Every executive I coach has at least one AI recommender running somewhere in the operation. A hiring pipeline that ranks candidates. A career tool that nudges people toward "roles like yours". A CRM shortlist.
Nobody describes those as market-shaping interventions. The vendor sold them as productivity tools.
At scale, the productivity tool quietly eats itself.
The CRO in Berlin
A chief revenue officer at a B2B SaaS firm, in Berlin, three weeks ago. The CRM had been nudging reps toward "accounts like your closed-wons" for a full quarter. Top-of-funnel looked rich. Reps were working.
Her close rate on recommended accounts had slipped by a few points. Enough that she was drafting a slide about rep quality for the board.
I asked how many reps were working the same shortlist. She did not know. I asked what the recommender did when four reps all chased the same account in the same month. She said it kept recommending it.
The accounts were crowded. Every rep was chasing the same whale, and the whale had picked one of them months ago.
The recommender was the problem it was being asked to diagnose.
What the Cornell team measured
Erica Chiang, Kenny Peng, Jon Kleinberg, Éva Tardos and Nikhil Garg at Cornell partnered with New York City Public Schools. They ran a live randomised trial inside the 2025-26 high school match. 444 treatment applicants, 822 controls. Real students, real seats, real year.
A congestion-aware recommender produced a 57 percent relative increase in applicants ranking a recommended program (16.4% vs 10.5%, p=0.011). It drove a 71 percent relative increase in matching to one (5.6% vs 3.3%, p=0.071). Zero treatment applicants were rejected from a recommended program.
The simulation of what the naive version would have done is the part I keep returning to. At 20 percent student adoption, the naive recommender holds the match rate at nudged programs at 0.70. The capacitated version holds it at 0.98. Scale eats the signal at exactly the items the system steered users toward.
Increased congestion can exacerbate existing disparities, disproportionately affecting applicants with the fewest existing opportunities.
When the capacity rule was tuned to protect under-served applicants, matches from Bronx Latin-eligible zip codes rose from 9 to 25. A 2.8× swing in equity outcomes from the same recommender logic, depending on whether the capacity constraint was named. The authors put it plainly: recommendation design is policy design.
This is the third paper in two weeks where a layer bolted onto production AI behaves differently from how the vendor named it. Classifier fairness scores on Oct 3. LLM-as-judge on Oct 5. A matching-market recommender today. Different units, same shape.
Your recommender is a market-shaping intervention whether you designed it that way or not. If nobody on your side has named the capacity constraint, the default has picked one for you. The default usually costs your most-constrained cohort first.
Ask what the capacity constraint is at the next tooling review. If the vendor cannot say, the recommender is already answering.