Snapsheet

The removal test places this with the claims platforms rather than with the vision specialists, despite a superficially similar origin. Strip the models out and a claims management system remains, covering first notice of loss through settlement with digital communications, workload management, configurable workflows, reporting and payment disbursement, and carriers buy it to run claims operations rather than to score them.

The company's own history reinforces the reading: the virtual appraisal franchise that made its name was originally staffed by remote human appraisers examining submitted photographs, so the business model was remote assessment before it was automated assessment, and the automation improved a workflow that already worked.

The current artificial intelligence layer is genuine and configurable, with agents running across summarisation, scoring, intake, liability checks, communications and claim evaluation, and it sits above a platform that predates it by more than a decade.

Both execution modes are published plainly, which is more candour than most of this lane offers, and the boundary between them is left entirely to the buyer. Agents can be triggered automatically inside event based workflows, meaning they act when the workflow fires rather than when a person authorises it, or initiated on demand by an adjuster in a side panel copilot.

The pre configured agent set makes the stakes concrete: liability checks determine who is at fault, claim evaluation and scoring shape what a claim is worth, and communications reach the claimant directly. The configuration model compounds this, since a carrier directs the actions an agent takes by writing its instructions in plain language, so the vendor supplies capability and the customer supplies the guardrails.

Nothing published names an action class requiring approval, describes a confidence threshold triggering referral, sets a default limit, or explains what a carrier is shown before an automated agent action takes effect.

The configurability that makes the product attractive also makes it the hardest thing in this lane to validate, and nothing published addresses that. Two passes located no accuracy or error rate for any agent, no validation methodology, no benchmark against adjuster performance, no drift monitoring, no revalidation cadence and no model documentation.

The structural problem sits above those omissions: a carrier configuring an agent at the prompt level in plain language is altering model behaviour, and a prompt change is an unvalidated change to a system making liability and evaluation determinations.

Nothing published describes version control over prompts, testing before an agent goes live, an approval workflow for configuration changes, or a record of which agent version produced a given output, which is what a model risk function would need before allowing any of it into production.

Volume evidence is specific and substantial, and its currency is the weakness. Published figures include more than 4.3 million claims managed and 15.3 billion dollars of indemnity processed, alongside more than 25 million automated tasks executed and a partner network exceeding 140 covering carriers, third party administrators, managing general agents and insurtechs.

Indemnity processed is a more meaningful measure than most vendors offer, because it sizes the money moving through the platform rather than the logins against it. Three reservations apply. The headline claim and indemnity figures attach to a 2023 report and have not been restated since. Named customers are thin for this scale, with IAT Insurance Group the clearest example located.

And no outcome is published for the artificial intelligence specifically, with no accuracy rate, no automation rate, no cycle time reduction and no loss adjustment expense saving attributed to any agent.

Two passes located no statement on whether carrier or claimant data is used to train or improve models, whether one customer's claims history informs capability delivered to another, what happens to data on termination, or what protections apply to claimant photographs and adjuster notes when an agent processes them.

A claim circulating in third party content that the models were trained on millions of past claims could not be verified against anything the company publishes, and a buyer should treat it as unconfirmed rather than as disclosure.

The configurable design creates a stewardship question most peers do not face: because a carrier designates which context and data an agent may use when writing its prompt, the boundary around sensitive claim data is set by whoever configures the agent, and nothing published describes what data classes are available for designation, whether any are restricted by default, or what prevents an agent being pointed at material it should not see.

The starting position is better than most of this lane and the verifiable content is not. A privacy policy and a Trust Center are both published and linked site wide, which means a buyer has a named destination rather than an email address, and the existence of a maintained Trust Center indicates a compliance function that assembles documents rather than improvising them per deal.

Two passes could not retrieve the contents of either, so no retention schedule, data processing agreement, subprocessor list or claimant data commitment could be verified. The sensitivity here is high in a specific way: the platform holds photographic evidence of loss, adjuster notes, liability assessments and payment instructions including disbursement details, which combines personal, evidentiary and financial data about claimants who are not the customer.

A Trust Center is published and linked from every page of the company's site, which is the correct structure and a better starting position than most vendors in this lane offer, since it gives a buyer's security team a named destination rather than a request queue. What could not be established is what sits inside it.

Two passes did not surface a named service organisation control report, information security certification, penetration testing statement or encryption description from the Trust Center or anywhere else, so no credential can be recorded here.

The grade reflects an unverified but properly structured disclosure route rather than an absence of one, and the practical instruction for a buyer is to open the Trust Center first and establish which reports exist, of which type and covering which period, because the platform holds claimant data and moves settlement money.

The company holds no insurance licence and does not claim one, operating as a technology supplier to the parties that do. Third party analyst description credits the platform with helping carriers maintain compliance with regulatory standards while handling claims, which is a capability claim rather than a status.

The regulatory surface it touches is real even so, since claims handling is examined at state level through market conduct review, timeliness and fair settlement requirements bind the carriers and administrators using it, and the company's own history included employing appraisers, an activity licensed in several states.

Two passes located no statement on how the platform supports state claims handling requirements, no market conduct or audit trail capability described, and no positioning against artificial intelligence guidance from the state commissioners association despite agents performing liability assessment.

Two passes located no responsible artificial intelligence statement, no governance framework, no fairness testing, no bias evaluation and no model card. The exposure is more direct here than at most vendors in this lane because of one item in the published agent list: liability checks.

An agent assessing fault is making a determination about a person's responsibility for a loss, which sets whether they are paid, whether they are pursued, and what happens to their premium afterwards, and fault assessment drawing on unstructured claim narrative is precisely where language models encode the patterns present in historical handling.

The configurable design puts the fairness question further out of reach, since carrier written prompts shape how each agent reasons and no two deployments behave identically, so even a vendor level fairness test would not describe what any given customer has built.

Recourse is unaddressed and the agent list makes the exposure unusually direct. A liability check assigns fault, claim evaluation and scoring set what a claim is worth, and communications agents write to the claimant, so a person outside the contract can be told they are at fault, offered a figure and corresponded with, all shaped by a model configured by the carrier and supplied by this vendor.

That person has no relationship with either party's technology, receives no notice that an agent contributed, and has no published route to obtain or contest what it produced. Between vendor and carrier the position is more tangled than usual because the carrier writes the agent instructions, so responsibility for a defective output plausibly divides between the party that built the capability and the party that configured it. Nothing published allocates it, and no error handling policy or artificial intelligence specific service commitment was located.

The architecture implies a commercial language model and the record never names one. Configuring agents at the prompt level in plain language, defining a persona and directing actions, is the interface of a general purpose model rather than a purpose built classifier, and two passes located no provider, model family, version, hosting arrangement or fine tuning description for any of it.

One third party dependency is disclosed by name, a partnership with Floatbot covering conversational agents, which is a partial disclosure of the wider stack and does not describe what powers the claims agents themselves. The consequences for a buyer are concrete: no way to assess concentration risk, no notification commitment if the underlying model changes beneath carrier written prompts that were tuned against its behaviour, and no answer for a regulator asking where claimant data goes during inference.

Within claims the estate is complete and reaches further than most competitors on one axis that matters: money. The platform runs first notice of loss through settlement with configurable workflows, adjuster workload management, claimant communications and reporting, and then disburses payment at the end of it, so the vendor sits in the settlement transaction rather than handing off before it.

Connectivity is evidenced rather than asserted, through open interfaces, no code automation, a partner network exceeding 140, a joint platform with the core vendor Socotra combining policy and claims stacks, and a named conversational agent partnership. The artificial intelligence layer is integrated on the same principle, described as operating with immediate context of the data, rules, permissions and logic already configured in the system rather than as a separate console. What holds the grade at this band is scope: no policy administration, billing or underwriting capability exists, so this is always one component of a wider stack.

Delivery is cloud based and described as modern architecture with flexible implementation, and the one deployment figure on record is a typical installation of 45 days, roughly 30 of which go to configuring business rules and training rather than to technical work, which is a useful and honest breakdown though it dates from an interview several years old. Beyond that the published record is empty.

Two passes located no hosting provider, no region list, no processing location statement, no tenancy description, no residency commitment and no private deployment path. Presence appears concentrated in the United States, which reduces the cross border weight of those omissions without removing them, since the platform holds claimant photographs, liability assessments and payment instructions whose handling location several state and federal regimes care about.

Pricing is absent from every vendor surface, with demo request as the only route, consistent with this lane. One historical detail is the sole basis signal on record: in its early appraisal business the company charged per estimate produced, which is a transaction model rather than a subscription, and nothing published indicates whether the modern platform retained that shape or moved to a licence.

The commercial argument made is smarter claims for less total cost without disrupting existing operations, which is a comparative claim carrying no baseline or figure. Two questions matter and neither is addressed publicly: whether the artificial intelligence agents are included with the platform or charged separately, which matters because they are configurable and therefore usable at very different volumes by different buyers, and how payment disbursement is priced given that money movement usually carries its own economics.

Buyer coverage is wide across the claims handling ecosystem, spanning carriers, third party administrators, managing general agents, self insured organisations and insurtechs, and the platform is described as flexing to multi line models and multi party needs, which is the right shape for a market where the party handling a claim is often not the party carrying the risk.

Line coverage runs from auto, where the depth clearly sits, through property and commercial claims, with appraisal extending across passenger vehicles, boats, recreational vehicles, motorcycles, utility and exotic vehicles. Two limits hold the grade at this band. The estate is claims only, so a buyer gets no underwriting, policy or billing capability and this vendor is always one part of a wider stack. And evidence of presence outside the United States was not located in two passes, despite international expansion having been stated as an intention some years ago.