AI Divergence Is Not Differentiation

There is a piece of advice working its way through enterprise AI rollouts right now. Audit your AI for convergence with competitors’, and engineer divergence as the strategic goal. The diagnosis is right. The prescription points at the wrong target.

Patrick van Esch, Yuanyuan Gina Cui, and J. Stewart Black describe what they call the Agentic Convergence Trap. Independent AI agents training on overlapping data, optimizing similar objectives at machine speed, reach near-identical decisions.

Researchers studying German retail gasoline markets found that when both competing firms deployed AI pricing agents, margins rose 38 percent across the market — without any data sharing between them. The federal antitrust action against RealPage in 2024 cited the same dynamic in housing. The pattern is documented in retail, hospitality, airlines, and housing, and it is accelerating.

The authors propose a four-step fix: keep humans in the loop on specific decisions, define what your AI optimizes beyond the platform default, feed your AI data competitors cannot access, and measure convergence as a governance metric. Each step is sensible. The metric the fix optimizes for — divergence — is not the metric that produces competitive position.

Different is not better

Divergence is one of four steps the authors prescribe — and the only one that gets measured. The trouble is with what gets measured, not with what is prescribed.

An AI strategy that produces decisions unlike competitors’ may be different and better, or different and worse. Engineering a custom objective function and feeding the model unusual data will produce different outputs. One of three things will be true of those outputs.

They will be more accurate readings of the market than competitors’ AIs produce. They will be less accurate. Or they will be accurate in a way the market does not yet reward.

Only the first is competitive advantage. The other two are failure modes.

A divergence metric does not distinguish between them. “Decision correlation with observable competitor behavior over the past 90 days” — one of the metrics the authors propose — tells you whether your AI sees the world differently than competitors’. It tells you nothing about whether the world your AI sees is closer to the market’s behavior. Optimizing for divergence optimizes for an indicator that may or may not track competitive position.

What AI competitive advantage requires

AI competitive position takes three forms — each harder to engineer than divergence, each producing divergence as a byproduct rather than a goal.

Data competitors cannot observe

Uber’s surge-pricing advantage over Lyft rests on 61 billion historical trips across more than 10,000 cities — a behavioral dataset Lyft’s smaller fleet has never matched. The advantage is not divergence — it is a sensor one company has and the other does not. The same model architecture trained on Uber’s data and on Lyft’s data will produce different decisions because the inputs differ. The advantage is upstream of the algorithm. Inventing the technology is the same kind of upstream asset — it does not automatically convert when downstream economics fight the deployment.

Objective functions that capture longer-horizon value

Starbucks’s Deep Brew optimizes for visit frequency and long-term relationship depth, with transaction value treated as a downstream consequence. A competitor optimizing for check size on the same data will see the same customer and surface a bakery upsize. Neither system is malfunctioning — they are optimizing different things. Compounded across millions of interactions, the objective choice — not the algorithm — produces different customer relationships.

Human judgment at the right decision points

Step one of the prescribed fix names this; the next three under-weight it. The decisions where an AI’s recommendation will be identical to every competitor’s are the decisions a human should re-enter the loop — not because humans are smarter than the AI, but because they introduce variation the system structurally cannot. The override, the strategic patience, the “we are not doing this even though the model says we should.”

Why the divergence audit will be adopted

The four-step fix will be adopted regardless of whether it builds competitive position. Divergence is measurable. “Decision correlation with competitor behavior over 90 days” is a number that fits on a slide. Boards adopting AI governance frameworks under regulatory pressure want metrics that look like risk management, and divergence audits look exactly like that.

Boards do not want to hear the harder questions.

What data do we have that competitors cannot replicate? What objective should we be optimizing that no one else on this platform would? Where do we still want human judgment in the loop?

The answers commit a company to inconvenient investments. Years of upstream data work that has not yet paid off. Custom objective functions vendors charge extra to implement. Decisions to slow down for human review when speed was the point of the deployment.

Antitrust defense is the second reason the audit will be adopted. The DOJ and FTC have begun citing AI-mediated convergence as evidence of coordination — RealPage was the named example, but the underlying theory generalizes. A company that runs a divergence audit can demonstrate to regulators that it was monitoring for collusion-by-learning. That is legal protection, and it will drive adoption regardless of whether the metric tracks competitive position.

The metric being purchased is governance hygiene and regulatory cover. It should not be confused with strategy.

• • •

The authors are right that AI agents trained on overlapping data converge. The fix they prescribe will produce measurable divergence on a governance dashboard. That divergence is not the same as competitive advantage.

An AI that produces different conclusions from competitors’ is one thing. An AI that produces conclusions competitors’ AI cannot produce is another. The data, the objective, or the human in the loop has to be upstream of theirs.

Competitive position in AI lives upstream of the algorithm.