Most critiques of enterprise AI stop at the technical complaint: these tools raise the floor and compress the ceiling. That framing assumes good-faith adoption. It assumes the org wants the tool to work.
The more uncomfortable possibility is that the tool working as intended is not always what the organization, or the people running it, actually want. Algorithmic outputs can be deliberately weaponized to launder managerial intent. And in some organizations, the people most exposed are not the underperformers. They are the top performers.
The inversion
The standard story: AI is a productivity tool. Adopt it in good faith, and you get gains across the curve. The top performers benefit less than the median, but nobody is harmed.
The inversion: a number is harder to contest than a judgment call. A manager's opinion can be challenged as biased. An algorithmic recommendation looks neutral. That asymmetry is what can make algorithmic output useful as a political instrument. Bad-faith actors don't need to fabricate anything. In organizations where governance is weak, they only need to select which outputs to surface, which metrics to report upward, and which exceptions to ignore.
Why top performers can be the most exposed
Top performers are the people most likely to override the tool. Reading past the pattern is part of what makes them exceptional. Closing the lead the system flagged as dead is the move that generates outsized revenue. But every override is a record.
Imagine a sales rep whose pipeline tool scores a lead as low-priority. She works it anyway, based on a signal the model can't see, and closes a large deal. In a healthy org, that's a story about judgment. In a hostile one, the same event reads differently: the system flagged this lead as low-priority and she ignored it. The act of judgment that generated revenue becomes the evidence that she is non-compliant, unpredictable, a process risk.
This is the move. The behavior that makes someone exceptional is reframed as the behavior that makes them a liability. The reframing is cheap because the algorithm gives the recharacterization an air of objectivity.
The scenario above is illustrative, not a documented case. But the structural ingredients (a model that scores, a human who overrides, a log that records the disagreement) are present in most enterprise deployments today.
The cage
Below-average performers get lifted. The tool's recommendations are better than their unaided judgment, so following the tool is a net gain. Average performers get normalized; they were already doing roughly what the model recommends, so compliance is easy and variance shrinks. Top performers get caged. Their value comes from non-obvious moves, which now require justification against a default that looks objective. In the worst case, they are targeted, with the same overrides that produced their results recharacterized as risk.
The cage is sometimes negligence, a side effect of bureaucratic process design. Sometimes it is the point. The malicious version and the negligent version can look identical from the inside, and distinguishing them usually requires evidence that is not available until after someone has already left.
How much this dynamic actually bites depends on the work. High-judgment, low-volume roles (enterprise sales, complex underwriting, senior IC engineering are the kinds of cases where this pattern would be most visible) are more exposed in principle than high-volume, low-judgment work where the model genuinely outperforms most humans most of the time. Whether any specific org is hostile or merely sloppy is an empirical question, not one this argument can settle.
What this changes
If you're building or deploying these tools, override logs are not a neutral artifact. They are a record of disagreement (the times the human said no), and they will be read by people with motives. They do not record the times the human deferred and was right to, or deferred and was wrong to. That asymmetry is structural, not a bug to be fixed in the next release. Designing the tool with this in mind, including how overrides are surfaced and to whom, is a governance decision, not a UX one.
If you are a top performer in an org adopting these tools, the question is not whether the tool is accurate. It is who controls the narrative around the gap between the tool's recommendation and your decision. The accuracy of the tool and your exposure to it are largely independent variables.
The useful frame is not is this AI good? It is who benefits when this AI's output is treated as objective, and who pays? The answers are rarely symmetric.