product of a service interaction. The other product is the person doing it, the accumulated result of every hard call they handled, every first draft they got wrong, every awkward meeting they survived. The quality of an industry ten years from now will be set by what we let that person do between now and then, and we are in the middle of deciding to let them do less of it, on the grounds that the AI does it more smoothly today.

Peachey asked whether AI raises the bar for humans. The question underneath it is whether we still know how to make humans, in the specific, patient, accumulated sense of how a firm turns a graduate into a partner, an agent into a manager, an analyst into someone whose judgement you would trust in a crisis. The firms that answer that deliberately will build seniors more capable than their predecessors. The firms that do not will not notice the cost for several years, and when they do, it will be the absence of people they had assumed would be there, with no quick way to produce them.

The argument is older than AI#

The clearest statement of this mechanism predates the personal computer. In Ironies of Automation (Automatica, 1983), Lisanne Bainbridge set out the paradox that automating a system removes the routine operation through which operators stay practised, while leaving them accountable for the difficult cases the automation cannot handle. In her words, "by taking away the easy parts of his task, automation can make the difficult parts of the human operator's task more difficult."

She was writing about process plants. The argument transfers to knowledge work without modification, and its age is the point. This is a well-documented property of automation rather than a reaction to generative AI that the industries with the highest cost of failure, aviation and process control, learned to design against decades ago, and that knowledge work is currently rediscovering the expensive way.

Independent evidence from the field#

Matt Beane's The Skill Code (2024) provides the closest thing to direct observation. Studying robotic surgery, he found residents losing the hands-on repetitions through which surgeons have always been made: the senior operates the console alone, the junior watches, the patient does well, and the training stops happening.

Beane's term for what juniors do in response is shadow learning, the informal and sometimes rule-breaking ways they scavenge practice the official system no longer provides. It is an important qualification to the argument on this page. The reps do not simply disappear. Some are recovered unofficially, by the people confident or well-connected enough to find them, which makes development less systematic and considerably less fair.

Further reading#

The graded evidence, including what each study does not support, is in the evidence base. The wider field map, including Bainbridge and Beane in full, is in the essential works.

Related SuperSkills research#

The structural version is the missing rungs; the individual result is synthetic seniority; the organisational accumulation is capability debt. On the learning evidence, see how humans learn with AI.

Development of the idea#

Stated directly in the Box of Amazing essay The Case for Being Bad at Things (18 January 2026), and connected to the pipeline argument in The Missing Rungs (21 September 2025) and Solving Synthetic Seniority (5 July 2026). See the timeline for the full record.

The learning science underneath this is desirable difficulty: the conditions that make work feel harder are frequently the ones that build durable capability. The older term for what happens when they are removed is deskilling.

What happens when the reps were never taken and the person is asked to supervise anyway: who supervises work they cannot do themselves?

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.

How this research works · Reviewed quarterly · Found an error? Tell me and it is corrected on the page.