Swallow

Two shipped lines where models carry the throughput: Herbie, described as the AI actuary inside the platform, builds a tested and deployable pricing model from a prompt and generates up to a hundred test cases from an instruction; and the filings line makes twenty five years of United States rate filings executable, extracting rating logic from documents rather than summarising them. Held at B rather than A because the residue test is unusually clean here.

The company sold the platform for roughly two years before the AI line existed, so a deterministic pricing engine, visual builder, test harness, approval workflow and rating interface all remain if the models are removed. That makes Swallow the pocket's clearest natural experiment on centrality, since the same product demonstrably worked without any of it.

This vendor holds the pocket's best specified human gate and its least supervised autonomous action, in the same product. The filing path meets the A bar on its own terms: a credentialed actuary, the customer's own or a partner, reviews, certifies and manages regulator objection responses, with the split quantified at eighty to ninety percent of artefacts generated automatically and professional judgement retained for the rest.

That names and staffs an adjudication layer, places it in the path, and ties the reviewer to an external professional credential, which is stronger than the idenfy precedent. Against it: the distribution path contains the most autonomous consumer facing action in this pocket, a voice agent that runs renewal campaigns at scale, generates updated quotes in real time and is described as negotiating live, with no threshold, review structure or escalation route published.

Deployment gates are real though (test suites that run automatically on any change, approvals, one click rollback with an audit record of who changed what and why). The general lesson is that oversight is a property of a path rather than of a vendor, and this record should be read that way.

Strong on pricing model governance, silent on the AI layer, and the split is the point. Published: a versioned interface created on every publish with previous versions staying live and callable, one click rollback with an audit record of who changed what and why, scenario and pivot testing, impact testing before production, what if analysis and back testing on historical data, and controlled experiments routing a share of quotes between model versions.

Better specified than the WTW versioning disclosure that earned the same grade. Held off A on the standing extraction question, which is central here rather than incidental: the filings product reads rating logic out of documents and a wrongly extracted factor table propagates silently into every quote.

Reconciled example calculations are offered against a filing's own worked examples, which is the closest thing to an accuracy control in this pocket, but no tolerance, pass rate, failure handling or human review rate is published, and nothing describes the models behind Herbie.

Five named customers (Open, Rivr, MGT Insurance, Ouch, ISC) with two full case studies carrying outcomes: Rivr launched complex life products in weeks rather than months, quoted by founder and chief executive Will Clarkson Webb, and Open brought a premium United Kingdom home insurance product to market in under a week. Named customer plus quantified outcome plus a named executive on the record is the A bar as written, and it is met.

Recorded so it is not relitigated: the scale claims are stronger than the attributed ones but sit apart from them, since the customer serving three million quotes a month and the client above one billion dollars of revenue are not identified. That is the familiar shape where the two halves never join, and it caps how much the scale claims can carry on their own.

Nothing published on whether customer quote data, pricing models or uploaded workbooks train anything. The claim shape is the unfalsifiable absolute and this is the sharpest instance yet catalogued, because the acknowledgement and the absolute sit on the same page: one answer states that AI can hallucinate quote questions, quote prices and the rules around quotes, and another states that the vendor's own channels handle new quotes, amendments and recall via AI without hallucination.

A vendor naming the risk and then exempting itself from it in the same document is a better example than a bare zero hallucination claim. Positive counterweight worth keeping: the market analytics product achieves peer comparison from filings that are public by law, so it delivers what pooled benchmarking delivers elsewhere without holding any competitor's confidential data. That is the first clean counterexample in the pooled corpus ladder, and the mechanism is a public transparency regime making confidential pooling unnecessary.

States that the platform is built with regimes such as the European data protection regulation in mind and that data can be exported, deleted or audited at any time, plus registration with the United Kingdom data protection regulator and a published privacy policy.

Held at C because built with a regime in mind is the gesture this index credits nowhere, the export and delete capability is a product feature rather than a commitment about the vendor's own processing, and there is no position at all on the United States financial privacy regime despite a product line built specifically for United States carriers. No data processing agreement, subprocessor list or statement about what the vendor does with quote data.

Certified to the international information security management standard, stated plainly in a support answer, alongside encryption in transit and at rest, role based permissions, audit logs and complete version history. Registered with the United Kingdom data protection regulator. The presentation is the catalogued failure though: the footer badge reads only certified with no numbered standard, which is the imprecision shape, and the numbered claim appears solely in a buried support answer.

This is the identomat pattern repeating exactly, where a real certification held by the company is absent from the badge row that exists to advertise it. Held off A: no trust portal, no certificate link, no scope or audit period stated, and no service organisation control report of any kind for a vendor whose output feeds premium calculations.

Unlicensed and unsupervised. The interesting shape is how close it operates to the machinery without being inside it: the product indexes the filing portal every carrier uses, generates the actuarial memorandum, rate manual, exhibits and rule pages, answers regulator objections, and encodes state specific rating restrictions, yet the party answerable to the supervisor remains the carrier and the certifying actuary.

Compliance guidance across three regions is a product feature, not standing. Direct contrast within this pocket: a company a fraction of the size of the incumbents holds a regulatory A because state supervisors examined and approved its models, while this vendor writes filings for a living and has never been through that door itself.

No fairness testing, no protected class position, no governance disclosure of any kind. This is the third vendor in the pocket selling price optimisation with silence beside it, after Earnix and the WTW optimiser, and it is the sharpest of the three on mechanism rather than on scale. The platform routes live customers into controlled pricing experiments and scores the versions on sales, and the voice agent negotiates renewal prices in real time with individual consumers.

Optimising against propensity to accept rather than against risk is exactly the practice consumer fairness rules address, and here it is not an inference from a product name but a described workflow with the success criterion stated.

No recourse route, no allocation of responsibility, nothing published beyond standard terms. The chain here is longer than most and the question follows it: an approved filing is machine extracted into an executable rater, the rater is published as a live interface, a voice agent quotes and negotiates against it, and a consumer is priced.

The certifying actuary signs the filing, not the extraction, and the carrier owns the customer relationship, so a pricing error introduced between the filed document and the live quote has no named owner.

One supplier is named, and it is the peripheral one: a direct integration with a speech synthesis provider for the voice channel. Nothing names the model doing actuarial work, reading rate filings or generating pricing models, which is where the buyer's exposure sits. Same shape as the frontier agents formulation elsewhere in this pocket, where a quality adjective substitutes for a name. Held at C rather than B on that basis: knowing the voice vendor tells a buyer nothing about the system reading its rate tables, and no model inventory, default or version appears anywhere.

Deliberately architected as a rating layer that operates independently of policy administration, sold as a way to move pricing out of a legacy core while leaving the core stable. Secure representational state transfer interfaces with a versioned endpoint per publish, webhooks, batch endpoints, software development kits, embeddable forms, chat and voice channels, and exports to Metabase, Redshift, Athena, Snowflake, Google Sheets and Excel. Listed on the Amazon Web Services marketplace.

Published performance envelope: one customer serving more than three million quotes a month, price comparison website volumes at low latency and high concurrency. Third instance of a published performance envelope in this run after Murex and the WTW rating engine. Held off A because no named core system connector exists, unlike the pocket incumbents which integrate with named policy suites.

Single tenant software as a service on Amazon Web Services with a public status page and a marketplace listing, custom service level agreements offered to enterprise customers. Nothing published on data residency, region choice, private cloud or customer hosted options, which matters for a vendor selling simultaneously into United Kingdom, European Union and United States insurers whose supervisors take different positions on where policy and quote data may sit.

The most transparent commercial disclosure in this pocket and one of the better ones in the index. A published entry price of 2,500 pounds a month, a published scaling basis (quotes processed and gross written premium), a published included volume of 100,000 quotes per plan, and a published segment discount for insurers under 10 million pounds of premium.

The filings line publishes its own methodology: priced per state and per product rater with a lower rate on each additional state, per filing for filing assistance with a lighter fee for each additional state and reduced fees for revisions, and annual subscriptions tiered by coverage scope for monitoring and market analytics.

The pocket inversion is worth writing up: the smallest and youngest vendor here is the only one that will tell a buyer what it costs, while the three largest publish nothing. Only softness is the up to fifty percent cheaper comparison, which names no baseline.

Genuinely broad for a company this size. Institution types named: carriers, reinsurers, managing general agents, program managers, captives, risk retention groups, brokers and embedded distribution partners. Geographies: United Kingdom, European Union and United States, with state level awareness for filings.

Lines enumerated at length across personal (private passenger auto, homeowners, dwelling fire, renters, condo, mobile home, umbrella, pet, motorcycle, recreational vehicle, boat, powersports, warranty), commercial (workers compensation, commercial auto, business owners policy, commercial property, general liability, professional liability, directors and officers, cyber, inland marine, equipment breakdown, surety, management liability) and specialty (medical malpractice, trucking, ocean marine, aviation, equine and agricultural, energy and environmental). The platform also states applicability to credit, mortgage, financing, leasing and membership pricing, though no customer outside insurance is named.