Agent Mediated Insurance (AMI) Standards
What are the AMI Standards?
The AMI (Agent-Mediated Insurance) standards are a proposed open specification for how an AI agent, acting on a customer's behalf, can quote, compare, and buy a regulated insurance product from any carrier. They define both the message format (a canonical request and response) and the conduct rules a compliant sale must follow (ground every quote in authoritative data, disclose material terms, minimise personal data, and log an audit trail). They are designed to provide a common language for agents to interact with underlying APIs built by different insurers to allow AI agents to naturally surface insurance products to users while remaining compliant.
Who is Marrow?
Marrow is a startup building the rails for agent mediated insurance, they are currently live in ChatGPT and binding policies at enterprise scale. Marrow has published the first open draft of AMI at v0.1 informed by our in-production development. While Marrow stewards the drafting, we claim no ownership of the interface. We are seeking leading insurers to co-author future versions, provide technical challenges, and signal adoption. The goal is to create a common language for the whole market to ensure insurance companies can benefit from the future of insurance distribution while remaining compliant.
Read the technical definition
Why are we proposing the AMI standards?
1. The rise of agent-mediated insurance
Over the past year, consumer AI surfaces have moved beyond conversation to include app stores, embedded products and agentic checkouts. Insurers all across the industry are starting to explore these new possibilities and avenues for distribution. The demand and importance of AI as a distribution surface is no longer hypothetical. Customers are already being invited to discover, compare, and begin buying insurance inside a conversation, and carriers, aggregators and MGAs are already rapidly releasing ChatGPT apps to serve this developing customer base.
This is the shape AI insurance distribution is taking today, but at maturity it presents a fundamental shift in how insurance is distributed. As we interact with AI agents to plan our lives and purchases, insurance products should be surfaced naturally, embedded in the moments where they are relevant, rather than waiting to be found.
This places AI agents in a regulated position - they will be introducing and selling insurance products to users. By proposing a universal and neutral architecture, these proposed standards anticipate regulation by the FCA and seeks to adequately prepare insurers for it.
2. Accountability and Consumer Protection
While the A2A and AP2 standards provide a useful starting point for establishing buyer identity, identifying that an agent is acting on a person's behalf and authorising payments - they only cover part of what a regulated insurance sale requires.
For regulated insurance products, applying the existing protocols raises questions that they are not yet equipped to answer:
- Lawful basis: What mechanism establishes and records that explicit, per-source consent was obtained before sensitive personal data was processed, in a form that satisfies UK GDPR?
- Grounding and accuracy: How can a carrier or regulator verify that the premiums, terms and exclusions an agent presented to a customer came from authoritative sources, rather than being hallucinated by the model?
- Disclosure: What evidence exists that the customer was shown the material exclusions, conditions and price before they committed, and that they understood them? Who is accountable if they were not?
- Suitability and fair value: If an agent recommends or selects a product, who is responsible for assessing whether it was suitable for that customer's circumstances and represented fair value under Consumer Duty?
- Accountability at the point of claim: If a customer makes a claim and finds their cover does not apply as they understood it, who is liable? The customer who delegated the purchase? The developer of the agent? The carrier who accepted the bind? The platform that hosted the interaction? Without a defined framework, this question has no clear answer.
- Regulatory supervision: How does the FCA supervise a market in which every agent integration makes its own choices about consent, disclosure and suitability? Without a common evidence base, there is no consistent way to assess whether regulated obligations were met.
These open questions lead to concrete risks. Without a common standard, the market fragments and the protection a customer receives depends entirely on the choices made in whichever integration they happen to use. Agent platforms and carriers must each negotiate their own framework for consent, disclosure and liability, multiplying cost and ambiguity across the market. And without a consistent audit trail, there is no reliable basis on which the FCA can supervise the channel.
By publishing these proposed standards we hope to open a discussion with the industry, understand the needs of multiple stakeholders and come together to define a set of standards that work for everyone.
Guiding Principles for a Trusted Agent-Mediated Insurance
The design of these proposed standards is rooted in five core principles.
I. Grounded in certainty
Language models are powerful tools for conversation and orchestration, but they are not built to calculate risk. A premium, an exclusion, an eligibility decision: these are not things an agent may estimate or infer. Every material fact an agent presents must trace to an authoritative source - a carrier's pricing engine, an enrichment bureau, a disclosed document.
II. The customer's right to safe delegation
Customers have the right to delegate their financial administration to an AI agent without fear of silent mis-selling. An agent acting on a customer's behalf must be held to the same standards of clarity, transparency and duty of care as a human broker. This means mandatory disclosure before commitment, suitability assessment under Consumer Duty, and protection for vulnerable customers. These are not optional features; they are preconditions for a valid sale, enforced at the rail level before a bind can occur.
III. Carrier sovereignty and risk integrity
AI platforms will become major distribution channels for insurance, but they must not become proxy underwriters. Carriers retain absolute control over their risk appetite, their pricing logic and how their products are communicated. The standard defines the interface, not the underwriting. A carrier maps its products once to the canonical schema; everything behind that boundary remains entirely theirs. The architecture is designed to protect the carrier's core IP and underwriting integrity, not to abstract over it.
IV. Compliance as architecture
Regulation is the foundation of consumer trust in insurance, not an obstacle to efficiency. In the agentic era, compliance must be encoded directly into the protocol rather than delegated to individual implementations. Every regulated action - consent, enrichment, disclosure, attestation, bind - produces an immutable, tamper-evident audit record. The evidence that a policy was lawfully sold does not depend on any single participant's record-keeping; it is a structural output of a conformant transaction.
V. Open protocols over walled gardens
The infrastructure connecting AI agents to regulated insurance must be open and interoperable. A single closed platform cannot and should not become the gatekeeper of regulated distribution. These standards are published as open drafts, stewarded by Marrow on behalf of the working group, with no proprietary claim over the interface. Any authorised agent running on any surface, and any carrier or MGA that maps to the canonical schema, can participate. The goal is a common language for the whole market, not a competitive moat for any part of it.
Developed to create standards for open and secure AI mediated insurance
As insurance purchases move from websites to AI agents, the industry requires a new set of standards to ensure openness and compliance. Marrow has developed the Agent Mediated Insurance Standards to define a common language for how agents and insurers transact. We built AMI as a collaborative, open source standard and we welcome your feedback on how to improve it.