Shipping AI features without burning your budget

The failure mode nobody warns you about

The scary AI bug isn't a wrong answer — it's a retry loop quietly spending $400 overnight, or a provider outage taking your dashboard down with it. AI features fail in ways CRUD features never do.

Three guardrails worth wiring once

1. A budget per product, not per request. Generation runs against a monthly allowance tied to the plan. Hit the cap and the user sees a clear limit message — not a surprise invoice.
2. Log every generation. Each call writes a usage event (model, tokens, cost). That's your audit trail when spend spikes and your input when you tune tiers.
3. Switch providers, don't hardcode them. Routing through the AI SDK means a failover is a config change, and a model deprecation doesn't become a rewrite.

Graceful degradation beats a clever prompt

When the model is slow or down, the right behavior is a soft failure: keep the rest of the page working, surface a retry, never block a billing flow on an AI call. VelocityKit wires the AI todo assistant this way by default — copy the pattern for your own features.

The point isn't to use less AI. It's to make AI spend observable and bounded so you can ship more of it with confidence.