Key Takeaways: Managing AI Costs with Ulrik Lehrskov-Schmidt

Big thanks to everyone who joined Office Hours last week.

This was one of our most popular sessions ever, and for good reason.

brought the heat with an absolute clinic on how to price, protect margins, and not lose your shirt as AI eats into cost structures.

Breaking down my 5 top takeaways below for those who couldn't make it:

1️⃣ The "double hockey stick" is the real AI pricing trap

We're all used to the idea that value is unevenly distributed across customers — some get massive value, most get medium, a few get none. What Ulrik nailed in this session is that AI introduces a SECOND hockey stick: cost distribution. He shared a real example from a client with a couple hundred million in ARR where 75% of their entire GenAI cost was driven by just 0.1% of users. Thousands of users driving three-quarters of all consumption.

The critical insight: there's no natural law that says your value hockey stick is proportional to your cost hockey stick. You can build a beautiful AI product that turns dollars into dimes. Or you can be like

@Steven Meyer

at Placer.ai, serving customers ~$50K/week of value at $150 in cost. Most of us land somewhere in the middle, but you have to KNOW where, and you only learn by putting product in the market and measuring.

2️⃣ Define fair use in dollars, not actions

This was the most tactical part of the session. Most companies set fair use policies as "more than 500 API calls per month" or some action count. Ulrik's argument: that metric will be obsolete the second your model changes or your prompt structures shift. Track REAL COST per user in real time, and trigger fair use at dollar thresholds.

The framework: bucket your users (malicious vs non-malicious, new vs tenured, monthly vs annual, by vertical, whatever matters). Build your fair use policy along a dollar cost line. Then run different policies per bucket and tune to the margin you want. He has customers where they literally use fair use policies to "decide the margin of the product" — turn the dial, get the gross margin you want.

3️⃣ Stack throttles like a customer journey, then measure each one

Best practice for fair use isn't ONE throttle — it's a series. For example: at $30 of cost, send a notification. At $100, cut speed in half. At $200, sever the connection. Or some variation that fits your product.

At each throttle point, four things happen: customers churn, customers ignore and keep going, customers adjust their behavior, or customers upgrade. Track the distribution of those four outcomes at each throttle and you have a pricing experimentation dashboard. Move the throttle from $30 to $40. Watch what changes. This is exactly what you do on a pricing page with conversion rates — same logic, applied to consumption.

4️⃣ Outcome-based pricing isn't the holy grail — fairness is

Everyone wants to be Intercom/Fin with "per resolved conversation." That's a great model when you have a single, clearly-defined outcome. But for most horizontal or multi-faceted products, chasing outcomes is the wrong frame.

Ulrik's reframe: what you actually want is for customers to ACCEPT your pricing metric. You can get there two ways — tie it to a clear outcome, OR make it so fair and so transparent that they don't care. Electricity is the perfect example. You pay per kWh because it's fair, not because there's one outcome. If you're an input to a bunch of value creation processes across many use cases, your job is to price the input WELL — not to invent an outcome that doesn't naturally exist.

5️⃣ Figma is the AI pricing case study to watch right now

@Farhan Manjiyani

asked who Ulrik admires for their AI pricing, the answer came fast: Figma. The numbers back it up. In their most recent earnings, 75% of their pro and enterprise accounts are already consuming credits above included limits. The average ACV for credit-consuming accounts is 3x the ACV for seat-based accounts.

The pricing power was already there in their enterprise accounts — the seat model just wasn't tapping into it. Now they have a monetization engine. And they're doubling down with MCP integrations, repositioning design as enterprise infrastructure rather than a tool for designers. This is the playbook every legacy SaaS company should be studying right now.

Huge thanks again to Ulrik for joining us, and to everyone who showed up and asked sharp questions.

Looking forward to the next one.

Rob

0 comments