The payments industry is racing to let AI agents transact. SDKs are shipping. Rails are being built in public. And almost all of it skips the question a regulated institution asks first: who governs the agent, who keeps the record, and who is liable when it pays the wrong thing?
An SDK gives an agent the ability to pay. It doesn't give the institution a policy envelope, a spend limit, a revocation switch, or a tamper-evident trail. Those aren't features you bolt on afterward — they're the architecture the payment has to run inside.
What governed agent payments actually require
- Identity: the agent bound to a verified human sponsor and a scope — so "who paid" always has an answer.
- Policy and spend scoping: per-agent limits, merchant and category rules, real-time revocation — checked before the payment, not reported after.
- Staged authorization: the agent prepares and stages a payment within consented limits; a human or a pre-authorized rule approves. On existing rails, under the existing payment-initiation regime, this works today.
- The record: a tamper-evident trail of every agent payment — the artifact an auditor verifies and a regulator relies on.
Why a neutral layer
A bank can build this for itself — but a bank cannot be the neutral layer for other banks; it routes to itself. The governed-payments layer wants to sit above the rails, across institutions and platforms, bundled with governance and audit. Neutrality alone isn't a moat. Neutrality plus governance, audit, and cross-rail span is.
That's the architecture we're building in AORBIT™'s Agent Payments stream — the assisted tier in design-partner pilot on national rails, and a fully autonomous tier that arrives only when the regulator says it does. Agentic payments are coming either way. The institutions that win will be the ones that governed them from the first transaction.