The shallowest moat in insurance: user data in the age of AI memory
Posted:
When price comparison sites catalogue their defences against conversational AI, data is frequently mentioned. Knowing the user and their details lets a comparison site offer a slick, seamless experience: your renewal dates, your past quotes, your family's dates of birth, all pre-filled. The stored profile that makes switching painful and coming back easy. The data is a moat.
But if users can shift this information into an AI assistant in seconds, and the assistant can remember it, re-use it, and even check it for mistakes and out-of-date details, does that moat still exist?
Why the data piled up in the first place
Third-party services collected user data because users had nowhere else to put it. Re-typing your details on every insurer's website was painful, so you let the comparison site store them. The comparison site's value was really two things bundled together: aggregation (getting quotes from many insurers at once) and memory (not making you repeat yourself). And when that wasn't enough, there were always expensive TV adverts, discount vouchers and cinema tickets.
But that bundle is now coming apart. AI assistants such as ChatGPT and Claude, built on large language models (LLMs), have persistent memory. They can hold your details, preferences, household information and purchase history on your side of the relationship. When your assistant remembers that you are a family of four, that you travel twice a year, and that you always want winter sports cover, the comparison site's stored profile stops being a convenience and starts being a duplicate.
The move takes one screenshot
The most striking thing about this shift is how little effort it involves. There is no export tool, no data request, no technical integration. Walk through a real journey.
Step 1: the screenshot. John, renewing his buildings insurance, opens his comparison site's quote summary, the screen showing everything the service knows about him, and screenshots it. The page is dense with hard-won detail: a semi-detached house on Beaver Road, Manchester, purchased June 2019; four bedrooms, brick walls, slate roof, built 1900; a 5-lever mortice deadlock on the main door; patio doors with a key-operated lock; two smoke alarms, no burglar alarm; never flooded, no subsidence; £356,000 rebuild cost; sole occupant, non-smoker, insurance adviser, no claims. The comparison site spent years building this profile. It leaves in one tap.
Step 2: the LLM transcribes. John pastes the screenshot into a chat and the assistant turns it into structured data, every field and every value, as cleanly as if the site had offered an export button.
Step 3: the LLM saves to memory. John says "add all this to memory" and his home insurance profile now lives on his side of the relationship, alongside everything else his assistant knows about him.
Step 4: the LLM checks for inconsistencies. This is where the assistant stops copying the comparison site and starts beating it. "Your profile says no burglar alarm, but didn't you have one fitted last spring? You mentioned replacing the patio doors, so is the lock detail still right? The purchase date says 2019 but the no-claims years show zero. Shall we check that?" The comparison site, which has no reason to raise doubts at the point of sale, never asked. Stale security details are not a small thing either: they feed into your price and can affect a claim. The assistant's check makes the data fresher and safer than the original.
Step 5: the LLM pre-fills the next application. On any insurer's form, not just the comparison site's. When the renewal comes round, John says "sort out the home insurance" and the assistant carries the verified profile to whichever insurer or comparison engine offers the best route.
Years of profile data, copied in around 20 seconds, and improved along the way. Notice what happened to the comparison site: it still gathers quotes, but it no longer owns the relationship or the profile. It has become an interchangeable back end, called on by an assistant that holds the context. The memory advantage that drove direct traffic, repeat visits and cross-selling has moved to the assistant, via a screenshot.
Why user-side memory wins
Four things make assistant-held memory stronger than service-held data:
It covers everything. A comparison site knows your insurance history. Your assistant knows your insurance history plus your calendar, budget, family circumstances and appetite for risk. Recommendations based on the full picture beat recommendations based on one slice of it.
It stays fresh. Profiles held by a service quietly go stale between annual visits. An assistant can question its own memory at the moment it matters: the changed door, the new alarm, the mileage that is no longer right. The result is data that is both more current and more honest.
It travels. Data held by a service locks you in. Memory held by your assistant goes with you to whichever insurer offers the best deal today. The switching costs that protected comparison sites turn into pressure on them.
It is held with your consent. People increasingly resent being the product. Memory held by an assistant working for you, which you can review, edit and delete, is an easier privacy story than profiles held by ad-funded middlemen. Regulators are pushing in the same direction.
It works to your advantage. Having your data inside an LLM memory allows that LLM to scour the whole market for better deals and better products. Your no longer limited by the commercial relationships and incentives of a particular comparison service.
User data is becoming a very shallow moat
This is not the end of comparison sites. Gathering quotes, insurer relationships, regulated advice, negotiated rates and consumer trust all remain real assets, and a memory layer replaces none of them. What disappears is the idea that holding the user's data is, by itself, a defence. If a profile can leave in a screenshot, it was never a moat. It was a queue users were too tired to leave.
The defensible position shifts from knowing the user to being the best execution layer for an assistant that knows the user: the service an AI calls on because it reliably finds the right product at the best price.
That is exactly what we are building at Prompted. Rather than spending heavily on marketing and capturing enough user data to make a service sticky, we're building to help a user, or their LLM, find the right product at the best price. Companies that take this position early will treat LLM memory as a new way to be found. Those that keep betting on the data moat will discover that their most valuable asset, the user profile, quietly walked out the door, one remembered preference at a time.