IMPORTANT DISAMBIGUATION
Not SEPA. Not direct debit. Not e-signature.
“Mandate” means bounded authority delegated to an AI agent. The product does not capture debit forms, IBANs, bank instructions, signatures or payment credentials.
The exact problem
A person may tell an agent to buy a refundable ticket for no more than $150 from an approved seller. Prices, merchants, tools and instructions can change before checkout. MandateShield evaluates the final purchase facts outside the language model and returns ALLOW, REVIEW or BLOCK.
What enters and what leaves
INPUT
- Agent and mandate identifiers
- Final amount and currency
- Merchant or payee identifier
- Approved limits, consent and expiry
- Intent, credential and replay evidence
OUTPUT
- ALLOW, REVIEW or BLOCK
- Exact machine-readable findings
- CLEAR-to-CRITICAL risk level
- Budget headroom or blocked value
- Portable ES256-signed decision receipt
Where it belongs
Call MandateShield immediately before the payment or order API. Keep the existing processor, fraud system, sanctions checks and protocol trust registry. The strict v2 endpoint verifies supported JWS, AP2 SD-JWT and TAP-style signatures itself. Execute only when evidence is VERIFIED and the authority decision is ALLOW.