The AI Agent Doesn't Work for You. It Works Around You.

When people talk about AI agents in the workplace, they tend to imagine a kind of digital assistant that handles the tedious background noise of a job — the scheduling, the data entry, the forms that exist only because some system somewhere demands them. The dream is that automation absorbs the friction and leaves the human free to do the work that actually requires them. It is a reasonable dream. It is also, according to a substantial new body of research, mostly not what is happening.

A large audit drawing on responses from roughly 1,500 workers across 104 occupations^[2] mapped where AI agents are being deployed most aggressively against what those same workers identified as the tasks they least wanted automated. The overlap was not incidental. In occupation after occupation, the tasks absorbing the heaviest AI intervention were the ones workers consistently flagged as central to their professional identity, their judgment, and the part of the job that made them feel like they were doing something rather than just executing a process. The agents were not filling gaps. They were displacing cores.

This is not primarily a story about job loss, though that question is never far from the surface. It is a story about what automation actually optimizes for, and whose preferences get encoded into the deployment decisions that shape millions of working days. The common framing — AI as a tool that augments workers — assumes a basic alignment between the people building these systems, the companies deploying them, and the workers living inside them. The audit suggests that alignment is largely fictional.

The mismatch has a logic to it once you understand whose problem the agent is solving. Companies are under pressure to reduce labor costs, accelerate throughput, and demonstrate that they are capitalizing on new technology before competitors do. Workers are under pressure to do good work, maintain professional standing, and not become unnecessary. These are not the same pressure, and they do not produce the same automation priorities. What gets built reflects who is driving the decision.

The Automation That Nobody Asked For

The research methodology matters here. The audit did not simply ask whether workers feared AI — a question almost guaranteed to produce either defensive minimizing or theatrical anxiety depending on the respondent. It asked something more granular: which specific tasks in your current role would you find it most acceptable for an AI system to handle, and which would you find least acceptable. Then it mapped those preferences against actual deployment data across the same occupational categories. The gap between preferred automation and actual automation was consistent enough to constitute a pattern rather than a coincidence.

Tasks workers were generally comfortable seeing automated tended to cluster around administrative overhead: documentation, routine data retrieval, scheduling logistics, boilerplate correspondence. These are real parts of the job, but workers understood them as instrumental — necessary for the work rather than constitutive of it. The tasks they resisted automated intervention in were the ones involving contextual judgment, relationship management, quality evaluation, and anything requiring a call that could go wrong in a way that mattered. A nurse comfortable with AI handling discharge paperwork was not comfortable with AI triaging patient concerns. A software developer fine with AI generating test scaffolding was not fine with AI making architectural decisions. The distinction was not between hard tasks and easy ones. It was between tasks that require accountability and tasks that do not.

“The agents were not filling gaps. They were displacing cores.”

What made the audit findings particularly sharp is that this preference structure was not random or idiosyncratic — it tracked closely with something researchers in occupational psychology have been documenting for decades under the heading of task identity^[4]. People derive meaning from work in proportion to how clearly they can trace the connection between their effort and a visible outcome. Tasks involving judgment, decision, or direct human consequence score high on task identity. Tasks involving data processing and administrative compliance score low. Workers, without any coaching from a researcher, were reliably trying to protect the high-identity parts of their roles. Companies, following a different set of incentives entirely, were reliably automating them.

What Gets Lost When Judgment Gets Outsourced

There is a familiar argument that workers are simply being sentimental — that resistance to automation is a normal human response to change, not a reliable signal about which automation is good or bad. The argument has some merit in narrow cases. But it does not hold up well against the specific tasks workers were flagging, and it ignores something important about what happens to skill when the occasions for using it disappear.

Cognitive skill does not persist independently of practice. A radiologist who reads^[1] thousands of scans develops a pattern-recognition capacity that a radiologist who rarely reads scans does not maintain. A financial analyst who builds models from first principles understands the assumptions in ways that an analyst who reviews AI-generated models may not. This is not about ego or professional gatekeeping. It is about what cognitive researchers call the generation effect: the understanding you build by producing something is qualitatively different from the understanding you build by evaluating something someone — or something — else produced. When AI agents take over the generative portion of a task and leave the human to review the output, that is not augmentation in any straightforward sense. It is a restructuring of cognitive labor that tends, over time, to shift expertise from the worker toward the system.

This matters for reasons beyond individual job satisfaction. When skill erodes at the worker level, organizational resilience erodes with it^[3]. If the AI agent fails, is wrong, or encounters a situation it was not designed for, the humans who are supposed to catch the error need to have retained enough expertise to recognize the problem. The audit found that many of the occupations experiencing the most aggressive AI agent deployment were also the ones where the gap between AI output and professional-grade human review was narrowest — meaning the workers best positioned to catch errors were the ones with the least practice at the underlying task. That is a fragility that does not show up on a quarterly productivity report until it becomes a liability event.

The Error Lands on the Person Who Didn't Make It

One of the underexamined features of AI agent deployment is what happens when the system gets something wrong. The accountability structure in most workplaces has not been redesigned alongside the technology. Workers are still formally responsible for outcomes that AI agents are now generating. In practice, this means the human in the loop — often someone who has been trained to review AI output rather than to independently produce the underlying work — carries the liability for errors they may not have the expertise, time, or access to detect.

“Companies get the efficiency. Workers get the accountability. The agent gets neither.”

Across several occupational categories in the audit, workers described a specific and recurring dynamic: the AI agent produces a draft, a decision, or a recommendation; the worker is given a narrow window to approve or flag it; approval becomes the default because flagging requires time and justification that the workflow does not budget for; and when something goes wrong downstream, the worker's approval is the last human signature on the chain. The agent introduced the error. The worker endorsed it under pressure. The worker absorbs the consequence. This is not a design flaw in some narrow technical sense — it is what happens when accountability structures built around individual human judgment are layered over systems that distribute decision-making in ways those structures were never designed to handle.

In knowledge work, this plays out through a subtle but persistent erosion of professional credibility. A lawyer whose filings are increasingly AI-drafted, a therapist whose session notes are AI-summarized, a teacher whose feedback is AI-generated — each faces a version of the same asymmetry. The efficiency accrues to the employer. The professional reputation rides on outputs the professional did not fully control. Over time, this reshapes what it means to be a professional in that field, and not in a direction most professionals would have chosen.

Platform Dynamics, Applied to Your Boss

The deployment mismatch is easier to understand if you stop thinking about AI agents as workplace tools and start thinking about them as platforms. Platforms do not optimize for user satisfaction in any deep sense. They optimize for engagement, retention, throughput, and the metrics that make them legible to whoever is paying. In the consumer context, this is well-documented: the platform and the user frequently have divergent interests, and the platform usually wins because it controls the architecture of the experience. AI agents in workplaces operate under a structurally similar dynamic, except that the "user" is the worker and the "platform" is shaped by employer priorities, vendor incentives, and procurement decisions made by people who do not do the job in question.

This helps explain why worker preference is such a poor predictor of deployment decisions. Workers have strong, contextually informed views about where automation helps and where it degrades the quality of their output and their work life. Those views rarely reach the people designing or purchasing the systems. The audit found that in most occupations, there was no formal mechanism for worker input into automation decisions — no structured feedback loop, no pilot period with meaningful opt-out rights, no post-deployment review that weighted worker assessments alongside productivity metrics. The preferences of the people most directly affected by the technology were simply not load-bearing in the decision architecture. This is not unusual. It is the norm.

Augmentation Is a Design Choice, Not a Default

“The promise of augmentation does not survive contact with a deployment process that never asks the worker what augmentation would actually mean for them.”

None of this means AI agents cannot improve working conditions or make demanding jobs more manageable. The audit found real cases where automation had absorbed genuinely depleting work and returned something more like professional capacity to workers who had been buried in administrative load. A social worker who spent forty percent of her time on compliance documentation and now spends twenty is not experiencing augmentation as a marketing slogan — she is experiencing it as two days of her week she can spend with clients. These cases exist. They are also not the dominant pattern the research found.

What distinguishes the cases that worked from the cases that didn't is rarely the sophistication of the AI. It is whether anyone asked the worker what was actually making the job harder, and whether the deployment was designed to address that rather than to hit a different target entirely. The technology is genuinely flexible enough to be deployed in either direction. The choice of direction is not technical. It is a decision about whose experience the system is being built to optimize, and right now, that question is mostly being answered by people who do not have to live with the answer.

The audit's most uncomfortable finding is not that AI agents are being deployed badly in isolated cases. It is that the systematic mismatch between worker preference and actual deployment suggests something closer to a structural feature of how these decisions get made — fast, from the top, with productivity metrics as the primary signal and worker experience as a rounding error. That is not a technology problem. It is a power problem wearing a technology disguise. And no upgrade to the model is going to fix it.

References

1. Does using artificial intelligence assistance accelerate skill decay and hinder skill development without performers' awareness? (pmc.ncbi.nlm.nih.gov)
   Provides evidence that radiologists who rarely read scans lose pattern-recognition capacity, supporting the article's claim about skill erosion from reduced practice.
2. Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce (arxiv.org)
   Provides the audit framework and data from 1,500 workers across 104 occupations comparing what workers want automated versus actual AI deployment.
3. The Vicious Circles of Skill Erosion: A Case Study of Cognitive Automation (doi.org)
   Establishes the link between skill erosion from automation and organizational resilience decline when workers lose practice on core tasks.
4. Job characteristic theory (en.wikipedia.org)
   Defines task identity concept showing how workers derive meaning from tasks involving judgment and direct human consequence.