CLARA Analytics

The company describes itself as an artificial intelligence as a service provider and the removal test leaves nothing standing. Image recognition and natural language processing read medical notes, medical bills and legal demand packages that arrive as unstructured documents, and predictive models turn them into severity forecasts, attorney involvement predictions, provider rankings, fraud referrals and subrogation flags.

Strip the models and what remains is a repository of claim documents with no product around it. The CLARAty.ai platform combines predictive and generative models built specifically for casualty claims rather than a general purpose assistant pointed at insurance.

The framing is consistent and it is decision support rather than substitution: the platform is described as augmented intelligence and as a claims professional co-pilot providing guidance and risk analysis, with the adjuster setting reserves and making the call. The fraud module is described as giving the adjuster confidence to refer a claim to an investigative unit, which places the machine before a human referral decision rather than in place of one.

What is absent is the specificity Federato publishes: no named automation tiers, no threshold above which a recommendation is acted on without review, and no sampling or audit of how often adjusters simply accept a triage score. Given that a triage flag drives reserve setting and adjuster attention, the practical question is whether guidance functions as advice or as a default.

No accuracy figure, no validation evidence, no false positive or false negative rate, no model documentation and no error analysis by line of business or claim type were located. The return on investment claim is an outcome assertion, not a measure of whether a prediction was right. In place of accuracy the company offers the size of the contributory database as the argument, stating that the larger it is the more accurate the predictions become.

That is a plausible direction of effect presented as a substitute for measurement. The gap matters most on the fraud module, where the failure mode is a claimant wrongly referred to an investigative unit, and on triage, where a mis scored claim changes both the reserve and the service an injured worker receives.

Two customers are named with announcements: Eastern Alliance, a workers compensation specialist and subsidiary of ProAssurance, and Merchants Insurance Group, which adopted document intelligence and triage together. Beyond those the customer set is described by category rather than by name, as companies from the top 25 global insurance carriers down to large third party administrators and self insured organisations.

Approximately 63 staff and 64 to 72 million dollars raised across four to six rounds depending on the source, with the Series C led by Spring Lake Equity Partners. The headline outcome claim is that most customers exceed 500 percent return on investment, which is a vendor aggregate with no methodology, no per customer figure and no independent study behind it. Two carrier venture arms hold equity, QBE Ventures and Nationwide, which is a buyer adjacent party backing the product, the same shape recorded for Elysian and American Family Ventures.

This is the clearest cross customer data pooling arrangement in the index and it is unusual because it is the product rather than an unstated side effect. The contributory database has accumulated casualty claims outcomes from customers since 2017 and the company states plainly that a larger contributory database produces more accurate predictions, and observes that insurers have historically resisted sharing data because they believe it surrenders competitive advantage.

Naming the mechanism openly is more candid than Feathery describing pooling as a roadmap item or Glia describing an unbounded learning loop. What is missing is everything that would make it governable: no de-identification method, no opt out, no statement of what a contributing carrier's data may be used for, and no account of whether one carrier's loss experience informs guidance served to a competitor. Compare DwellFi tenant containment and Rulebase zero training on customer data, which are the opposite architecture.

The payload here is the most sensitive in this index. The platform reads medical notes, treatment records and medical bills for injured workers, alongside legal demand packages, and those records are contributed into a shared database. Workers compensation medical data sits in a partial and contested position relative to federal health privacy rules, which makes an explicit vendor statement more important rather than less.

No data protection agreement, retention schedule, de-identification method, subprocessor list or deletion commitment was located, and the only privacy artifact found is a cookie consent banner. The injured worker whose medical history is read, modelled and pooled has a relationship with an employer and a carrier and none at all with CLARA.

No attestation, certification, trust centre or security page was located, and no service organisation control report or international information security standard certificate is announced or offered on request. The only compliance artifact present is a cookie consent banner. For a platform holding injured worker medical records and legal demand packages from multiple carriers in one shared environment, this is the assurance gap that would surface first in any carrier vendor risk review. Compare Akur8, which announces its attestation with a named chief information security officer, and Rulebase, which names attestations plus residency.

One named federal regime is addressed as a product rather than as compliance language: CLARA MSP Compliance handles Medicare Secondary Payer submissions, an obligation with mandatory reporting duties and civil money penalties for late or inaccurate reporting, and the module is sold on cost of submission. That is concrete and specific.

Against it, nothing addresses the state layer where casualty claims handling actually lives: no engagement with state unfair claims settlement practices statutes, no reference to the model bulletin state insurance regulators have adopted on insurer artificial intelligence use, and no mention of state workers compensation regulators. CLARA holds no licence and needs none, which is the correct posture for a technology supplier under the index convention.

The exposure is wide and the only published statement about it is wrong. Four separate model classes touch people who cannot answer back: triage flags a claim as high risk which shapes reserve setting and adjuster attention on an injured worker, provider scoring ranks named medical practitioners in a way that can affect referral volume, litigation models predict whether a claimant will retain an attorney, and fraud models refer claimants to investigative units.

None has a described appeal, notice or correction route. The company's own published position is that artificial intelligence is known to introduce bias because it learns from humans, but that a larger dataset helps to minimise that risk. That is not how it works. A larger corpus drawn from historically uneven claims handling entrenches the pattern with more statistical confidence rather than diluting it, and scale is not a control.

Publishing an unsound answer signals the question was considered and closed. No fairness testing, no outcome analysis by claimant characteristic, no provider score methodology or appeal path. Same unfair claims settlement practices exposure recorded for Reserv and Elysian, on a much larger installed base.

No guarantee, indemnity, service level commitment on model quality or published falsifiable accuracy claim was located, so nothing commercially binds the vendor to its output. The recurring shape in this index is sharper here than almost anywhere because there are two classes of affected party who are not the customer.

The injured worker is triaged, scored for litigation likelihood and potentially referred for fraud investigation, without notice that a model was involved and with no correction route. The treating medical provider is ranked on performance by a system it does not subscribe to, in a way that can influence whether claims are directed to it, with no published methodology and no appeal. Adjuster review is a control for the carrier, not recourse for either of them.

The proprietary layer is well described: models are built specifically for casualty claims and trained on the contributory database, which is the company's own accumulated asset rather than a licensed third party corpus, and that shortens the chain at the point that matters most for the predictive modules.

Generative artificial intelligence is named as part of the platform and as the basis of the fraud module, and no model provider, hosting arrangement or data boundary for that layer is disclosed anywhere. Whether medical records and legal demand packages pass to an external model provider during document analysis is unanswered and is the first question a carrier privacy review would ask.

The platform is stated to integrate with the customer's claims system or risk management information system, and naming the second of those is the right detail because it is how a self insured employer without a carrier claims platform actually consumes this. Delivery is positioned as an intelligence layer over existing systems rather than a replacement, which lowers the adoption barrier for a carrier that has already validated its claims platform.

No individual claims administration system is named, so a buyer cannot confirm whether its own environment is supported without a sales conversation. That is what separates this from the A held by Federato for native connections to Guidewire, Duck Creek and Sapiens.

The delivery model is described as artificial intelligence as a service and more recently as intelligence as a service, which implies a multi tenant hosted platform, and the contributory database architecture requires customer data to reach a shared environment by design. No cloud provider, region selection, private deployment option or residency commitment was located.

The absence is more consequential here than for most vendors because the pooled architecture means data movement is not incidental to the product but constitutive of it, so a buyer cannot evaluate the arrangement without terms that are not published.

No pricing, no packaging tiers and no basis of charge are published anywhere, and every route into the product is a demo request. The only figure resembling a price is a claim that the Medicare Secondary Payer module saves over a third of the cost of submissions, which describes the buyer's existing spend rather than what CLARA charges for it. Category norm rather than a failing, but it sits well behind Sumsub's published per verification rates, and behind the disclosed billing bases from Akur8, Glia and LeapXpert.

Five distinct buyer types are named and served: insurance carriers, managing general agents and managing general underwriters, reinsurers, third party administrators and self insured organisations. That last one matters because a self insured employer buys claims intelligence with no carrier in the chain at all, and the platform is stated to integrate with risk management information systems as well as carrier claims systems to reach them.

Three casualty lines are covered end to end, workers compensation, commercial auto liability and general liability, with general liability completing the set. The stated range runs from top 25 global carriers down to individual self insured employers.