The customer doesn't know what they don't understand

We have been saying the same thing in insurance for years. Customers don’t understand the policy they buy. They don’t read the schedule of cover. They find out what is excluded at the moment of a claim. Then they say, “I thought this was covered.”

That is a fair observation, and it is the industry’s oldest gap.

Over the past year a second gap opened next to it. Customers now fail to understand the policy and also fail to understand where it came from. The second one is harder to see, because it does not show up in any claim file yet.

Buying insurance now fits inside one conversation

For a long time, buying insurance worked like this. You searched online, collected quotes, compared them, decided, and bought. Several steps, several screens, and usually an intermediary somewhere in the middle.

All of that now fits inside a single conversation. The customer asks an AI assistant, gets an answer, and decides right there.

The examples have moved past the pilot stage.

In Spain, a digital insurer called Tuio moved policy sales directly into an AI assistant. Six months later, more than a fifth of its new customers were arriving through that channel. Mapfre then bought 39% of the company. Mapfre bought a sales channel that did not exist eighteen months earlier.

In the US, Insurify quotes inside the same assistant across eleven states, with live prices from more than one carrier. The customer sees their own price, sees the payment schedule, asks coverage questions on the spot, and moves onto the path that ends in a policy.

In the UK, Aviva put its home insurance quote in the same place. Compare the Market did the same. The second one carries extra weight, because Compare the Market is already a regulated intermediary. A firm under state supervision and a piece of software under no supervision at all now sit side by side on the same screen. To the customer they look identical.

The market took this seriously. Intermediary shares fell in February 2026 after the Insurify launch. Willis Towers Watson dropped 12%, its worst session since 2008. Arthur J. Gallagher lost 9.9% and Aon 9.3%. The S&P 500 Insurance index closed the day down 3.9%.

A startup called Waniwani began selling the plumbing: a kit that puts any quoting company inside these assistants, with a paid compliance layer on top. EverQuote took a minority stake in it in August 2026.

No regulator has set a rule for this channel yet. The market is already charging for something it calls compliance. The industry has decided on its own that responsibility lives here.

The line between information and advice is disappearing

Insurance has an old, clean distinction. Giving information is free. Giving advice needs a licence.

A price comparison site falls under supervision because a real sale happens through it and it earns money from that sale. An AI assistant answers questions and passes on information. It sells nothing and gives no personal recommendation, so it stays outside supervision.

The distinction holds up on paper. In the field it is losing meaning fast.

Take Brokly. You enter your trade and your location, and it tells you which insurance the law requires you to carry. It reads the insurance clause in your contract and shows you where your current policy falls short. That function sells you nothing. It tells you what to do.

Is that information or advice? And can we expect a customer to draw that line?

The sales side has stopped being clear too. “AI can’t complete a sale” no longer holds. In banking, a platform called Meow lets AI agents open a company account, issue cards and move money from inside assistants like ChatGPT and Claude. The technical plumbing for the insurance version exists as well. It is waiting its turn.

The customer sees none of these distinctions. They ask, they get an answer, they decide. “Does whoever gave me this answer owe me anything?” never enters their head.

One study shows how far this has gone. A fifth of UK adults say they would hand financial decisions, savings and borrowing included, over to an AI. They are not describing a second opinion. They are describing delegation.

The same study makes a technical point. An AI answer can look reasonable and still be wrong. Models drift, and their reliability drops over time. Producing a confident answer is easy. Producing a correct one, and proving it, is hard.

What the tests show

Independent tests of these apps have started appearing. Two of them cover the ground well.

The first walked through Insurify’s ChatGPT app across four turns of conversation. The firm running the test works in the same space, so set its score aside and look at the findings. All four land on the subject of this article.

ChatGPT turned the comparison data into a direct instruction: “Go with State Farm, don’t over-optimize.” That reads as advice, no longer as comparison.

The model produced percentages of its own invention: “80% chance State Farm is your best deal.” No source sits behind that number and no method either.

The driver details collected in the conversation, meaning age, accident history and whether the person owns a home, did not carry over to the website. The user had to start again there.

The disclaimer arrived after the recommendation. Not before the decision.

The second test asked whether you can buy car insurance through ChatGPT, Claude and Gemini. None of the three can quote on its own, since they hold no live carrier connections. Through the Insurify app, though, the tester received quotes inside the chat and bound a policy with the carrier they picked. In their own words, they never had to use Insurify’s website at all.

The gap between the two tests matters. The February version gave a rough price and sent the customer elsewhere. Insurify added personalised pricing, payment schedules and a path to bind in August. The “don’t worry, it only gives information” defence lasted six months. Working out where it lands in the next six takes no imagination.

The customer doesn’t know what they gave up

A customer who sits down with an intermediary has a structure behind that conversation, whether they notice it or not. The person opposite has to recommend a product suited to their needs. They have to keep a record of the conversation. If something goes wrong, there is somewhere to complain. The intermediary carries professional indemnity cover, so a mistake gets paid for from somewhere.

The customer knows none of this. All of it exists.

Ask the same question of an AI assistant and none of it exists. The customer doesn’t know that either.

Companies have already spotted the gap on their own side. Klaimee’s founders built the business around one question: when an AI acts on its own and gets it wrong, who pays for the damage? They review and certify the software first, then insure what is left. Chaucer and Armilla launched a product covering losses caused by AI. Vouch covers losses from information an AI invents.

Responsibility is already changing hands. It moves from company to company, through insurance. On the consumer side nothing corresponds to it.

One possibility gets little airtime. These tools may sell better than a person. They stay patient, stay consistent, never tire, and adjust to how the customer speaks within seconds. Steering someone into more cover than they need requires no bad intent. Well-built software is enough.

This is not the end of the intermediary

I can guess my colleagues’ first question. If this channel grows, is there room left for the intermediary?

There is. But not if the intermediary keeps doing today’s job the same way.

The market panicked in February and then walked it back. Analysts called the selloff overdone, and their reasoning describes the intermediary’s ground well. KBW read the integrations as lead generation in personal lines for now. Evercore ISI pointed to the data and scale an intermediary holds, which both inform risk management decisions and win better terms from insurers. TD Cowen expects any pressure to arrive gradually and stay modest.

The same technology squeezes the intermediary and strengthens them. The difference sits in what the intermediary sells. I have argued before that the right question is how AI strengthens the intermediary, and that the real competition is over the customer interface. This channel puts both arguments on the same screen.

Software already handles price comparison. The part it cannot copy is different: understanding the customer’s real situation, asking the right question, standing beside them at the moment of a claim, and owning the outcome. The value was always there. Losing the convenience half has made it visible.

A shared rule works in the industry’s favour

The industry’s first reaction here is to defend. Bring in a rule and innovation stops, costs rise, we slow down.

I think the opposite.

Waiting for everyone to do the right thing voluntarily does not work. Take one insurer or intermediary who stands up alone and says, “in this channel I will spell out to the customer what they are buying and what they are not.” That firm lengthens its own process, makes the sale harder, and falls behind its competitors. Explaining loses. Staying quiet wins.

Nothing changes in that equation until a rule binds everyone at the same time. Expecting the industry and consumer representatives to close this gap by agreeing among themselves does not look realistic to me after nearly twenty years in the business.

A shared rule does something simple. It brings everyone to the same line and stops punishing the firm that behaves well. Insurance history has plenty of examples, from compulsory covers to standard policy wordings. None of them shrank the industry. All of them grew trust.

The limit of the rule needs stating too. The goal is not a ban on the chat interface; it is that the customer can see what the answer they got is, and what it is not. A rule like that does not choke innovation. It gives it legitimate ground.

“Did the customer understand” is now measurable

The industry’s favourite objection runs like this: customer understanding cannot be measured, the whole thing is too abstract.

That objection has aged too.

A London startup called Consentfully pulls out the important clauses of a policy and puts the customer through a short comprehension check. At the end it produces a record you can show a supervisor. The “I have read and accept” box gives way to a document showing what the customer took in.

I see that as an opportunity. The same AI can simplify policy language and explain to a customer what falls outside their cover. An industry that uses this tool only on the selling side and not on the explaining side will widen the gap instead of closing it.

Markets where this channel has not arrived yet, mine included, still have time. Building the discussion from scratch after the fact is avoidable if the thinking starts now. Three questions belong on the table before the fact: whether intermediation rules cover this new interface, what a disclosure obligation means inside a chat window, and which protections the customer loses there.

Where I stand

I am not against this channel growing. Whatever is easy for the customer wins, and this may be the easiest interface anyone has built.

My objection sits elsewhere. Selling convenience while removing the protection without saying so.

A customer understanding what they bought is not a decorative line in a responsibility statement; it is the operating condition of this business. A customer who says “I thought this was covered” at the moment of a claim takes a piece out of the industry’s trust.

We did not solve this in the policy wording. Now we have to solve the same problem on a screen where the customer decides without asking us at all.

Being late there will cost far more.