12-CONTROL IMPLEMENTATION GUIDE
Agentic commerce security checklist
A production checklist for the moment between an AI agent choosing a purchase and your system executing the order or payment.
Published by MandateShield · Updated July 24, 2026The boundary in one sentence
Re-evaluate the final purchase against explicit human authority immediately before execution, outside the language model, and retain the outcome.
Convert intent into explicit limits
Store a numeric maximum, currency scope, permitted merchant or payee and a short expiry.
Identify the acting agent
Use a stable agent ID and protocol-native credential binding instead of a display name.
Bind the final merchant
Compare the actual seller or payee at checkout with the approved merchant scope.
Evaluate the final total
Include taxes, fees, shipping and currency before checking the hard spending cap.
Prove current consent
Tie explicit human consent to the same mandate and material checkout facts.
Keep authority short-lived
Reject expired mandates and review unusually long authorization windows.
Bind intent to execution
Carry a digest of the approved intent or cart so material changes cannot pass silently.
Enforce outside the model
Do not ask the language model to decide whether its own payment should bypass policy.
Consume each attempt once
Persist an idempotency key behind an atomic uniqueness constraint across retries and regions.
Fail closed on missing evidence
Use REVIEW or BLOCK when identity, intent, merchant or currency evidence is absent.
Retain minimal decision evidence
Store decision, reason codes and receipt without card numbers or personal shopping data.
Keep fraud and processing separate
Run authorization alongside—not instead of—fraud, sanctions and processor controls.
Turn the checklist into a test
Score the current architecture, then run a live purchase against overspend, merchant substitution, expired authority, missing consent and prompt injection.