Kalepa

Strip out the models and Copilot shows an empty screen. The workbench's contents are produced by models. They make the structured read of a messy submission document and the cross reference against billions of external data points. They also find the exposures the broker did not disclose and set the priority score against appetite. The company states the product works on day one with no integrations, so there is no system of record beneath it supplying anything. Some orchestration shaped workbenches in this market predated their models and would survive their removal. Kalepa had no product before the models.

Oversight rests on three named mechanisms, each with a stated place in the path. An ensemble engine selects the model or method for a task in real time. It is described as knowing when to automate rather than amplify underwriter judgment, so the automation level is itself a decision rather than a setting. Independent judging agents validate critical extracted fields, measuring confidence continuously and catching errors before output reaches the underwriter.

The division of labor is stated narrowly: Copilot reads, cross references and scores, and the underwriter decides, with the reasoning behind every flag visible. One limit belongs to the verification approach as a whole, not to Kalepa alone. The judging agents are independent of the extraction agent but not of the model class, so a failure mode shared across both passes verification unnoticed. The same pattern appears in health AI self verification and in any product that validates model output with another model.

The architecture is real product design rather than assertion. Independent judging agents validate critical fields, confidence is measured continuously rather than sampled, and the company frames consistency at scale as the objective. The gap is one the vendor draws attention to itself. Kalepa claims market leading underwriting accuracy and verifiable results but publishes no accuracy figure or error rate. There is no benchmark and no independent evaluation.

An unquantified superlative about accuracy has the same shape as a compliance vendor advertising a large cut in false positives while publishing nothing about the cases it misses. The measurement infrastructure described is exactly what would produce the number, which makes its absence more conspicuous.

The carrier behind the hard numbers is described rather than named, and the three named carriers carry no figures, so the two never join. The named carriers are Munich Re Specialty, Bowhead and SECURA, spanning reinsurance specialty, excess and surplus, and regional mutual business. A published case study covers a mid market division of a top 15 global property and casualty carrier.

Within two months, it reports, time to quote fell 40 percent and bind rate rose from 30 percent to 34.5 percent, while compliance with appetite and guidelines passed 95 percent. All of it was achieved with 15 percent fewer underwriters on staff. A separate figure claims complex risks quoted 58 percent faster. Kalepa has been recognized in the InsurTech100 four years running.

The cross customer question is open. The company states that Copilot automatically learns what the best underwriters are doing, which makes that question unavoidable rather than theoretical. Nothing published says whose underwriters, or whether behavior learned inside one carrier's deployment informs scoring or prioritization served to another. The platform also aggregates third party and open web sources into a single view. Nothing describes how source reliability is assessed or what happens when an external record is wrong. United States server location is stated, but a data location is not a stewardship position.

A privacy policy is published. There is no public data processing agreement or subprocessor list, and no retention schedule or lawful basis statement. Exposure would normally be low for a commercial lines product, since the subject of a submission is a business. It is higher here because the platform deliberately retrieves criminal activity and pending legal action records from third party and open web sources.

Those records concern identifiable individuals connected to the insured business. Material of that kind attracts consumer reporting and data protection questions in several jurisdictions, and nothing public addresses them.

A SOC 2 Type II report stands alone. Kalepa states it clearly on the product page rather than burying it in a footer, alongside a commitment that data sits on United States based cloud servers. Type II is the audited operating effectiveness report, not the point in time design report, and it is the version enterprise buyers want. There is no ISO 27001 or ISO 42001 certification, no trust center, and no subprocessor list or penetration testing statement. At least two competing underwriting platforms hold three audited certifications each, and a carrier running a vendor security review would notice the difference.

As a technology supplier to carriers and managing general agents, Kalepa holds no insurance license, which is the correct posture. No statement of regulatory position is published in any form. There is no classification under the EU AI Act and no engagement with insurance supervisors. Nor is there sandbox participation, program admission or a supervised test. Compliance with appetite and underwriting guidelines is a product capability measured against the carrier's own rules, not regulatory standing.

Nothing is published on fairness testing, differential outcomes or governance review, and the exposure is more concrete than usual for commercial lines. Copilot surfaces criminal activity and pending legal actions from external and open web sources as underwriting exposures. The vendor's own case study describes the target exposures as violent incidents at insured premises.

Screening businesses on crime data attached to a location, in a United States market, imports whatever geographic patterns that data carries. Two otherwise identical restaurants in different ZIP codes will not screen alike. Two facts soften this. The insured is a business rather than a consumer, which weakens the protected class analysis considerably, and state unfair discrimination duties fall on the carrier rather than the supplier.

The recourse gap has a specific shape. Copilot pulls criminal records, legal actions and open web material about a business into the underwriting file. A business declined or surcharged on the strength of a third party record it never sees has no route to inspect or correct it. The confidence and judging agent architecture guards against Kalepa misreading a document, a control against the vendor's own error.

It does nothing about an upstream source being wrong, and that is the failure mode most likely to reach an applicant. No error rate or remediation commitment is published, and there is no liability position or correction path.

Multiple models are in use and none is named. The ensemble engine is described as selecting the right model or method in real time. No provider or model family is disclosed, nor any version or licensing arrangement, and the claim that the models are best in class carries no supporting detail. The United States server commitment does establish country of processing for the data, though it says nothing about which models process it.

Copilot deliberately trades integration depth for speed, and the company says so. It works out of the box on day one with no integrations, and email inbox, portal and policy system connections are available afterwards. That is a real commercial advantage against implementations that take a year. A carrier gets value quickly but does not get Copilot wired into the systems where the policy is eventually issued.

No policy administration system, rating engine or workbench is named as a supported integration, and no integration partner is named at all. Competing platforms name specific core platforms, rating engines and data providers.

All data is stored on United States based cloud servers. That is an actual residency statement rather than a gesture at one, and several far larger insurance platforms leave the question open. The answer is single valued, and the company operates across more than ten countries. No European or United Kingdom region, single tenant option or self hosted path is offered or mentioned. For a carrier subject to European data requirements, the clarity of the disclosure reveals the problem rather than solving it.

The published case study gives a carrier every input for a return calculation except the cost. A 40 percent reduction in time to quote, a bind rate moving from 30 to 34.5 percent and 15 percent fewer underwriters are all quantified. No pricing or tier structure is published, nor any billing basis or indicative range, and the route to a number is a demo request.

This is depth within one half of the insurance market. Kalepa serves commercial and specialty property and casualty carriers and managing general agents, and life, health and personal lines are absent. The named client base spans reinsurance specialty, excess and surplus and regional mutual business, a wider spread of carrier type than the company's size would suggest. Line coverage handles multi location, fleet and additional insured complexity, not only simple risks. Scale is a limit too, with roughly 54 staff and a Series A balance sheet against peers serving hundreds of billions in premium.