Socotra

The company markets itself as built for an artificial intelligence driven world, and the removal test still lands it with the rest of the core platform tier. Take the models out and a policy administration system, a billing engine, a rating engine and document generation all continue to run, because that is what carriers buy and what they migrate onto.

The learned components are real and recent rather than decorative: an underwriting assistant generally available since March 2026 that imports documents, produces risk assessment insights and writes summaries against live policy data, and an agentic configuration capability that turns natural language into deployed product configuration. Both change how the platform is used rather than what it fundamentally is. The architectural claim deserves partial credit and does not lift the grade, because being designed for models to operate inside is a readiness property rather than a dependence on them.

The oversight position is published as an absolute rather than a preference, and it sits at the top of this band: the assistant makes changes or takes actions only with explicit human approval. Stating the constraint at action level is more precise than the human in the loop language used across this lane, which usually describes an intention without naming the unit of control, and it is paired with an audit log that records every operation, so approval is evidenced rather than asserted.

The claim covers a real capability rather than a trivial one, since the assistant operates on live policy and billing data inside the underwriting workbench, and the agentic configuration capability deploys product configuration from natural language. What separates this from the top band is scope rather than sincerity.

Nothing published describes what the assistant may draft, stage or prepare without approval, whether approval can be granted in bulk across many items, or what a carrier is shown at the moment of approving, and those are exactly the details that determine whether an approval step is a control or a formality.

Transparent reasoning is claimed as a property of the underwriting assistant, and the platform level auditability behind it is real, with every configuration change and data operation logged and exposed to the customer. That gives a carrier a record of what happened, which is a different thing from an account of why the model produced what it did, and the second is what a model risk function needs.

Two passes located no validation methodology, no accuracy or error rate for document import or risk assessment output, no monitoring or drift detection description, no revalidation cadence and no model documentation a carrier could file.

The insurer specific learning design adds a question nobody has answered publicly, since a capability that adapts to each carrier's criteria produces a different fitted behaviour per customer, and nothing describes how that per customer variant is tested before it operates on live applications or how a carrier would detect it degrading.

This is the weakest side of an otherwise well disclosed record, and the contrast with its own architecture claims is the useful finding. Deployment evidence is stated in aggregate rather than named: more than 70 product launches and 15 migrations for insurers globally across property and casualty and life.

Those are meaningful numbers for a challenger and an order of magnitude below the incumbent cores in this lane, and the figures are undated, so the rate at which they are accumulating cannot be established. One customer is documented at implementation level, a property and casualty insurer that self implemented a product in five months, which supports the openness claim specifically.

The quantified artificial intelligence outcomes carry no method at all: halved configuration timelines, development costs reduced by 75 percent and prototype iteration cycles shortened by 90 percent are published without a baseline, a customer attestation or a measurement period. Independent peer review carried no published customer reviews at the time of research, so third party corroboration is thin in both directions.

One published commitment does most of the work here and it is stated plainly enough to hold the company to: the model does not learn from an insurer's secure data, and what it learns instead is that insurer's own risk assessment criteria and product workflows.

That distinction matters commercially as well as ethically, because it forecloses the arrangement where a vendor improves a shared model on one carrier's book and sells the result to a competitor, which is the live question at several data advantaged peers.

Supporting controls are architectural: completely independent production environments per customer, customer specific encryption keys, encryption at rest and in transit, and an audit log covering every operation on customer data exposed through an event stream. What is missing keeps it out of the top band.

The commitment appears as a design principle in release material rather than in published terms, no subprocessor list exists, and nothing states where inference happens or whether prompts and document contents leave the platform boundary during processing.

Privacy is documented at the level a buyer's counsel can actually work from, which is uncommon in this lane. Published positions include compliance with European data protection regulation and with United States health information rules, a data processing agreement setting out technical and operating procedures for handling customer data, and data ownership defined explicitly in the terms and conditions rather than left to inference.

Isolation is architectural rather than contractual, with each production customer in a completely independent environment holding its own configuration, its own data and its own generated private keys. Auditability is the strongest element: every configuration change and every operation involving customer data is recorded, and the log is exposed to the customer through an event stream rather than held internally, so a carrier can reconstruct who did what without asking. Absent are a retention schedule, a subprocessor list and any region by region statement of where data is held.

Security is documented on a dedicated published page rather than promised on request, which separates this record from most of its lane. Certification to the international information security management standard is held and its scope is published explicitly, covering business processes, locations, technology, people and information, and scope is the detail that makes such a certificate meaningful rather than decorative.

Practices are described concretely: encryption at rest and in transit, 256 bit encryption with strong private keys generated per customer, automated secrets management, network monitoring for unusual behaviour, and regular penetration testing by an internal information security team to identify and remediate vulnerabilities. Compliance positions are stated for European data protection and United States health information rules.

One qualification a buyer should carry into diligence rather than skip: the service organisation control report cited is of the first type, which addresses controls relevant to financial reporting, so the report specifically examining security, availability and confidentiality is not among the credentials published. There is also no self serve portal from which the underlying documents can be obtained.

The company holds no insurance licence and does not claim one, which is the ordinary and correct position for a core platform supplier. Its regulatory language is about operating environment rather than authorisation, describing the platform as built to support production artificial intelligence at scale in regulated environments, with security, auditability and traceability offered as the evidence for that.

Nothing published addresses the filing machinery that surrounds a policy administration system in practice, with no rate, rule and form maintenance capability described, no statement on state by state filing support in the United States market, and no positioning against the European artificial intelligence regulation despite the platform being sold into life insurance, where risk assessment and pricing are named in that regulation's high risk schedule. The audit log is the one published feature a regulator would actually find useful.

Governance content exists and is entirely procedural, covering approval, auditability and the boundary around training data, with nothing on fairness. Two passes located no bias testing description, no fairness metrics, no disparate impact analysis, no model card and no responsible artificial intelligence statement of any kind.

The gap is sharper here than the grade alone conveys, because the flagship capability is underwriting: an assistant producing risk assessment insights on live applications is operating at exactly the point where a proxy variable produces a discriminatory outcome, and the company sells into life insurance where that use is treated as high risk under European law.

The published commitment that the model learns each insurer's own risk assessment criteria cuts both ways on this axis, since it means the model inherits whatever bias already sits in that carrier's criteria, and nothing published describes testing for it.

Accountability is structurally clearer here than at most of this lane, and formal recourse is still unpublished. Because the assistant acts only with explicit human approval and every operation is logged and exposed through an event stream, any outcome reaching an applicant can be traced to a named person who approved it at a recorded moment.

That is a materially better position than a system whose decisions cannot be attributed, and it locates responsibility with the carrier rather than diffusing it. What remains absent is everything contractual and everything applicant facing. Nothing published allocates liability between vendor and carrier when the assistant summarises a submission wrongly, imports a document incorrectly or produces a risk assessment insight that proves unfounded.

The applicant declined or priced on an assisted assessment has no relationship with this vendor, no notice that a model informed the underwriter's view, and no published route to obtain or contest what the model contributed, which is the same third party exposure carried across this lane.

Disclosure is asymmetric in an instructive way. The agentic integration surface is documented openly, with a published Model Context Protocol server, setup instructions and named commercial artificial intelligence tools it connects to, so a carrier can see exactly which external agents can reach core data and through what interface. The model behind the vendor's own underwriting assistant is not named anywhere.

No provider, model family, version, hosting arrangement or fine tuning approach is published, and the statement that the model does not learn from an insurer's secure data implies a base model supplied from elsewhere without identifying it. That leaves a carrier unable to assess concentration risk, to know whether a provider change would alter behaviour mid contract, or to answer its own regulator on where inference occurs. Documenting the third party agents thoroughly while leaving the embedded model unnamed is the specific question to raise.

Integration openness is the product's central design claim and it is evidenced concretely: publicly documented open application programming interfaces described as fully open, modules that can be added, removed or swapped, a marketplace through which some connect as applications, a stated ability to integrate with distribution platforms, data providers and third party systems, and construction in common programming languages so that a carrier's existing engineers can extend it without proprietary tooling or specialist training.

One customer self implemented a product in five months, which tests that claim rather than restating it. A published Model Context Protocol server extends the same openness to external agents, giving a documented and auditable route for agentic tools to operate against core data, which very few cores in this lane currently offer. What holds the grade below the top band is estate breadth rather than architecture.

The core covers policy, billing, rating, document generation and workbenches, and a claims system is not part of what is published, so a carrier consolidating the full lifecycle onto one vendor cannot do it here.

Deployment is cloud only and single tenanted in the way that matters, with each production customer given a completely independent environment holding its own configuration, its own data and its own generated private keys, which is a stronger isolation position than the shared multi tenancy usual at this price point.

Every customer runs the same version and is upgraded automatically with no downtime, and the company states that keeping all customers current is what makes new capabilities available to everyone at once, which is an operational commitment with a real cost attached. Geographic support is claimed broadly, with the platform and the underwriting assistant both stated to support any geography, and European data protection compliance backed by a data processing agreement.

Left unpublished are the region list, the choice of processing location, any residency commitment a carrier could contract for, and any on premises or private cloud path for a buyer whose own policy forbids public cloud.

Pricing is absent from vendor surfaces, in line with the rest of this lane, and the commercial claim it does make is comparative and unquantified, positioning the platform as delivering the lowest total cost of ownership in its market.

Independent analyst directory content describes the model as subscription based with cost determined by usage and chosen modules, and with tiers available, which is more shape than most peers offer and is third party description rather than vendor publication.

Two structural facts do bear on cost and are published: every customer runs a single version with automatic upgrades, which removes the version upgrade project that dominates ownership cost elsewhere in this lane, and the module design allows components to be added, removed or swapped. Neither is priced. A buyer can therefore reason about the shape of the bill without seeing a number.

Breadth of product model is genuine and the customer base behind it is modest, which is the tension this grade reflects. The platform is stated to support any line of business, any geography and any distribution model, with a product agnostic data model rather than line specific templates, and deployment evidence spans both property and casualty and life markets, which is a wider segment claim than most cores in this lane make.

Buyer types run from insurers replacing a legacy core to those augmenting one, and the underwriting assistant is stated to support all insurance products and geographies rather than launching in one line first. What holds it below the top band is volume rather than reach: more than 70 product launches and 15 migrations is a fraction of the installed bases at the incumbent end of this lane, so the segments are addressed by design and evidenced thinly in several of them.