AI Burnout Isn't About How Much You Work. It's About Who Decides.
A software team adopted an AI code review tool in early 2025. By Q3, ticket velocity was up 40 percent. By Q4, three senior engineers had quietly started interviewing elsewhere. One of them, in an exit conversation, said something that didn't fit the standard burnout script: "I'm not overworked. I just don't feel like my judgment matters anymore."
That is not the burnout the industry is measuring. And it's the one that's spreading.
The Variable That Predicts AI Burnout Better Than Workload
Burnout research has been workload-centric for decades. The Maslach Burnout Inventory — the field's standard diagnostic tool — treats exhaustion as the primary metric. Reduce hours, reduce exhaustion. The intervention logic follows from this: if workers are burning out, they're doing too much.
But a 2026 ScienceDirect paper, "Safeguarding worker psychosocial well-being in the age of AI: The critical role of decision control," found something that disrupts this model. Across AI-augmented work contexts, decision control — the degree to which workers can shape, override, or meaningfully understand the decisions they're responsible for — predicted burnout outcomes better than workload volume.
Workers in high-workload, high-decision-control roles showed lower burnout indicators than workers in lower-workload, low-decision-control roles. The relationship wasn't subtle. When workers couldn't explain why the AI recommended what it recommended, couldn't modify the output without going around the system, or couldn't override decisions they believed were wrong, they burned out — regardless of how many hours they were working.
This matters because most AI augmentation strategies are designed around workload reduction. That's the promise: AI handles the routine, humans handle the judgment calls. But the promise relies on the judgment calls actually being preserved. When they're not — when the AI handles the judgment calls and humans execute against them — you've removed the psychological buffer the research identifies as load-bearing.
Why Self-Determination Theory Predicts This
Self-Determination Theory, developed by Deci and Ryan at the University of Rochester, identifies three psychological needs as fundamental to sustained motivation and wellbeing: competence, relatedness, and autonomy. All three matter. Autonomy — the sense that your actions reflect your own judgment rather than external control — is not optional for wellbeing. When autonomy is persistently frustrated, motivation erodes and exhaustion accumulates even when the work itself is not demanding.
AI-augmented workflows frequently frustrate autonomy in a specific way. The work isn't being taken away. The worker is still responsible for outcomes. But the decisions driving those outcomes are being made by a system the worker doesn't fully understand and often can't override without friction. Responsibility without authority is the psychological burden profile that produces the fastest burnout.
The distinction between executing a decision and making one turns out to matter enormously to wellbeing, even when the behavioral output looks identical from the outside.
The Three Patterns Where Decision Control Loss Is Hardest to See
There are three common AI workflow patterns where decision control loss accumulates without being recognized as the cause of the burnout it produces.
The recommendation executor. The AI surfaces a recommendation. The worker's role is to review and approve. In practice, the review is cursory because the approval rate is high and the penalty for rejection is friction. Over time, the worker stops making the decision — they're ratifying it. The sense of judgment exercised diminishes. Responsibility stays intact. Control does not.
The override-hostile system. The AI produces an output. The worker can technically modify it, but the workflow is structured around the AI's output as the default, and modification requires justification, documentation, or manager approval. Workers stop overriding — not because they agree, but because the friction cost exceeds the perceived benefit. Over time, they stop noticing what they'd override if they could.
The legibility wall. The AI makes a decision the worker is responsible for, but the reasoning is opaque. The worker can't explain the output to a client, a colleague, or themselves. They're responsible for a decision they don't understand. That is a specific cognitive and psychological load that workload measurements don't capture — and it accumulates.
All three patterns produce workers who are technically doing less, in terms of cognitive labor, and burning out faster than the workload math would predict.
What the Research Says About Intervention
The ScienceDirect study's prescription is structural, not therapeutic. The researchers found that decision control had to be preserved in the workflow architecture — individual coping strategies (mindfulness programs, flexible scheduling, EAP access) did not compensate for systematic decision control loss.
This shifts the design question. The relevant unit of analysis is not the individual worker's resilience. It is the workflow's accountability architecture: at what points does the worker exercise genuine judgment? Which outputs can be overridden, easily, without friction? Does the worker understand the reasoning well enough to explain it?
For managers, the diagnostic question is simpler: if a member of your team thought the AI was wrong on a given decision, what would they do? If the honest answer is "probably nothing" or "they'd need to escalate," you have a decision control problem.
For teams designing AI augmentation, the principle is: preserve judgment capacity explicitly. Not as an afterthought or a UX consideration. As a primary design requirement. The features that let workers override, modify, and understand AI outputs aren't friction reducers. They're the mechanism that keeps the psychological contract with the worker intact.
The Exit Interview Pattern Nobody Is Connecting
Across organizations that have deployed AI augmentation aggressively, a pattern is emerging in exit interview data that HR analytics teams haven't systematically connected to AI adoption. Workers who leave describe variations of the same experience: a feeling of diminishment, a sense that their expertise stopped mattering, a disconnection from the work that predates any specific burnout event.
The language these workers use is not the language of exhaustion. It is the language of irrelevance. And irrelevance, in the SDT framework, is exactly what decision control loss produces: a persistent signal that the worker's judgment is not the point.
The engineer who started interviewing elsewhere wasn't overworked. He was experiencing the specific form of burnout that accumulates when you remain responsible for outcomes you no longer control. That pattern has a research name now. Organizations that know what to look for will catch it before the exit interview.
Related reading: AI Is Burning Out the People Who Embraced It Earliest covers the cognitive overload dimension of AI-related burnout — a companion problem to decision control loss.
Cover photo by RDNE Stock project via Pexels.