How to avoid disaster when vibe-coding a billing engine

Official Schedule Context

This talk covers what that infrastructure looks like in practice: which primitives matter, where the

human checkpoints belong, and what changes when your billing system needs to be legible to machines

instead of configured by humans clicking through a UI. When building AI products, billing and

pricing should be directly tied to the products themselves. They're in the hot path. Every token,

every agent action, every inference is a billable moment, and if your entitlement checks aren't

keeping up, a single runaway agent can rack up thousands of dollars in seconds with no one to send

the bill to. Get metering wrong and you're either eating costs or overcharging customers. Get ledger

consistency wrong and your invoices don't add up. Get tax wrong across 47 jurisdictions and you find

out from a regulator, not a user. Here's the thing, though — agents are legitimately good at billing

strategy. They can pick pricing models, configure plans, run simulations, and iterate on packaging

way faster than a human team could. You want them doing that work. But proration, multi-currency,

revenue recognition, tax — this stuff took the industry years to get right, and it's unforgiving

when you get it wrong. The question then becomes not whether agents should be making billing

changes, it's what they should be operating on when they do. Agents need tight, composable building

blocks where the correctness is already baked in, human-in-the-loop checkpoints before anything

irreversible goes out the door, and sandbox environments where they can experiment freely without

torching production. That's the architecture that lets you move fast on pricing without waking up to

broken invoices. Target audience: Engineers and technical founders building AI products that charge

for usage — whether that's per-token, per-action, or per-seat with consumption overages. If you've

ever hard-coded a pricing tier, duct-taped metering onto an existing system, or wondered how your

billing setup is going to survive your next pricing change, this talk is for you. Audience

takeaways: - A clear understanding of why billing for AI products sits in the hot path — and what

specifically goes wrong when metering, entitlements, or ledger consistency can't keep up. - A

practical architecture for making billing agent-operable: composable primitives with correctness

baked in, human-in-the-loop checkpoints on irreversible actions, and sandbox environments for safe

experimentation. - A framework for deciding where agents should be empowered to move fast on billing

strategy and where guardrails need to be non-negotiable.

No official session recording transcript was found by exact title match on the AI Engineer YouTube channel during this run.

Pending transcript synthesis when an official recording or confirmed matching video is available.