Guidewire

The Clearwater precedent at its clearest, and an independent analyst quoted in the launch coverage supplies the framing without meaning to: the value of this vendor's AI comes from deep integration with core business processes rather than from isolated experimentation. That is a description of models added to a system of record, not of a system of record built from models.

PolicyCenter, ClaimCenter and BillingCenter have been the books of record for property and casualty insurers for two decades and continue to issue policies, pay claims and bill premium with every model removed. The pricing capability that is genuinely model led arrived by acquisition in 2025.

Asserts that controls exist without describing them, which is the settled B. The published language is operational control, governance, security and regulatory compliance, plus secure access to operational data and workflow controls, none of it specified.

The agents genuinely act rather than advise, executing multi step workflows across policy, claims and billing, so the gap matters: nothing states what an agent may do without approval, what value or severity thresholds apply, who reviews an agent built by a customer before it reaches a policyholder, or how an erroneous action is reversed. Recorded as the same unresolved question raised by the no code agent builders on the banking roster, in a second industry.

Nothing published about how the models perform. The claims models make consequential predictions at first notice of loss, including severity, litigation likelihood and subrogation potential, and no accuracy figure, validation methodology, false positive rate, drift monitoring or versioning policy appears for any of them.

One phrase is worth flagging as a claim shape rather than evidence: pre built analytics are described as market validated, which asserts that the market has done the validating and names no test, no result and no validator.

A deep named customer library and a revealing gap inside it. Named institutions with described deployments: Zurich Insurance in Japan and the United Kingdom plus a global selection of the cyber analytics product, Nationwide, Basler, and Yuzzu of the AXA group. A named executive on record with a quantified outcome exists, a chief information officer reporting 90 percent straight through processing within six months at a farm bureau mutual.

Held at B rather than A because that outcome, and most of the quantified ones, belong to a partner integration operating inside the platform rather than to the vendor's own models. Worth noting as an observation about scale: a platform this large increasingly publishes case studies about what its partners achieve within it, which is evidence of ecosystem strength and weaker evidence of its own capability than it first appears.

Seventh instance this session of a pooling claim made as a product feature rather than a data commitment, and one of the most explicit. The claims benchmarking product compares a carrier's performance against peer and industry benchmarks, and the data platform ships curated industry datasets. Benchmarking a carrier against its peers necessarily means the vendor holds the peers' data and computes across it.

That is stated as a selling point and never as a term: nothing says what is contributed, whether contribution is optional, how it is aggregated or anonymised, or whether one carrier's claims experience informs models another carrier runs. On a platform holding the books of record for a large share of an industry, that is the highest concentration of pooled operational data anywhere in this index.

Nothing product specific located. The exposure is larger than for a pure software licence because the cloud platform means the vendor hosts policyholder and claimant records for a large share of an industry, and the analytics products curate that data into datasets for reporting and prediction. Nothing published states retention, vendor personnel access, or the boundary between a carrier's own data and the curated datasets built from it.

No security page, trust portal or enumerated certification located in the searches run. Recorded as a queued question rather than a finding of absence, and the expectation runs the other way: a listed company operating the system of record for 570 insurers will hold certifications and very likely publishes them somewhere not reached here. This is the cheapest available grade move on this vendor and should be the first thing checked on any revisit.

A software licensor and cloud host holding no insurance licence and claiming none. The exchange listing is an issuer obligation and earns nothing. Recorded as a deliberate contrast within this pocket rather than a neutral C: ZestyAI, built the same session and a fraction of the size, reaches an A because state regulators examined and approved its models for use in rate filings. The same regulators oversee the carriers running this platform, and nothing here has been through that door.

No fairness testing, protected characteristic handling or disparate impact analysis published, and the shipped capabilities make this the sharpest bias gap in the pocket. One claims model identifies claims with a high likelihood of litigation from claim data and adjuster notes.

That is a prediction about whether a claimant will assert their rights, used to route how their claim is handled, and it is the kind of model where correlation with income, geography, language and legal representation is the expected outcome rather than an edge case. Alongside it sit severity prediction and automated straight through settlement of low severity claims. Two other vendors in this same pocket disclose bias testing and this one does not.

Nothing published on responsibility for a wrong outcome, and the agent framework sharpens a question the analytics products already raised. When an agent executes a multi step workflow across policy, claims and billing and gets it wrong, responsibility divides between the customer who configured the agent, the vendor who wrote the framework and supplied the pre built agents, the model the customer selected for that task, and the carrier that acted on the result.

That is the fourth published instance of the agent marketplace liability gap and the first in insurance. For the claimant on the other end of a mis triaged claim, nothing addresses notification, correction or appeal at all.

Third instance of the route to B that hands the choice to the customer, after Fioneer and the reasoning applied to the banking platform tier. The company states that insurers can select different models depending on the business task while retaining secure access to operational data and workflow controls, which tells a buyer three useful things: the model layer is swappable, no undisclosed provider is locked in beneath the product, and the dependency can be placed with a supplier the carrier has already assessed.

Held at B: no default model named for the company built claims and underwriting agents, no model inventory, and nothing on which capabilities work with which models or how behaviour varies between them. The underlying cloud is named and earns nothing on its own under the standing rule.

This vendor does not integrate with the core system, it is the core system for a large share of the market, which is the same structural position that earned the banking platform tier its grades. Beyond that, the surrounding architecture is published as products rather than claims: a cloud platform, a data platform with curated datasets and built in integration to the analytics tools, a partner marketplace of validated integrations, and application programming interfaces that partners build against.

The agent framework is described as executing multi step workflows by interacting with policy administration, claims and billing systems on the platform, which is integration depth stated as a capability rather than asserted.

Delivery is documented as products rather than described as options. Guidewire Cloud Platform is a named, separately marketed offering with its underlying public cloud named, a separate hosted suite exists for smaller carriers, and self managed installation remains for carriers that have not migrated.

Independent implementation material describes cloud migration as a distinct programme with quality gates, which corroborates that these are genuinely different delivery modes rather than one product with labels. Held at B on the standard reason: no residency statement, no regional hosting map and no sovereignty commitment across 43 countries.

No pricing published for any product, and no commercial model stated for the analytics line or the agent framework. Notable only because the buyer population includes small regional carriers served by a separate hosted product, where a published entry point would be commercially normal and is absent.

More than 570 insurers across 43 countries, spanning global multi line carriers, national insurers and, through a separate hosted product, smaller regional carriers. Named users run from one of the largest multi line groups in the world down to a state farm bureau mutual. Coverage within property and casualty is complete rather than partial: policy administration, claims, billing, underwriting, pricing, analytics and catastrophe response are all addressed. Deliberately confined to one industry and graded on reach within it.