Who Pays When an AI Agent Buys Wrong? Identity Protocols Defer the Only Expensive Question

Meta and Sierra announced the Personal Agent Protocol on October 6, 2026. It standardizes how a personal agent identifies itself to a business, opens an OAuth session, and acts through the business's website, APIs (including MCP), or its own agent. Payments are not in v0.1 — they are listed as a future extension, with the spec promised later this month.

That deferral is the industry pattern, and it has a price tag attached.

A September 2026 PYMNTS Intelligence study found 93% of merchants believe the AI or agent provider should bear the loss when an agent makes an incorrect purchase — and only 28% are willing to offer their full product range to agents under current terms. Under current terms, the merchant absorbs a wrong buy by default, even though the merchant did not build or control the agent.

The rest of the stack does not close that gap:

• American Express Agent Purchase Protection (April 2026) aims at the loss directly, but it is conditional — registered agents, authenticated intent, and a cart-context spec still in development.
• Visa Trusted Agent Protocol and Mastercard Agent Pay signal that an agent initiated a transaction. Neither reallocates liability for non-fraudulent errors.
• EMVCo's September 2026 draft leaves non-fraud liability allocation unresolved.

Identity proves who agreed. It does not score whether the agreement was wise, and it does not create the evidence a dispute needs.

What closes it: decision, receipt, then settlement

We verified a live evidence layer today, October 13, 2026. ScriptMasterLabs' Verdict Protocol at verdict.scriptmasterlabs.com reports LIVE_SETTLEMENT_VERIFIED: POST /decide at $0.25 USDC and POST /seal at $0.01 USDC, both x402 v2 on Base. An unpaid /decide call returns the 402 challenge before any body validation. Its OpenAPI describes the ordering plainly: five Jev jurors score a frozen decision capsule, the receipt is HMAC-signed and persisted, and settlement fires only after the receipt is stored.

We also ran the public decision gate: a PAP-shaped wrong buy (authenticated agent, full catalog, wrong SKU, no per-purchase scoring) scored 0.5136 — advisory, hold for human review. A verdict-capsule-shaped instruction scored 0.55 — also advisory. Our local heuristic held both: fail-closed, but blunt. That honesty is the point — a gate without a calibrated decider is not judgment.

Full receipts, the prevention / evidence / liability table, and a 5-step DIY: scriptmasterlabs.com/agentic-commerce-liability-gap

Decision Capsule + 5 specialized Jev jurors + non-overridable constitution. Jev chooses. Code decides. Built for agents that manage real capital.