The boring 80% nobody warns you about when an AI demo becomes a real product

Short summary

Moving an AI product from demo to production isn't blocked by ML—it's blocked by billing, fraud control, and accounting. This post covers five major pitfalls: usage metering and ledger reconciliation, free-tier abuse, acquisition attribution tracking, affiliate payouts, and supporting infrastructure. The authors open-sourced their solution as Velobase Harness (MIT, Next.js + Postgres).

• •Usage-based billing requires separating billing events from model calls and reconciling balances to the cent
• •Prevent free-tier abuse with rate limits, CAPTCHAs, disposable-email blocking, and credit clawbacks
• •Track acquisition attribution server-side (stamp source at signup, fire conversion from backend) for reliable budget decisions

Generated with AI, which can make mistakes.