The Quiet Lie of Delegation
The question that haunts every conversation about artificial intelligence is almost always the wrong one. We ask whether the technology is smart enough, fast enough, or safe enough, as if the tool itself were the variable that matters. But a new wave of research, drawing on over 53,000 agent skill specifications from the Manus Skills Marketplace, suggests the more uncomfortable question is about us: what are we actually willing to hand over, and what do we tell ourselves about what we have handed over? [1] The gap between what AI claims to do and what it actually does is not a technical bug. It is a social arrangement, and we are the ones who keep signing the contract.
The study, published as an arXiv preprint, introduces the Agentic Adoption Index (AAI), which measures not where AI could work but where workers have actually committed tasks to automated workflows. [1] The finding that should stop us cold is that the occupations where delegation concentrates look almost nothing like the jobs we spent years warning were at risk. The clerks and the data processors, the ones the early doom-laden forecasts said would be replaced first, are not the ones building these workflows. Instead, the adoption peaks among workers with bachelor’s degrees, in the middle of the wage distribution, and then falls off sharply at the top. The most educated, the most compensated, the people with the most to gain from leverage, are the ones holding back.
This is where the deception begins, and it is not the kind of deception we usually worry about. We tend to imagine AI lying to us, hallucinating facts or fabricating sources, and that happens, but it is almost a sideshow. The more pervasive deception is quieter: AI presents itself as a faithful executor of intent, and we want to believe it, because believing it feels like progress. When a worker builds a task into an automated workflow, they are making a statement about that task, that it can be specified in advance, that its steps are knowable, that its outcome is measurable. The act of delegation is an act of self-deception whenever those assumptions are wrong.
Consider what the data actually shows about the shortfall at the top of the wage distribution. The researchers note that technical availability explains most of the variation in adoption, but it does not explain why the most educated occupations lag behind — a gap that persists even after controlling for industry and firm size. Something else is going on, and the paper offers two possibilities: either the work itself resists advance specification, or the professionals who do it are exercising discretion over the pace of codification. Both explanations are really the same confession. The people who understand their work best are the ones most aware that the workflow is a lie, that the task cannot be captured in a prompt, that the judgment involved is not a step in a sequence but a stance toward the world.
The historical parallel is uncomfortable but instructive. Every labor-saving technology has arrived with a promise of transparency, a claim that it simply does what it says, and every one of them has required a period of adjustment during which we learned to see what it was actually doing. The assembly line did not just make cars faster; it made workers into extensions of the line, and we called that efficiency. The spreadsheet did not just make accounting faster; it made certain kinds of questions unaskable, because they did not fit in the grid, and we called that rigor. AI is following the same arc, and the deception is structural, built into the interface that invites us to delegate without asking what we are giving up.
The most telling detail in the research is the gap between what the index tracks and what workers actually use. The AAI measures how closely an occupation’s tasks match the agentic routines practitioners have built and shared, and it tracks what AI could do more closely than what workers currently use it for. That sentence is worth sitting with. The potential outruns the practice, and the reason it outruns the practice is not fear of technology or lack of access. It is the slow dawning recognition, among the people who have tried it, that the tool does not do what it appears to do. It performs a version of the task, a flattened replica, and the difference between the replica and the real thing is exactly the difference that expertise was supposed to protect.
This is why the question of the standard matters so much. When we ask whether AI is good enough, we are really asking who gets to define good enough, and the answer is usually the people who built the benchmark. The researchers behind the AAI are careful to note that their measure is not a judgment of quality, only a record of adoption, but the distinction is hard to maintain in practice. Every workflow that gets shared becomes a template, and every template becomes a standard, and every standard quietly redefines what the task is. The worker who delegates is not just using a tool; they are accepting a definition of their own work that someone else wrote, and that acceptance is the deepest deception of all.
The wage data makes the stakes concrete. The adoption peak at the bachelor’s level suggests that the people delegating are the ones whose work may reflect tasks that resist advance specification, or professional discretion over the pace of codification. The decline at the top is not Luddism; it is the last defense of a kind of work that has not yet been flattened. The professionals holding back are not afraid of being replaced. They are afraid of being misrepresented, of having their work reduced to a routine that can be automated, because they know the routine is not the work. The routine is the shadow the work casts, and they have spent their careers learning to tell the difference.
We should also be honest about the way the technology itself invites this deception. The interface of a modern AI agent is designed to inspire confidence, to present a smooth surface of competence, and the smoothness is itself a kind of lie. A human collaborator who does not understand a task will ask questions, will show hesitation, will reveal the edges of their ignorance. An AI agent will produce output with the same confidence whether it has understood the task or merely recognized its shape, and that uniformity of confidence is the most dangerous feature it has. The worker who delegates cannot see the uncertainty, because the uncertainty has been designed out of the interface.
The paper’s authors call for repeated measurement over time, and that is the right instinct, but we should be clear about what such longitudinal tracking will show. It will show the gap narrowing, as more tasks get codified and more workflows get shared, and we will be tempted to read that narrowing as progress. But the narrowing might just be the spread of the deception, the gradual acceptance of flattened versions of tasks as the real thing. The question we should be asking is not whether AI can do the work, but whether the work we are asking it to do is the work that matters. The goal we pursue is not the goal we started with, and the technology is not the only thing that has changed.
This is the point where the research stops being a study of adoption and becomes a study of us. The 53,000 agent configurations are not just data points; they are decisions, made by people who looked at their work and decided that a part of it could be handed over. [1] Each configuration represents a specific task decomposition, a moment where a worker judged their own process reducible to code Each decision was a judgment about what the work really was, and each judgment was a small act of self-definition. The people at the top of the wage distribution are not refusing to make that judgment; they are making a different one. They are saying that the part of their work that can be specified is not the part that matters, and they are willing to leave the efficiency on the table because they know what the efficiency would cost.
The deception, in the end, is not that AI fools us about what it can do. The deception is that we fool ourselves about what we are doing when we use it. We tell ourselves we are saving time, but we are also surrendering definition. We tell ourselves we are leveraging our expertise, but we are also outsourcing the judgment about what expertise is. The tools do not lie to us; we lie to ourselves, and the tools make it easy. The question is not whether AI is good enough. The question is whether we are honest enough to see what we are becoming, one delegated task at a time — and whether the version of ourselves we are building is one we can live with.