AI compliance layer

In the A2A protocol, the compliance layer is formally called the Rail. The agent talks to the Rail; the Rail talks to the carrier. This separation means the agent never directly accesses a carrier's systems, a data enrichment source, or a payment handler. All such interactions are mediated and governed by the Rail, which is purpose-built to enforce compliance without re-implementing it in each agent.

The Rail performs five distinct functions. First, enrichment: it calls authoritative data sources such as DVLA, MyLicence, CUE, BCIS, and Companies House to populate the risk schema, keyed from the canonical identifiers the customer provides. Second, validation: it checks every field in the assembled quote request against carrier-defined product rules before making any carrier call, blocking hallucinated, estimated, or out-of-appetite values. Third, PII protection: it strips, tokenises, and protects personal data before it enters AI model context or logs, with heightened protections for special-category data. Fourth, compliance enforcement: it enforces the disclosure-and-acknowledgement gate, the step-up confirmation requirement, and any applicable signposting duties. Fifth, audit: it writes an immutable, tamper-evident record of every enrichment call, every enriched-versus-attested fact, every disclosure, and every authority exercised.

Marrow is the AI compliance layer for UK insurance. The compliance, enrichment, and audit implementations behind the A2A open standard are Marrow's; the interface standard itself is open.