Companion to the Harbor brief · secondary research · 9 September
This scan exists to answer one question before proposing anything: has anyone already tried competing on comprehension at the moment of decision? Everything here is secondary — desk research, clearly separated from the field evidence in the brief.
The long-running industry bet: price the individual more accurately using observed driving. Every major carrier now runs a program.
Growing ~34% CAGR inside automotive OEM platforms. Deloitte estimates that if 20% of US personal auto moves to embedded by 2030, $50bn in premium shifts away from traditional channels. Dealers adding insurance to the sales process report a ~20% lift in F&I gross profit.
Automating intake, triage and settlement — the cost side.
A layer of AI products aimed at converting purchased leads faster, and at catching quotes that expire before anyone follows up. Aimed at the funnel, not at the buyer’s understanding.
The share of premium paid back out in claims ran around 80% through the 1970s and 80s. It reached 62% across property & casualty in 2024, and 55% in homeowners in 2025. Advertising and agent commissions take 16.3% of premium. And in 2025 the sector earned $112bn investing premium it had not yet paid out, against $69bn from underwriting itself. The money is made holding the premium, not settling the claim. That is the arithmetic underneath §05’s question — and it is a reason the result can look like this without anyone hiding anything.
secondary Figures as stated by the report’s author in a published interview. I have not read the report itself or checked the numbers against NAIC filings, and I would not assert them in front of someone who has.
The insurtech cohort competed almost entirely on cheaper, faster, prettier. The results are public:
| Where it stands | |
|---|---|
| Lemonade | Cumulative losses since founding around $1.27bn; FY2025 net loss ~$165m on ~$738m revenue; gross combined ratio still above 100% |
| Hippo | Gross combined ratio around 130% for four straight quarters; the original homeowners book still loss-making |
| Root | Same cohort, same long climb to profitability |
The stated lesson from the sector is that AI alone could not fix it — these companies began without the historical data incumbents had, and needed to buy that data through losses. A decade and well over a billion dollars has been spent proving that faster and cheaper does not, by itself, win personal lines.
This matters for the intervention. The obvious moves — a slicker quote, a lower price, a faster bind — are the exact moves that cohort made.
I have to report this against my own finding rather than for it.
There are more tools promising to read a policy than at any point in history — tabulating coverage, flagging gaps, reasoning about what a policy does when a claim lands. But the ones being marketed are overwhelmingly built for insurance agents, not for consumers. The comprehension layer is being sold to the intermediary.
Policygenius pairs AI quoting with a licensed human advisor who reviews coverage afterwards to catch gaps. That is the nearest existing thing to the intervention this brief points at — and it needs a person to work.
States began mandating policy readability in the early 1980s, with Flesch reading-ease requirements. Forty years of mandated clarity, and the comprehension problem is undiminished. That is strong evidence the fix is not better prose in the contract.
This is the item my first scan missed, and it is the strongest thing in it. Consumer-permissioned insurance data APIs are a mature commercial category. Canopy Connect — which markets itself as “Plaid for insurance” — returns structured P&C policy data, coverages, limits, deductibles and declaration-page PDFs from 300+ carriers, about 95% of the market, in under thirty seconds, off a sign-in the consumer performs themselves. InsurGrid and MeasureOne sell adjacent versions.
Who pays is not the difference, and it matters to be exact about that. The buyers are agents, lenders, mortgage originators, dealerships and carriers — the same side of the table that would fund this. The difference is which way the answer points. Today the customer signs in and the data goes to the seller, who quotes, underwrites or verifies with it; the customer never sees a comparison. The infrastructure that would let a shopper compare what they hold against what they are being sold is built, cheap and generally available — and in a decade nobody has built the version that hands the answer back.
That closes off every explanation that depends on difficulty. It is not a capability gap. It is an allocation of one. Which is the brief’s finding, arriving from the supply side instead of the evidence side.
secondary Vendor-published coverage figures, not independently verified. The claim that matters is not the percentage — it is that the capability is general, commercial, and has never had its output addressed to the buyer.
Nothing in this scan shows an insurer positioning on we will make sure you understand what you are buying, before you buy it. Carriers compete on price, brand, speed and claims service.
About a third of auto shoppers now use AI while shopping. About a third of them find it unhelpful. And the ones who use it are significantly more likely to switch.
Read that three ways. People are already trying to close the comprehension gap themselves. They are reaching for a general-purpose tool because no insurer offers them one. And when it works, it moves them — which is the only evidence in this whole scan that comprehension changes binding behavior.
secondary Survey data, not observed. But it is demand demonstrated and supply failing, in the same statistic.
The ground is narrower than “nobody does this,” and it is still open.
The strongest version has to survive one question, and it is the question a partner will ask: if this were valuable, why hasn’t an incumbent done it? The honest answer available from this scan is that incumbents have no incentive to make a renewal legible — a legible renewal is a renewal a customer might decline. That is a reason a challenger can act on and an incumbent cannot, which is exactly what a new entrant needs.