AI in Commerce1 min read
Consent, mandates, and trust in autonomous agent payments
How an agent can pay on your behalf and audibly prove why. A short taxonomy of the guardrails, and where each is easy to get wrong.
The gap between "an agent bought me groceries" and "an agent drained my card at a gaming site" is entirely a design gap. Both flows use the same underlying rail. The difference is the guardrails, and — critically — whether every decision left an auditable trace.
Written September 2026.
The minimum guardrail set
At the smallest useful shape, an agent payment mandate must express:
- Which agent (identity binding)
- On whose behalf (principal binding)
- In what currency
- With what per-transaction cap
- Cumulative caps: daily, monthly
- Explicit merchant category allowlist and blocklist
- An expiry
- A step-up threshold above which fresh consent is required
Deny loudly
The temptation, when a check fails, is to short-circuit the pipeline: "MCC not allowed, deny." Don't. Run every check, produce a verdict per check, and only then decide.
There's a working example at /payment-lab/trust-simulator.