Tesora

The agents are the product and the company describes itself as AI native rather than as software with models added. Strip the models and there is no workbench left, only the spreadsheets and email the product exists to replace. Worth recording as the sharpest contrast in this pocket: the two largest vendors here both sit at B because a deterministic rating engine remains underneath their models, and this one has no such residue because it never had a rating engine to begin with.

Strong B and better specified than most vendors ten times its size. The division of labour is stated explicitly rather than implied: the product handles the assembly and the actuary stays in charge of every decision, which is a positional claim about where the human sits rather than a general assurance that humans are involved.

Reviews are described as aligned to actuarial professional standards, which names an external standard the review is measured against, and that is the same class of specificity that separates an auditable control from guardrail language.

Held at B and not A on two points recorded rather than smoothed: alignment to a standard is not enforcement of it, and nothing describes what a review actually checks, what an agent may do without approval, or what prevents an agent from deploying a rater the actuary has not examined. The agents genuinely act, since building and deploying a rater is acting.

Earned on two named mechanisms rather than on assertions, and the vendor frames the problem correctly. It states that generic AI was untrustworthy because its analyses were often unreproducible, and answers with reproducibility as a product property plus provenance: the agents record where data and logic came from so they can cite the source later, and rebuilt raters become versioned models.

Reproducibility and source citation are exactly what a reviewing actuary or examiner needs and almost nothing in this index publishes either. Held at B and well short of A: no accuracy figures, no validation methodology, no error rate on the document extraction the whole pipeline depends on, no drift monitoring and no external assessment.

One claim is flagged rather than credited, and it is the most checkable thing on the site: when an existing spreadsheet rater is rebuilt, existing logic is said to be preserved exactly. That is an absolute stated about a translation process, with testing mentioned and no test method, tolerance or failure rate published.

No named customer, no quoted executive, no case study, no quantified outcome, no independent evaluation. Design partners are referred to and never identified. Two things exist that are deliberately not credited because they are not customer evidence: an accelerator backing, which is investor validation, and a former chief actuary working with the team, which is a credential of the vendor rather than a reference from a buyer.

Nothing states whether a carrier's loss experience, rating logic or submissions inform anything beyond that carrier's own work. The question matters more than usual for this shape of product because rating logic is among the most commercially sensitive material an insurer holds and the agents ingest it directly, and because the same agents also read competitors' public rate filings, so the product sits between public and proprietary rating information by design.

Nothing published on retention, access or processing commitments, for a product whose inputs are a carrier's loss runs and broker submissions containing insured and claim level detail.

No certification, security page or trust portal. One claim is recorded as a shape rather than credited: the audit trail is described as meeting the grade expected for financial reporting controls, invoking the American corporate reporting statute as a quality standard. Nobody certifies against that statute, and it governs the filer's controls rather than a supplier's product.

This is a milder version of the catalogued pattern where statutes and uncertifiable frameworks are listed alongside real certifications, and it is worth noting because the claim is doing credential work while naming no assessor, no scope and no report.

No licence or standing claimed, and none would be expected for a software supplier at this stage. Recorded with one distinction worth keeping for this pocket: the product reads state rate filings as an input, which is consuming public regulatory data and is a different thing entirely from having a model examined and approved by a regulator, which is what earns an A on this axis elsewhere in this pocket.

No fairness testing, protected characteristic handling or governance disclosure. The exposure is direct rather than incidental: the rating agents produce base rates, factors and increased limits factors, which is the construction of the price a consumer is charged. Two vendors in this same pocket disclose bias testing and this one, whose agents build the rating structure itself, does not.

Nothing published on responsibility for a wrong outcome, and the exposure is unusually concrete for a company this young. The agents build and deploy raters, so an error does not stay in an analysis, it reaches the market as a rate charged to policyholders.

The spreadsheet rebuild path is the sharpest version: a translation error in parsing formulas or identifying factor tables produces a rater that prices differently from the one the carrier approved, and nothing addresses detection, correction, or who bears the cost of business written on it. Reproducibility means the error can be traced afterwards, which answers a different question from who is responsible for it.

No model, provider, family or version named. The capability is described as frontier agents, which is a phrase that gestures at model quality while disclosing nothing about the dependency, and it is worth cataloguing alongside proprietary models as a term that substitutes for a name. Notable in a product whose entire value proposition is that generic AI could not be trusted for this work: the claim that this system is different rests on the model layer and the model layer is unnamed.

Two real capabilities and no named integrations. The product reads state rate filings, loss runs, statements of value and broker submissions in almost any file format, and it exposes a callable pricing interface so a carrier can invoke a rate from its own systems, which is a genuine production path rather than an export.

Against that: no policy administration system, rating platform or data provider is named anywhere, no interface documentation is published, and the deployment target for a rebuilt rater is the vendor's own runtime rather than a carrier's existing stack. Ingesting documents is not the same as integrating with the systems of record.

Nothing published. No hosting model, no cloud named, no options, no residency statement, no customer responsibility split, for a product that ingests a carrier's loss experience and broker submissions.

No pricing, no commercial model and no indication of whether the product is licensed by seat, by rater, by call volume or by subscription. Every route is a contact form.

A genuine C rather than a thin one, and rare on this axis. The company was founded in 2025, has a founding team of two to four, and describes working with design partners rather than a customer base. Target buyers are stated with precision, namely specialty property and casualty carriers, reinsurers and managing general agents, and a stated target is not coverage. No institution of any size is named anywhere and no count is claimed. Graded on what exists today, with the note that this is a startup at the beginning of its commercial life rather than a vendor failing to disclose.