AI readiness
Sometime in the next eighteen months, a customer's personal assistant is going to ask your business for a quote. The request will come from their software, not from them. It will put the same structured question to three or four insurers and show the customer the answers. There are three practical ways to get ready, and all three run on Subcontext today.
Consumer assistants such as ChatGPT and Claude now support third-party apps. A customer can ask about protection inside the tool they already use, and brands with an app there are the ones that get asked. For a large insurer this is defensive. For a mutual or a mid-sized provider it is a route to a mass-market audience that does not go through an aggregator or a broker panel.
We build and run your app inside ChatGPT and Claude. It can quote, check eligibility, explain an exclusion and book a callback, using your products and your rules. Every answer is recorded with a full audit trail.
Here's a quote from the YourBrand app:
Three quick health questions would confirm this price. Want to go ahead?
An agent visiting your website today has to guess. It reads the page the way a person would, works out what the forms probably do, and often gets it wrong. WebMCP is a new browser standard, backed by the major browser vendors and already shipping in Chrome, that lets a site tell an agent what it can do: how to get a quote, what the inputs mean, and what comes back. The standard is young, but it costs little to adopt and most sites will take years to catch up.
Sites and quote journeys built on Subcontext support WebMCP. The pages declare the same operations that power every other channel, so a visiting assistant gets a quote by calling your functions rather than working out your forms.
WebMCP covers agents that arrive through the browser. A2A is the equivalent for agents that call you directly: an organisation publishes a card describing what its agent can do, and another agent can open a task with it, server to server. We publish our own at subcontext.com/.well-known/agent-card.json.
Your first live agent should not be your sales flow. Pick a channel you control, where a wrong answer is recoverable and the compliance surface is small: nurturing leads that have gone quiet, following up unfinished quotes, or a win-back campaign to lapsed customers. The point is what you learn. How consent gets captured, what the audit trail needs to contain and when the agent should hand off to a human become the operating model for everything that follows. Organisations that do this well treat it as a governance exercise with a commercial by-product rather than an IT project.
Deploy an agent behind web chat, WhatsApp, SMS, voice or email, one channel at a time. Boundaries, guardrails, full transcripts and handoff to a named person are built in. Nothing goes live until someone with authority publishes it, and any release can be rolled back. Add the next channel when the first one has earned it.
A channel where a wrong answer is recoverable.
Why now
The FCA has already described this channel. The Mills Review named personal AI agents as a plausible route to closing a protection gap that leaves 15 million UK adults uninsured and 31 million more under-protected. When the regulator describes a distribution channel before the industry has built for it, that channel is arriving whether or not you are prepared. An agent shopping for protection will prefer brands that can answer structured questions and show their working.
All three moves rest on a product a machine can read: eligibility, pricing and underwriting rules held as configuration, and every decision recorded so it can be explained later. The Subcontext platform provides that. Define the product once and the ChatGPT app, the WebMCP site and the channel agent all work from the same definition, with the same audit trail behind every answer.
App, website or a small test behind one channel. We'll tell you which to fund first.