Verisk

The removal test is unambiguous here despite the volume of artificial intelligence activity. Strip the models out and what remains is one of the most durable franchises in insurance: a licensed advisory organisation with 34.5 billion statistical records, standardised coverage forms maintained against thousands of legislative and judicial changes a year, projected loss costs filed with state regulators, and an estimating platform that is the working tool of the property claims industry.

None of that depends on a model. The company's own framing is accurate and does not lift the grade, since it describes artificial intelligence as infrastructure placed inside workflows teams already run, and states that value arrives without changing how those teams work. That is embedding into an established franchise rather than dependence on learned components. Actuarial and statistical modelling has been central for decades, and the axis measures dependence on the current generation of models rather than analytical sophistication in general.

The oversight position is stated plainly at product level and the language of the products themselves supports it. The claims capability is described as designed to assist rather than replace claims professionals, framed as removing low value tasks so that people concentrate on judgement and craft, and the outputs are named as recommendations rather than decisions, with estimating recommendations, line item suggestions, generated descriptions and summaries all landing in front of an adjuster who is doing the estimate.

In underwriting the described role is first pass risk assessment and data standardisation ahead of a person rather than in place of one. Two things keep it below the top band. The assist not replace statement appears in product marketing rather than in a governance commitment or contractual term, so it describes intent rather than an enforced control.

And nothing published names autonomy modes, referral triggers or a threshold above which a recommendation cannot be accepted without review, which is what a carrier would need to configure the boundary deliberately.

Model discipline is evidenced through several channels rather than claimed in one. The actuarial franchise carries its own rigour, with projections of future claim costs and loss adjustment expenses built by staff actuaries and data scientists on a database of billions of records, and those projections are filed with state regulators, which subjects the modelling to external review no commercial vendor process would supply.

Catastrophe outputs are described as defensible and regulator ready, and the underlying data as normalised, validated and governed at scale. Validation and documentation practices on the claims document product run through an independent governance vendor.

What is not published is the model level detail a carrier's own model risk function needs to complete its inventory: no accuracy or error rates for photo labelling, document extraction or estimating recommendations, no model cards, no drift monitoring description, and no revalidation cadence for the generative capabilities embedded across the product estate.

Evidence here is audited, current and quantified, which separates it from everything else in this lane. The company files with the United States securities regulator, so revenue, segment performance and business mix are examined rather than asserted: second quarter 2026 revenue of 806.3 million dollars, up 4.3 percent and 5.8 percent on an organic constant currency basis, subscriptions at 83 percent of revenue, and full year guidance of 3.19 to 3.24 billion dollars.

Artificial intelligence adoption is reported as a specific number rather than a claim, with licensees of the claims capability rising nearly tenfold between March 2026 and mid 2026 to approximately 7,000, and module delivery is tracked against a public target, seven shipped in the second quarter against 25 planned for the year. Product level outcomes are quantified, including a claimed reduction of up to 98 percent in research time on premium audit documents. The reservation worth recording is that the adoption figure counts licensees rather than measuring outcomes, and the efficiency percentages remain vendor stated without a published measurement basis.

Four separate published items support this grade and no other vendor in this lane offers more than two. A Commitment to Ethical and Responsible AI disclosure is published, described as setting out human centred values, ethical principles, an artificial intelligence governance structure, standards and review processes, and stated to sit within the same corporate governance machinery used for strategy and risk management rather than beside it.

A dedicated data stewardship function is published under corporate responsibility. The protocol connectors carry an explicit and checkable data commitment, that retrieved data stays inside the client session and outside model training, which is the precise assurance a carrier needs before letting an external model reach its book.

And the industry data pooling that underpins the analytics happens under licensed advisory organisation status subject to periodic state examination, which answers the contributory data question that sits unanswered at several data advantaged peers. The reservation recorded is that the ethical artificial intelligence disclosure was verified as published and enumerated rather than read in full, and no retention schedule or subprocessor list was located.

The pooling structure that raises this question is answered better here than at any comparable vendor, and answered by statute rather than by contract. Carriers submit loss and policy experience, the company aggregates it and returns projected loss costs, and it does so as a licensed advisory organisation and statistical agent whose activities are subject to periodic examination by state regulators, which places the pooling under supervision rather than leaving it to private terms.

A dedicated data stewardship function is published under corporate responsibility with stated principles for responsible data gathering and analytics. One artificial intelligence specific commitment is precise and checkable: data retrieved through the protocol connectors is stated to remain within the client session and outside model training. Absent are the operational documents a buyer's counsel would still want, with no data processing agreement, retention schedule or subprocessor list located in two passes.

Two passes across the company's site, its newsroom, its artificial intelligence and stewardship pages and third party coverage located no trust centre, no service organisation control report, no information security certification and no dedicated security page, which is a thin published position for a listed company holding industry wide claims and policy data.

The grade rests on the inference standard applied elsewhere in this index rather than on published evidence, and the supporting facts are stronger here than in most applications of it: the company is subject to periodic examination by state insurance regulators in its licensed advisory capacity, files with the United States securities regulator under disclosure obligations that now extend to cybersecurity risk management, and supplies most large United States property and casualty carriers, each of which conducts vendor security review with the leverage to demand evidence. The controls therefore almost certainly exist and are not published, and a buyer can assess nothing before entering a confidentiality agreement.

This is the only vendor in this index holding statutory status inside the industry it serves, and the difference is one of kind rather than degree. It operates as a licensed advisory organisation and statistical agent, subject to state rather than federal regulation by operation of the McCarran Ferguson Act, with many products and operations subject to review or approval by state regulators, licensed advisory activities subject to periodic state examination, and operations additionally exposed to state antitrust and trade practice statutes enforced by attorneys general and insurance regulators.

It submits forms and loss costs to state regulators on behalf of insurers, meaning the regulatory interface is the product rather than a constraint on it. Its coverage language is maintained against roughly 10,000 bills, 8,000 regulatory actions and 2,000 court decisions annually.

Most tellingly for this index, the company authors the standard endorsements through which the industry excludes generative artificial intelligence liability from commercial general liability cover, filed by carriers with their own regulators during 2026.

Governance is documented to a standard the rest of this lane does not approach, and the fairness limb is still evidenced by framework alignment rather than by results. What is published: an ethical artificial intelligence commitment naming fairness and transparency explicitly, a governance structure, standards and review processes, and, for the claims document product specifically, governance aligned to the state commissioners association model bulletin and the federal artificial intelligence risk management framework through an independent governance vendor's control library.

That last item is the strongest single element, because it introduces a third party's controls rather than relying on self assessment, and it is the only instance of independent governance tooling found on a named product anywhere in this lane. Three reservations hold the grade at this band. The independent alignment covers one named product rather than the estate. No bias testing results, fairness metrics or disparate impact analysis are published for any model. And alignment to a framework establishes process rather than demonstrating an outcome, which matters because loss costs feed carrier rates that reach consumers.

The sharpest observation in this record sits on this axis. This company authors the standard endorsements through which the property and casualty industry excludes generative artificial intelligence liability from commercial general liability cover, language carriers began filing with their own regulators during 2026, and it publishes nothing about liability or recourse for the artificial intelligence it sells. The party writing the market's exclusion clauses has not written its own.

What holds the grade above the floor is a structural transparency route no peer offers: loss costs and forms are filed with state regulators and those filings are public documents, so the pricing base that reaches a consumer's premium can be obtained and examined by that consumer's counsel or their regulator without the company's cooperation. That route covers the filed products only.

It does not reach the claims capability, where a policyholder whose settlement was shaped by a generated estimate, an automated photo description or a pricing recommendation has no notice and no published route to contest.

The external dependency is named and its data properties are stated, which is more than most of this lane manages. The protocol connectors released in May 2026 were built with Anthropic and bring the company's datasets into that provider's workflows, with an enterprise integration giving natural language access to loss cost trends and estimating intelligence, and the governance property published alongside it, that retrieved data remains inside the client session and outside model training.

The company also states explicitly that a customer's artificial intelligence strategy should not depend on a single vendor's roadmap, which is a deliberate position on concentration risk rather than silence about it. Its own analytical models are built on proprietary data it owns and files, so that part of the chain is self evident.

What remains unpublished is the model layer beneath its embedded generative features, with no provider, family or version named for the photo labelling, summarisation, document extraction or plain language estimating capabilities, and no notification commitment when a model changes.

This vendor integrates into cores rather than being one, and within its own estate the integration is deep and evidenced. The claims capability spans the estimating platform, the assignment and analysis product, the contents product, the estimate review product and the customer engagement product as a single connected flow, and the company states data moves across them without disrupting existing processes.

Generative capability is live inside four named products rather than offered as a separate console, which is integration into the working tool rather than beside it. Outward reach is real in two directions: loss costs, forms and rules flow into carrier rating and policy issuance systems across the market, and the protocol connectors released in May 2026 give external agents a governed route to underwriting and claims data. What holds the grade at this band is that policy administration, billing and the system of record belong to someone else, so a carrier's core estate is a dependency here rather than a deliverable.

Delivery is cloud based and stated in general terms, with the claims capability described as running securely in the cloud and embedded in platforms customers already use, requiring no special hardware or software and available to most users within days of approval. Beyond that the published record thins considerably for a company of this size.

Two passes located no region list, no processing location statement, no data residency commitment a carrier could contract for, no tenancy description and no infrastructure provider named for the core estate. Operations spanning more than 20 countries make those omissions more consequential rather than less, since a European or Asia Pacific carrier has nothing published against which to check its own residency obligations. The one deployment fact with real weight is ease of adoption rather than control, which serves the buyer's convenience more than their compliance function.

No price list is published, and considerably more commercial information reaches the market here than anywhere else in this lane because the company reports to a securities regulator. Revenue mix, subscription proportion, segment performance and margin guidance are all disclosed and audited, and pricing strategy is described in earnings commentary, with price increases taken at contract renewal alongside data and artificial intelligence enhancements, which tells a buyer plainly that renewal is a price event.

One specific commercial term is published at product level and it is the kind peers never state: two artificial intelligence estimating capabilities are included with an existing professional licence at no additional purchase. What remains absent is the buyer facing detail, with no rate card, no unit of billing, no contract term and no indication of how the newer artificial intelligence capabilities are priced where they are not bundled.

Coverage approaches the structural limit of what this axis can measure. On institution type the reach spans property and casualty carriers of every size, reinsurers, brokers, managing general agents, restoration contractors working inside the estimating ecosystem, and capital markets participants through catastrophe risk assessment used in insurance linked securities.

On line of business it covers commercial and personal lines across the underwriting estate, with 8.2 billion commercial and 21.5 billion personal lines records behind the loss cost work, and claims coverage running from property estimating through casualty and anti fraud. Geographic reach is stated as teams in more than 20 countries with catastrophe modelling addressing global perils.

The defining coverage fact is structural rather than commercial: standardised forms and filed loss costs are the base layer from which most United States property and casualty carriers construct their own rates, so the market position is closer to industry infrastructure than to a vendor with a customer list.