On the loop and the surface, with the user cap kept as a smoke alarm. Ulrik Lehrskov-Schmidt (via Jason Lemkin, October 23, 2026) reports 1% of users driving 50 to 75% of AI costs and recommends brakes moved in until about 4% of users are affected; Kevin Stumpf at Databricks caps at maybe $500 a day against a $300 power user; LangSmith's gateway returns a 402 per user or key. By Lehrskov-Schmidt's own rule that cost and value are independent, a brake on the heaviest users lands on users, and the Larridin cohort at around $1,300 a week was the one at 11.8x output. The unit that runs away is the loop, so write a stop rule per loop, attempts, spend, and the name it escalates to, and read the per-outcome line per surface, where a 380x customer is a negative-margin workflow and gets a routing row. My own loop without a stop rule drifted 14% in five weeks, and the flat-usage half of the base was 12 accounts on plan, two lost at sunset.
On the loop and the surface. The user cap stays as a smoke alarm, set above anything a real customer has done.
The three brakes side by side, the independence rule that moves the kill point, and the vendor-side numbers are in 1% of Users, 75% of the AI Cost. The Brake Goes on the Loop, Not the User. The three fields are from The Stop Rule. The per-outcome line and the Larridin cohorts are in Three Token Bills, One Unit.
Last reviewed 2026-10-23 · 1 min read