Your next customer is an AI assistant: what MCP means for insurers and service businesses
By Vishal Kumar, Founder and CEO · Last updated · 8 min read

In short: More and more people ask ChatGPT, Claude or Gemini before they ever visit your website. MCP (Model Context Protocol) is an open standard that lets those assistants use your services directly: get a quote, start a claim, order a certified valuation. An API (application programming interface) lets software talk to your systems. An MCP server does the same job but also tells the AI what is on offer and what it needs to provide. Insurers and other service businesses that publish one can be found, and used, inside the conversation.
Take a photo of a painting. Upload it to Claude. Ask what it's worth.
You'll get an answer. A reasonable one, even.
But you can't insure a painting on a chatbot's opinion. You need a valuation from a qualified appraiser, the kind that stands up when you make a claim.
That gap, between "the AI gave me an answer" and "I have something with legal standing", is where a lot of service businesses will win or lose customers over the next two years.
I explained it at dinner last week to a founder in the insurance space:
"Earlier, your customers were human. What's happening more and more is your customers are humans through ChatGPT, through Claude, through Gemini."
What is MCP, in plain words?
MCP stands for Model Context Protocol.
The official definition: "MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems." The project's own analogy is a USB-C port for AI applications: one standard way to plug things in. [1] Claude and ChatGPT both support it. [1]
My plain version:
"MCP is nothing but a bridge, or a door, for an external AI to come and use our services."
You run an MCP server. An AI assistant connects to it. Now the assistant can do things with your business on the customer's behalf, instead of just talking about you.
| Step | What happens |
|---|---|
| Customer | Asks their AI assistant |
| AI assistant | Claude, ChatGPT and others support MCP |
| MCP server | A door with a menu: Get a quote for fine art cover; Start a claim; Book a valuation |
| Your service | Does the part that carries legal or professional weight |
What is the difference between an API and an MCP server?
You probably already have an API, or your vendor does. So why another thing?
An API is a set of commands one piece of software uses to talk to another. It assumes the caller already knows what to send. A developer reads the documentation, writes code, and the two systems talk.
An MCP server is built for a caller that hasn't read the documentation. Here's how I put it:
"An MCP is like an API, but with a self-explanatory menu. So it's like: oh, you want to insure? Here is what I need. You want to run a claim? Here is what you should upload. So it's self-instruction."
That self-description is the whole point. The AI assistant can look at your menu, see what you offer and what each service needs, ask the customer for the missing pieces, and call you.
No developer in the middle. No integration project for every new partner.
If your team says "we already have an API", they're half right. The API is usually the engine. The MCP server is the shopfront that AI assistants can walk into.
| API | MCP server | |
|---|---|---|
| Assumes | The caller already knows what to send | The caller has not read the documentation |
| How it is used | A developer reads the docs and writes code | The assistant reads a self-explanatory menu |
| New partner | An integration project each time | No developer in the middle |
What does "insurance as an API" mean?
It means turning an insurance product into something software can buy.
One of the two platforms we're building for the US market does exactly this. An art appraisal firm values artworks for collectors. Once a collector knows a piece is worth, say, US$50,000, the next question is how to insure it. The firm wanted to offer cover on its own website instead of sending people elsewhere.
The chain behind that one button is long. A carrier in the UK willing to write fine art cover. Its partner in the US. A wholesaler. A broker of record. Then the appraisal firm. Most of them run on email and spreadsheets.
We partnered with the managing general agent (MGA) in that chain, the specialist firm authorised to sell on the carrier's behalf. The deal is simple to describe:
- The MGA gives us the product specification and rating rules.
- We turn them into an API that can return a quote.
- The appraisal firm's website calls that API at the point of valuation.
- The same rules can then sit behind an MCP server, so an AI assistant can ask for a quote too.
Once the product is an API, it can be sold anywhere: on the appraisal site, by a broker's portal, or inside a customer's AI assistant.
Why would an AI assistant need a certified answer?
Because some answers need someone accountable behind them.
An AI assistant can estimate what a painting is worth. It can't sign a valuation an insurer will accept, and it can't issue a policy.
The appraisal firm's view, as their CEO put it to us, is: yes, your AI can answer the question, but if you want a certified appraisal, which has legal standing through an expert, connect to our service via MCP.
That pattern applies well beyond art:
- Insurance: a binding quote, a policy document, a claim lodgement.
- Valuations and inspections: a signed report from a qualified person.
- Certification and compliance: a certificate that a product, site or person meets a standard.
- Professional services: a booked appointment with a licensed adviser.
In each case the AI handles the conversation. Your business provides the part that carries legal or professional weight. MCP is how the two connect.
What should a service business put behind its MCP server?
Think of it as opening a store for AI shoppers. My shorthand for it:
"First, you need an MCP that basically is like your online store. It's like Shopify for agents."
Here's a practical order of work:
- Be findable. AI assistants recommend businesses they know about. That means long-form content on your site, reviews and mentions elsewhere, and structured answers. People call this AEO (answer engine optimisation) and GEO (generative engine optimisation). Without discovery, nobody connects to your MCP server.
- List your services as clear actions. "Get a quote for fine art cover", "Start a claim", "Book a valuation". One action per thing a customer actually wants.
- Say exactly what each action needs. Documents, photos, values, dates. The AI will collect them from the customer.
- Decide where a human steps in. Anything that needs judgement, a licence or a signature goes to a person, and the AI tells the customer that clearly.
- Log every call. What was asked, what you returned, and what the customer agreed to.
- Add payment last. Agent-initiated payments are still early. Get the first five right before you let an assistant pay.
The tempting choice and the better choice
build a chatbot for your own website. It feels modern. But it only helps people who already found you, and they have to leave their own assistant to use it.
make your services usable from the assistants your customers already have open. Then your chatbot, your website and every AI assistant all draw on the same set of actions.
What are the risks?
An MCP server lets outside software act on your systems. Treat it like any other public entry point.
- Authentication and permissions. Decide who can call which action, and what an unauthenticated assistant can see. The MCP project publishes security best practices worth reading before you build. [2]
- Regulated advice. In insurance, a quote is fine; personal advice may need a licensed person. Design the menu so the AI can't cross that line.
- Wrong or stale data. If your rating rules change, the MCP server must change the same day.
- Cost. Every AI call costs money somewhere. Keep responses short and structured. Our guide to reducing AI costs covers this.
- Data location. If customer documents pass through your server, the same privacy rules apply as for any other channel. Our private AI guide goes deeper.
How soon will AI assistants buy on behalf of customers?
My view, and it is a view, not a forecast with data behind it: sooner than most people think.
The scenario I describe to founders goes like this. Someone tells their assistant, "I just bought this laptop. Get it insured at the best rate." The assistant finds three providers, compares them, picks one with good reviews and a fair price, buys the cover and emails the policy.
I expect purchases made entirely by agents to start showing up in the second half of 2027. When a founder at dinner suggested business procurement would come first, my answer was: "You'll be surprised at how quickly it will come."
The businesses that will be ready are the ones that can already be found, and used, by an assistant.
For the bigger picture of where this is heading, read my piece on AGI and the agentic operating system.
- 2nd half of 2027
- when Vishal expects purchases made entirely by agents to start showing up
- Vishal's view, not a forecast with data
Your next step
If a customer asked their AI assistant for your service today, what would it find, and could it do anything about it?
Write down the three actions a customer most often wants from you. That's the first draft of your MCP menu. Get your AI roadmap if you want help turning it into a working server.
Frequently asked questions
What does MCP stand for?
Model Context Protocol. It is an open-source standard for connecting AI applications, such as Claude and ChatGPT, to external data, tools and workflows.
Is MCP the same as an API?
No, but they work together. An API lets software call your systems. An MCP server describes your services to an AI assistant, including what each one needs, so the assistant can use them without custom code.
Do insurers need an MCP server now?
Not every insurer needs one today. But if customers already ask AI assistants about cover, quotes or claims, an MCP server is how you show up as an option they can act on, not just read about.
Can an AI assistant bind an insurance policy?
Technically, an assistant can call a quote or purchase action if you offer one. Whether it should, and when a licensed person must be involved, depends on the product and the jurisdiction. Design those limits into the menu.
How is this different from SEO?
SEO (search engine optimisation) helps people find your pages. AEO and GEO help AI assistants find and cite you. MCP goes one step further: it lets the assistant use your service once it has found you.





