AI shopping agents can read what a brand claims about a product.
But they still can't tell which claims have supporting evidence, where that evidence came from, and what's not verified.
TrueChoice makes product evidence understandable to AI agents.
https://t.co/OC48xs0fR0
I keep hearing the same question about agentic commerce: who pays when an AI agent buys the wrong thing?
Right now, agent purchases follow the same liability rules as any other online card purchase:
• Buyers: Protected from fraud. But if their agent misinterprets an instruction they gave it, the purchase was technically authorized, so fraud protections don’t apply. Their fallback is a return or a dispute.
• Merchants: Pay for most online fraud and incur chargeback fees in the case of a dispute. Merchants also tend to be liable for the gray-area disputes when an agent gets an order wrong.
• Card issuers: Take on the fraud loss when a purchase was properly authenticated.
• AI platforms: Absent from card network rules. Their terms of service typically put the risk of agent mistakes on the user.
The problem with keeping these rules as-is: agents will make mistakes humans wouldn’t and most of that cost will land on merchants. Buyers will get hit to some extent too, with returns and disputes. Yet, the AI companies that built the agents' core decision-making logic won't carry any liability.
New liability rules for agents may come eventually. Until then, a few stopgaps are emerging:
• Network signals and tokens: @Visa’s Trusted Agent Protocol lets merchants confirm they’re dealing with a registered agent, not a bot. @Mastercard Agent Pay gives each agent its own payment token, and Mastercard’s Verifiable Intent creates a record of what the user actually approved. This mostly helps merchants: they can trust agent traffic and have proof of what the customer signed off on if a dispute comes up. It doesn’t change who pays yet, but it gives everyone better evidence.
• Agent purchase protection: @AmericanExpress has committed to covering cardholders when a registered agent buys something that doesn’t match what they approved.
• Scoped credentials: Agent payment tokens can be limited to a set $ amount or merchant, which caps how much damage a single mistake can do.
As agentic commerce volumes ramp up and existing liability frameworks are tested, more edge cases will emerge and we'll likely see more debate between what constitutes agent fraud vs agent mistakes.
If you’re building liability frameworks, or insurance solutions to protect merchants and buyers until these exist, reach out! Would love to hear from you.
Anthropic’s IPO filing is a milestone for AI.
But as agents move from making answers to decisions, intelligence isn't enough.
They need evidence, not just plausible outputs.
The next infrastructure layer makes real-world evidence machine-readable, attributable, and verifiable.
@abrilzucchi Felt. Except I had to do it for safety.
I got my first job at 13.
Started living on my own at 15.
Graduated hs a year early so I could support myself financially.
I have to remind myself sometimes when I get caught up comparing myself to ppl who come from supportive homes.
Money is moving faster than the evidence underneath it.
AI-driven purchases grew over 15x this year.
Meanwhile a supplement seller was sentenced to 70 months in federal prison for forging lab certificates.