Here is the pitch you will hear at your next renewal meeting: underwriting is going agentic. The pricing software lives inside the core platform, the agent calls outside risk data at the moment of decision, and insurers running it are improving loss ratios by three to five points. It sounds like a reason to expect better terms — or to accept worse ones and call it the future.
One question cuts the pitch down to size: on which named book, tested how? Nobody can answer it yet. That is not an argument against the technology. It is an argument for reading the claim like a submission: the number is only as good as the evidence under it.
This piece works through why, drawing on the Connected Insurance Hub’s November 7 analysis, “Agentic Underwriting Has a New Owner. Its Evidence Still Has to Be Earned” — that analysis in broker terms, plus five questions for any renewal where an agent helped set the price.
What actually changed
On October 27, 2026, Guidewire signed a definitive agreement to acquire hyperexponential, the London underwriting and pricing software company, for $1.7 billion in cash, plus up to $400 million more if recurring revenue targets are met. The release names Aviva, Allianz, Sompo, Beazley, Markel, HDI, Cincinnati Financial, Convex, and Bowhead as customers, and says the platform has supported underwriting decisions across more than $75 billion of annual commercial P&C premium. Guidewire frames the deal as positioning it to “lead the transformation to agentic underwriting.” Closing is expected in Guidewire’s second fiscal quarter of 2027.
On October 28, Verisk announced connectors that let eligible customers call Verisk content inside general-purpose AI assistants — including permissioned, property-level hail and wind history through a connector for Claude. Existing entitlements still apply: you can only call what your subscription covers. No pricing or availability dates were stated.
So the direction of travel is real. But what it establishes is a fact about the software market: a platform buyer valued underwriting workflow at $1.7 billion. A deal price is evidence of what a buyer expects to earn from selling software. It says nothing about what an insurer earns from using it — still less about what your client should pay.
The figure under the microscope
Now the three-to-five-point claim itself — the number that will end up in renewal decks. The figure comes from a January 2026 hyperexponential article stating that commercial P&C insurers implementing agentic AI are achieving loss ratio improvements of three to five percentage points. The Hub’s finding is blunt: the figure is attributed inconsistently — including to McKinsey work on advanced analytics, a broader category than agentic AI — and published with no methodology at all. It is vendor-produced. As published, it sits at the bottom of the proof ladder: anecdote, a percentage that cannot be tested from what has been published.
The honesty tell is the Hub’s own record: its Technology Radar repeated the figure in March without the caveat — and has since corrected it.
State it plainly, because your clients will not hear it from the platform side: no published evidence yet shows agentic underwriting improving loss ratios on a named book. Not “the effect is small.” Not “the studies disagree.” Until an insurer publishes outcomes on its own portfolio, the three-to-five points carry no weight in a pricing conversation — whichever direction someone tries to spend them.
Integration is not evidence
The pitch’s quiet move is to treat integration as proof: the agent is wired in, so the signal it calls must be worth something. That confuses the plumbing with the water.
Integration lowers the cost of wiring a signal into a price. It does not move that signal up the proof ladder — from correlation (this score moves with losses) to attribution (the losses it predicts would not have happened anyway). A score inside the pricing workbench is still, at best, a correlation until someone has shown the second thing.
That is how risk signals have always reached a price, and agents have not repealed it. A signal enters as an underwriting variable, applied case by case within an underwriter’s authority. It becomes a filed rating factor only after being quantified against credible experience. The filed rate comes last. Better plumbing speeds the early stages — a signal an agent can call at triage is easier to try — but it does not skip them.
The Hub’s framework names five gates a signal has to clear. In broker language:
- Attribution — the losses the signal predicts would not have happened anyway. The master gate.
- Honest Baseline — the improvement is measured against a real comparison book, not the year the pilot happened to run.
- Persistence — the effect holds past the first renewal, not just the honeymoon.
- Exportability — the result travels from the pilot portfolio to the wider book, not just the chosen slice.
- Materiality — the effect is big enough to actually change a price, not just to decorate a slide.
A signal that has cleared none of these is a story. One that has cleared all five is a rating factor waiting for a filing. Everything a carrier says about its agent sits somewhere on that ladder — and “it’s integrated” is not a rung.
The amplification problem
Here is why this matters more with agents than with scores. A human underwriter applying an unvalidated score influences only the risks that underwriter touches. An agent applying the same score at submission triage, across a whole portfolio, shapes thousands of decisions before anyone checks whether it predicts loss at all.
That cuts both ways. A signal that genuinely predicts loss, at portfolio scale, is worth real money. One that merely correlates manufactures mispricing at industrial speed — risks declined that would have run clean, risks priced for reasons nobody can reconstruct, a book that looks fine until the claims develop.
The protection is a record. For every decision an agent touches, the carrier needs to keep — unit by unit — which external signals it relied on, so each signal’s contribution can be tested against the claims that came in. Without that record, an agent-priced portfolio cannot climb past correlation however good its results look. With it, the ladder becomes climbable. Ask a carrier about its agent, and this record is the thing to ask for. It either exists or the story ends there.
A word on the money behind all this. Gallagher Re put Q3 2026 global insurtech funding at $1.36 billion, down 44.6% quarter over quarter, with AI companies taking 83.2% across 72 deals. CB Insights counted $1.4 billion across 104 deals — the lowest deal count since 2016. Capital is concentrating hard on AI. But neither tracker separates loss-prevention technology from workflow software, so how much of that money backs things that actually reduce losses is unreadable. Say so when the slide cites it.
Five questions to ask before an agent prices your client’s book
Carry these into the renewal. Each is quotable as written, and maps to a gate or to the record.
- “Which outside signals did the agent call on this account — and what proof supports each one?” (Attribution.) Asking for the list ends the generic pitch; asking for the proof starts the real conversation.
- “Compared to what book?” (Honest Baseline.) An improvement figure without a named comparison book is a number about nothing.
- “Has the effect held past the first renewal — on accounts like mine?” (Persistence.) First-year results are the easiest an agent will ever produce.
- “Was that proven on a pilot slice, or on the wider book?” (Exportability.) Pilot portfolios are chosen. Your client’s account lives in the wider book.
- “If we dispute a price the agent helped set, can you show us — line by line — which signals drove it?” (The governance record, and Materiality together.) A reconstructable decision can be argued with. One that cannot means your client pays for a black box — worth saying out loud in the room.
None of these is hostile, or anti-technology. They are what any of us would ask of a human underwriter’s judgment call. The agent does not get a lower standard for being faster.
What would change the view
This is a proof-first position, not a permanent one, and it should say what would retire it. Two developments, either visible by the end of 2027, would do it: platform owners certifying third-party signals against published tests of loss lift, so integration itself carries evidential weight — or insurers granting documented pricing credit on integration alone, and those books then outperforming. Either would show the platform, not the evidence, had become the gatekeeper.
Until then the rule is simple. Integration is a fact about software. Evidence is a fact about losses. Renewals get priced on the second — and the three-to-five-point question is how you keep it that way.
The practice
This week, pick one upcoming renewal where the carrier has mentioned AI or its new platform in the run-up. Write the five questions into your file notes beforehand, and ask at least the first — which outside signals the pricing relied on, and what proof supports each. You are not picking a fight; you are establishing, on the record, whether the price came with evidence or with a story. Do it on one account and you will know how to do it on the rest of the book.