A manager's opinion can be called biased. A score cannot. That asymmetry is the story of algorithmic management in the enterprise.
The previous piece argued that AI tools trained on population data encode median judgment as ground truth, and that this structurally caps the top of a sales team. That framing assumes good-faith adoption. The structural problem is real, but it is not the whole problem. There is a political one underneath, and it is uglier.
Algorithmic output is an excellent laundering instrument. It takes managerial intent, runs it through a model, and returns the same intent dressed as objectivity. A sales lead classified as low-priority by a scoring model looks like a fact. "I don't think she should work that account" looks like a judgment. The first is far harder to contest than the second, even when both express the same preference.
This is not hypothetical machinery. It is the natural behavior of any organization given a tool that produces numbers. The bad actor does not need to tamper with the model. The model is fine. What the bad actor controls is selection: which outputs get surfaced in a review meeting, which metrics get pulled into the quarterly dashboard, which override gets flagged as non-compliance and which gets quietly ignored. The model is the exhibit. Selection is the argument.
Top performers are the most exposed
The uncomfortable implication is that the people most at risk are the ones the organization least wants to lose.
Top performers override tool recommendations more often than anyone else. That is most of what makes them exceptional. They read past the pattern, close the lead the model wrote off, escalate the account the routing rule filed away. Every one of those acts produces a paper trail: tool said X, the rep did Y. In a healthy culture that paper trail is a feature, because the outcome vindicates the judgment.
In a hostile culture the same trail is evidence. The argument writes itself: the system flagged this lead as low-probability and she worked it anyway. The system recommended pausing outreach on this account and she escalated. Her pattern of non-compliance has been consistent across the quarter. None of those sentences are false. They are also a case for managing her out, built entirely from the behavior that generated her revenue.
A manager cannot say "I don't like her judgment" without being challenged. A manager can say "she is consistently at odds with the tool" and sound reasonable. The laundering is complete before anyone notices the input.
Three outcomes, and one of them is the point
Revisit the three-outcome picture with this in view.
Below-average performers get lifted. Average performers get normalized. Top performers get caged; in the worst case, they get targeted. The first two outcomes are what vendors pitch. The third is rarely discussed, partly because it is awkward to discuss and partly because it is not always accidental.
Sometimes the cage is negligence: a manager adopts the tool's frame because it is easier than thinking. Sometimes the cage is structural: the routing layer absorbs the top rep's best leads before she ever sees them. And sometimes the cage is the intent. A manager who has wanted a particular person gone for reasons unrelated to performance now has a rationale that survives HR review. The tool did not tell anyone to do this. It does not need to. It just has to produce enough numbers to point at.
What this means for buyers
The implication for anyone evaluating enterprise AI is not "do not deploy it." The implication is that deployment is a governance problem, not a modeling problem, and the governance questions are harder than vendors make them sound.
Three are worth asking before signing.
First, who decides what counts as a tool override, and who reviews those decisions? If the answer is "the same manager whose quota the rep affects," the audit trail belongs to one side of the table.
Second, are overrides tracked against outcome, or against compliance? A system that measures "percentage of recommendations followed" is measuring obedience to the model. A system that measures "outcome of overridden recommendations" is measuring judgment. Only the second tells you whether the tool is actually helping.
Third, what is the mechanism for contesting a tool-generated flag? If an employee cannot push back on a score in a way that creates its own paper trail, the score is not information. It is a verdict.
These questions do not make a deployment safe. They make it legible. That is the floor, not the ceiling.
The broader pattern
Algorithmic management did not arrive with AI. Performance dashboards have been used for this purpose for twenty years. What AI changes is the surface area and the plausibility. Scores used to cover pipeline, activity, and attainment. Now they cover judgment itself: which lead to call, which objection to handle, which customer to escalate. Anywhere a rep previously exercised discretion, a score can now sit where the discretion was, and any deviation from the score is reviewable.
The defense is not a better model. It is an organizational commitment to treat algorithmic output as prior, not as verdict, and to make that commitment expensive to walk back. Without that, the tool does exactly what its best users fear it will do: it gives the organization a cleaner way to do what it was already inclined to do.
A number cannot be accused of bias. That is the feature the vendor sells. It is also the feature that should make any serious buyer pause.