Introducing Malcolm

AI agents are already having millions of insurance conversations every day. Consumers are asking ChatGPT which home policy to pick, whether their car cover is fit for purpose, what income protection makes sense for the self-employed. The questions are regulated-product questions, and the volume is only going one way.

The problem is that very little of the infrastructure beneath those conversations is built for regulated products. Today, an AI assistant fielding an insurance question is mostly pattern-matching on public web content: brochureware, old blog posts, forum threads. There's no direct line to the insurer's actual pricing engine, no view of current underwriting appetite, no record of what was said or how a quote was constructed. For a casual query, that's fine. For a regulated transaction, it's a structural gap.

That's why we're building Malcolm.

The gap

When an AI agent today recommends an insurance product, three things need to be true, and currently none of them reliably are:

• Accuracy - does the product actually cover what the agent says it covers, at the price the agent implies?
• Compliance - does the recommendation meet the regulatory standards a human broker would be held to?
• Connectivity - can the customer move from conversation to bound policy without dropping back into a legacy funnel?

Close any one of these and you have a better demo. Close all three and you have a distribution channel.

What Malcolm does

Malcolm sits between AI agents and insurers as a compliant middleware layer. Every AI-generated quote is validated against the carrier's own product rules in real time. PII is stripped in-flight. Underwriting questions are routed back to the carrier's systems rather than answered by the model. Compliant checkout URLs are generated automatically, and every interaction leaves an immutable audit trail.

The shape of the integration matters. Insurers have spent decades investing in pricing engines, underwriting logic, and product IP. An AI distribution layer that bypasses that work - guessing at prices, paraphrasing policy terms - is a liability, not a channel. Malcolm is designed so that the carrier's existing systems remain authoritative. The language model orchestrates the conversation; the carrier's engine generates the quote.

What we believe

A few principles shape how we build, and they're worth naming up front:

• Zero hallucination on financial data. Language models are powerful orchestrators, but a hallucinated quote isn't a UX bug - it's a regulatory failure. Pricing has to come from deterministic, carrier-verified systems.
• Carrier sovereignty. AI platforms will become distribution channels, but they must not become proxy underwriters. Carriers keep control over risk appetite, pricing logic, and how their products are communicated.
• Compliance as code. Regulation isn't friction to route around - it's the bedrock of consumer trust. Audit trails, suitability checks, and duty-of-care standards belong inside the protocol, not bolted on after the fact.

We set these out in more detail in our Agent-to-Agent Manifesto.

What's next

We're working with leading UK insurers and backed by Founders Factory. Over the coming weeks we'll be sharing more about our technical architecture, our compliance framework, and the integrations we're building with early carrier partners. We'll also be writing about the broader shift in distribution - how AI assistants are changing where insurance intent surfaces, and what that means for anyone retailing regulated products.