Agent payment protocols: what's converging
Multiple industry proposals for how AI agents should initiate payments have emerged in 2025-2026. Where they agree and where they don't.
Several proposals for how AI agents should initiate payments have emerged in 2025 and 2026 — from the card schemes, from wallet providers, from the AI vendors themselves. The proposals differ in the details. What's converging is the shape.
Written September 2026 from a survey of published specifications and implementation notes.
The convergence
Nearly every serious proposal agrees on:
- A mandate object as the primary abstraction — an issued permission that the agent presents at transaction time.
- A signed intent from the principal, verifiable independently of the merchant or the agent.
- A step-up mechanism where the principal reconfirms consent for out-of-mandate transactions.
- An audit trail that can be produced independently by both the agent and the merchant.
The disagreements are on transport (bearer token vs. cryptographic proof), on scope (per-merchant vs. per-category), and on step-up mechanics (out-of-band notification vs. inline confirmation).
What integrations should build for
- Mandate as the abstraction, not the transport. Whatever scheme wins, the mandate concept is stable. Build against that.
- Independent audit trail from both sides. The merchant's log of what the agent claimed, and the agent's log of what it did, should be reconcilable. If they aren't, disputes will be impossible.
- Explicit refusal handling. An agent that presents an invalid or expired mandate should be refused with a clear reason, not silently degraded to a different flow.
What to avoid
Locking your integration to one vendor's proposal. The 2027 landscape will look different. Building for the shape rather than the specific spec is the durable choice.
Agent payments are one of those areas where the discipline of shipping cautiously matters more than the discipline of shipping quickly.