The burden of proof has moved onto the deployer.

The EU AI Act and the UK's Smart Data Framework both push in the same direction: transparency, auditability and explainability for automated decision-making. Institutions can no longer point at a model and describe its output as an unavoidable property of the system.

What supervisors are asking for is narrow and concrete. Who authorised this agent. What limits was it operating under. Why did it act as it did. A firm that cannot answer those three questions with tamper-evident evidence is carrying an exposure that sits with a named individual, not with a department.

Standards are still forming. Firms that adopt a protocol-level answer now avoid the cost of retrofitting one into deployments that have already scaled.

The preceding years were defined by copilots — systems that suggested, with a human making the commitment. The shift underway is to agents that act, and the difference is entirely a governance one.

Financial infrastructure was built for human-to-human or API-to-API interaction. It has no negotiation layer: no space where two competing agents at two different institutions can settle a price or hedge a risk without exposing proprietary information or creating a compliance and anti-trust problem in the process.

Nothing in the existing stack fills that role, because nothing in the existing stack was designed for a counterparty that has to prove its own authority.