ChatPRD founder Claire Vo walks through inference abuse, the controls she added, and the customer friction they created in this 40-minute How I AI episode.
Seat pricing still exposes a compute bill
- Vo says token fraud cost her bootstrapped business tens of thousands of dollars over roughly a year.
- Her six observed patterns are trial farming, promo-code abuse, card fraud, resale relays, custom-MCP abuse and model probing.
- Examples include a map-making studio consuming trials and a reseller proxying ChatPRD to its own paying customers.
- A custom MCP, a connection from an agent to external tools, turned useful integration access into a persistent compute worker.
Put limits where the money leaves
- Vo closed free trials and is abandoning promo codes, sacrificing a source of sales leads.
- Enforce account and organization caps, upfront billing, model allowlists and rapid feature shutoffs.
- Monitor dollars at least hourly or live; message counts alone hide expensive-model consumption.
- Use payment rules and card fingerprints to detect repeated trials; review agent-written rules in the payment dashboard before enabling them.
- Her roughly $200,000 card-testing incident involved failed attempted purchases, not revenue or a loss of that amount.
Verify clients, tools and support permissions
- Vo uses request-level bot checks, firewall rules, authentication protections and restrictions on custom MCP hosts.
- Valid credentials can still power an unauthorized relay; bot detection adds a signal rather than proving authorization.
- Metadata thresholds trigger Jev classification; an hourly review agent investigates unusual signups and activity.
- Give support fraud context and response playbooks while restricting its ability to restore access or issue promotions.
- Feed each new abuse pattern back into rules and operational checks.
Enforcement has a customer cost
- Vo acknowledges one corrected false positive and the legitimate customers excluded by broad country blocks.
- She distinguishes commercial resellers from people seeking unaffordable AI access, while saying she cannot subsidize either indefinitely.
- Caveat: this is a founder’s operating account. The episode supplies no classifier accuracy, thresholds or controlled comparison of defenses.
“If you sell AI, you sell compute.”