1% of Users, 75% of the AI Cost. The Brake Goes on the Loop, Not the User.

Three posts this week put a brake on AI spend, and all three put it in the same place. Ulrik Lehrskov-Schmidt's, via Jason Lemkin at SaaStr this morning, is the one with the numbers, so it goes first, and it is also the one whose own rule says the brake is in the wrong place.

The short version

In one window, three roster sources shipped a runtime brake on AI spend. Ulrik Lehrskov-Schmidt, as Jason Lemkin reports his session on October 23, 2026, sees about 50% of users at near-zero AI consumption and 1% driving 50 to 75% of costs, and recommends a kill point plus brakes moved in slowly until about 4% of users are affected, which he says addresses 60 to 80% of the cost problem; companies recovered up to 22 points of gross margin that way, with no new pricing. Kevin Stumpf at Databricks (October 22) recommends a hard cap somewhat above the typical power user, maybe $500 against $300 a day, and cutting frontier access at 70% of a team's budget. LangSmith's LLM Gateway sets spend caps per organization, workspace, API key, and user and returns a 402 to the agent at the cap. All three brakes land on the person or the key. By Lehrskov-Schmidt's own rule that cost and value are independent, a brake on the heaviest 4% of users lands on users, and Larridin's deeply AI-native cohort, as Lemkin reported it in September, was at around $1,300 a week and 11.8x output with no ceiling. The unit that runs away is the loop (LangSmith's own example is one agent making 10,000 calls into a four-figure invoice), and the unit that carries margin is the outcome. Keep the kill point and move it: per loop, a stop rule with attempts, spend, and a name; per surface, the per-outcome line. The 50% at near-zero are the other hockey stick, and from the vendor side that was 12 accounts on plan, paying minimums, two lost at sunset.

What Lehrskov-Schmidt has that nobody else published

Lemkin's write-up is a recap of a session by Lehrskov-Schmidt, founder and CEO of Willingness to Pay, who has worked on pricing full time for about ten years. The figures below are his, as Lemkin reports them, and I did not watch the session.

The shape he calls the double hockey stick: customer value follows a power law, and with AI, cost does too. Flat pricing used to hide that because every customer cost roughly the same to serve. Across a couple dozen high-growth AI pricing projects his team regularly sees about 50% of users with zero or near-zero AI consumption, 1% of users driving 50 to 75% of total AI costs, and in extreme cases 0.1% driving 75% or more. One client's least profitable customer was spending 380x more in tokens than it paid in revenue. Another client with roughly $500K ACVs had its largest enterprise accounts at about $20K a year in AI cost, and he told them that was a good business and to keep going.

His rule under all of it: cost and value are independent. High cost does not mean high willingness to pay, and the ratio moves as models, prices, and behaviour change. The four combinations get company names. Clay, spiky value and flat cost. Figma, spiky on both. Netflix, flat on both. Higgsfield, flat value and spiky cost, with $19, $49, and $99 tiers listing unlimited features.

Then the process, which is the part people will copy. Segment users, with new users getting headroom and fraud handled separately. Set a kill point, the usage level where the product shuts off. Set brakes that reduce utility at a threshold. Communicate through email, in-product, or the MCP. Move the brake in slowly, touching 0.01% of users, then 0.1%, watching who continues, uses less, churns, or upgrades. At around 4% of users affected, he says, a company might have addressed 60 to 80% of its cost problem, and then a second, stronger brake goes in at the kill point. His evidence that it works: companies that went from profitable to unprofitable after adding AI and recovered up to 22 points of gross margin within weeks, with throttles that affected a few percent of users and no new pricing. Figma built fair usage limits before launch, then turned on credit counting, then enforcement, then monetization; gross margins fell from 92% to 82% across that stretch and growth accelerated at more than $1B in ARR after credits. Clay added action credits as a second pool and its heaviest users now pay several times what they paid before.

On agents, his line is that charging per agent seat "worked for about five minutes," because agent output has no range the way human output does. Price on what the agents do.

That is the most useful public account of the cost side I have read, and I am going to argue with one word in it.

Two more brakes, same place

Kevin Stumpf, who leads Unity Gateway at Databricks, gave the enterprise version in an interview the day before. Set the hard cap somewhat above what your typical power user spends, his example being $500 against $300 a day, so a looping agent gets caught without capping a legitimate heavy day. Prefer graduated responses to an abrupt shutoff: alerts, then reduced capacity, then, in his example, cutting off default access to frontier models once a team has burned through 70% of its budget and routing the rest to open-source models. His estimate is that roughly 5 to 10% of tasks need a frontier model.

Martha Janicki and Ramon Petgrave's post on the LangSmith LLM Gateway gives the plumbing. Time-bound spend caps at four levels, organization, workspace, API key, and user. When a cap is hit, the agent receives a 402 with a clear error. A custom header identifies each end customer so a reseller can cap per customer on one key. The runaway case that motivates all of it is one agent making 10,000 LLM calls and producing a four-figure invoice.

Read the three side by side and the unit is the same in each. Lehrskov-Schmidt brakes users. Stumpf caps users and teams. LangSmith caps keys and users. The brake goes on the person.

The person is the wrong unit, by his own rule

Cost and value are independent. That is Lehrskov-Schmidt's sentence, and it is the reason the brake cannot go on the user.

A brake set at the heaviest 4% of users lands on users, selected by cost. By his rule, that selection says nothing about their value. Some are the 380x customer. Some are the account he told to keep going. The brake cannot tell them apart, because it is reading the one axis he just said is independent of the other.

I have a number for who is in that 4%. Three Token Bills, One Unit carries the Larridin cohort data as Lemkin reported it in September: same company, same tools, same starting spend, and the deeply AI-native engineers reached 11.8x output at around $1,300 a week with no ceiling in sight, while the low-AI cohort stayed flat at about 1.9x across a 20x spend range. The heaviest spenders in that data were the ones producing. A brake on the top 4% by spend lands on them first.

And the 50% at near-zero are the other hockey stick, the one on the revenue side. I have that one from the vendor's chair. Lovable Isn't Ready for Outcome Pricing. Here Is the 21-Month Bill is the $40M ARR migration where our tracker had 12 accounts classified as on plan. They were paying their committed minimums and using less than the minimum. Nobody escalated, because the dashboard said fine. When the legacy tier sunset, we lost two of them. A flat-usage user is breakage with a renewal date, and no brake touches it, because no brake is looking there.

The unit that runs away is the loop

Look at LangSmith's own example again. Ten thousand LLM calls and a four-figure invoice. That is not a heavy user. It is one agent looping.

A user-level cap catches that at the invoice, or at $500 if Stumpf's cap is set, which is still the invoice with a smaller number on it. The Stop Rule catches it at attempt three. Three fields, written before the loop runs: attempts, spend, and the name it escalates to. Salesforce's Agent Designer ships with two fix attempts and a $3 ceiling before it stops and escalates. My own pricing agent had no stop rule, drifted 14% under benchmark in five weeks, and took two quarters to recover. Stumpf's graduated response is the stop rule's escalation column with a budget where the name should be, and the budget is the weaker of the two, because a budget does not know who to call.

The unit that carries margin is the outcome, and that is where the 380x customer belongs. Per surface, the line from Three Token Bills: outcomes this month, tokens per outcome, cost per outcome, realized price per outcome, gross margin, with a red dot where margin is negative or falling. On that line the 380x customer is not a person to slow down. It is a workflow with a negative margin, and the fix is a routing row, a cheaper model where the outcome tolerates it, a stop rule where the loop is retrying, or a price where the outcome is real and underpriced. 41 Brands Price AI Five Ways. The Credit Is the New Seat has the three numbers that last fix needs: the exchange rate, the burn per action, and what expires unused.

So keep the kill point. Keep the sequence, 0.01% then 0.1% then watch, which is the best description of a careful rollout I have seen in a pricing post. Move the kill point from the user to the loop and the surface.

One more thing about the 402. It lands mid-task. A brake on the user is a half-finished job at 3 a.m., and the person who finds it at 8 does not know which step it stopped on. A stop rule on the loop stops at a step it can name, with a name attached to the escalation.

Under SaaS to AI Business Models the argument is that the price follows the work and the margin follows the price. The brake has to follow the same unit as both, and the unit is the outcome and the loop that produces it, never the person who asked.

This week: find where your first AI spend cap is set. If the field says user or API key, write the stop rule for your most expensive loop instead, attempts, spend, name, and put the per-outcome line next to it for that surface. Then set the user cap, somewhere above anything a real customer has done, as the smoke alarm it should be rather than the brake.

Related answer: Where should the brake on AI spend go, the user or the loop?

Sources: Jason Lemkin, "The Double Hockey Stick: Why 1% of Your Users Now Drive 75% of Your AI Costs, and What Figma, Clay, and Higgsfield Are Doing About It," SaaStr, October 23, 2026, read in full; every figure attributed to Ulrik Lehrskov-Schmidt, including the Figma, Clay, and Higgsfield examples, is as Lemkin reports his session, and the session itself was not watched. "The new economics of AI cost management," an interview with Kevin Stumpf, Databricks blog, October 22, 2026, read in full. Martha Janicki and Ramon Petgrave, "LangSmith LLM Gateway: Runtime Controls for Agents," LangChain blog (the page is dated July 30, 2026; the feed carried it on October 22, 2026), read in full. Falk Gottlob, "Three Token Bills, One Unit" (September 28, 2026), with the Larridin cohort figures as Jason Lemkin reported them there; "The Stop Rule" (September 29, 2026); "Lovable Isn't Ready for Outcome Pricing. Here Is the 21-Month Bill"; "41 Brands Price AI Five Ways. The Credit Is the New Seat" (October 15, 2026).