Your Agent is Too Expensive: Taking Control of Your AI Bill
Your AI bill is the fastest-growing line in your P&L, and most teams can't say what one request costs them.
Agents make it worse: one user action becomes 3 to 30 model calls, each with its own context, tool calls, retries and reasoning tokens. The curve isn't linear, and credits only hide it for a while.
This is a hands-on evening for pre-seed and seed founders and early technical employees who are building with LLMs today.
No product slides.
Three people who see hundreds of startup AI bills will tell you where the money actually goes and what the teams that fix it do differently.
Then we split into small rooms so you can bring your own numbers.
Speakers
Philip Andersson, Applied AI @ OpenAI
Andy Kane, Specialist, GenAI Models @ AWS
Andree Hultgren, AI Engineer @ Lovable
Moderated by Deina Kellezi, Investor at byFounders.
What we'll cover
1. Know your cost per outcome
Cost per task as a first-class engineering metric. Tag spend by feature and tenant, set budgets and alerts, attribute per key and service. Catch the runaway test script and the leaked API key before the invoice does.
2. Cut tokens without cutting quality
The levers in effort order: model right-sizing, cascading and routing, prompt caching and batch, context trimming and output caps, prompt compression, fine-tuning or distillation. And the rule that makes them safe: no cost change ships without an eval.
3. Agents: where the bill goes non-linear
Per-run budgets and step limits, circuit breakers on loops, tracing cost per trajectory rather than per call, managing context across steps, and evaluating on outcome plus cost. Closing with the hard question: how do you price something whose cost varies 10× per task?
Agenda
16:30 Doors open, drinks
17:00 Panel (30 min)
17:30 Breakouts: each speaker hosts a room on their topic, you choose the room. Bring your questions.
18:15 Walking dinner and drinks
21:00 Close
Who should come
You're a founder, CTO or early engineer at a pre-seed or seed company, you have LLM calls in production (or will within a quarter), and you'd rather argue about trade-offs than watch a demo. ~50 people, approval-based, so the rooms stay small enough to be useful.
Before the night
When you register we'll ask one question: what do you want answered about your AI bill? The most common ones go straight to the speakers.
Hosted by AWS Startups and The Shack.