Bdeo

The removal test leaves nothing behind, which is the clearest case for the top band. Take the computer vision away and there is no product: no damage detection, no severity determination, no pre existing damage check against historical images, no automated underwriting inspection of a used vehicle. What would remain is a photograph upload form.

The company was founded in 2017 specifically to advance this form of learned capability in insurance rather than adding it to an existing business, its own account of the market opportunity rests on the model being good enough to replace a physical inspection for the roughly 70 percent of claims that are minor, and it states that 65 percent of its staff work on the technology itself. Every regional deployment described, from telematics triggered capture in Germany to underwriting inspection in Spain, is a different application of the same model rather than a different product.

The published positions differ by region and nothing reconciles them. In the United Kingdom the framing is assistive, described as adding a layer of visual verification to the process and supporting claims decisions, with the model giving insurers vision rather than judgement.

The company's founding thesis points the other way, resting on roughly 70 percent of claims being minor enough to assess without a physical person present, and the Spanish underwriting position implies automated inspection at volume rather than case by case review. Both can be true if the boundary is configured per client, and the boundary is exactly what is missing.

Nothing published names what the system determines alone, what routes to a human assessor, what confidence threshold triggers referral, or who reviews a severity determination before it becomes a settlement figure. For a product whose outputs reduce or decline payments, no automation rate is published and no appeal or overturn rate is published beside it.

The number that would matter most is not published anywhere. A model that detects damage, classifies it and determines severity has a measurable accuracy against expert assessment, and two passes located no accuracy rate, no false positive or false negative rate, no confidence scoring description, no benchmark against human adjusters and no validation methodology.

Investor commentary refers to the reliability of results as a factor in funding the company, which is a third party's impression rather than a published measurement. Also absent are drift monitoring, revalidation cadence and any account of how performance is maintained as vehicle models, construction materials and capture devices change, which is a live issue for vision systems whose training data ages against the physical world it observes.

Customer evidence is unusually strong for a company of this size and outcome evidence is absent, which is the split that sets this grade. Named clients include several of the largest carriers in Europe and Latin America, with Reale, Mapfre and Generali in Spain, Ageas and Fidelidade in Portugal, Zurich, Allianz and BBVA in Latin America and Hollard in South Africa, and one is documented as a joint build rather than a deployment, with Mutua Madrileña named as co creator of a system to automate policy underwriting.

Aggregate scale is more than 50 insurers in more than 25 countries. Two reservations apply. The most striking penetration figures, handling more than half of Spanish motor underwriting and working with 8 of the 10 leading Spanish motor insurers, date from mid 2023 and have not been restated since, so a buyer should treat them as historical. And no outcome is quantified anywhere: no automation rate, no accuracy figure, no cycle time reduction and no cost saving, for a product whose entire proposition is measurable efficiency.

This grades below the fraud specialists in this lane because those vendors publish data handling positions and this one does not. Two passes located no statement on whether client images are used to train or improve models, whether one insurer's captured material informs capability sold to another, what de identification is applied, or what happens to image libraries on termination.

The training data question is the central one for a computer vision business, because model quality here comes from volume of labelled damage imagery and the only source of that volume is customer claims, so a buyer should assume the question is live until answered.

The historical image capability sharpens it further: checking for pre existing damage means the system retains and re queries images captured for earlier, unrelated claims, which is a repurposing of policyholder data that no published document describes or bounds.

Two passes located no privacy policy content, data processing agreement, retention schedule or subprocessor list, and the absence is more consequential here than for most vendors in this lane because of what the product ingests.

The input is photographs and video captured at the scene of a loss or at a vehicle inspection, which routinely contain material beyond the damage itself: registration plates, vehicle identification, the interior and exterior of a policyholder's home, documents held up to a camera, and frequently people.

That is personal data under European law by any reading, and the company is a European vendor operating across more than 25 countries including Latin American markets with their own regimes. Nothing published states how long images are retained, whether they are used beyond the assessment they were captured for, what is redacted, or where the historical image library used for pre existing damage checks is held and who may query it.

Two passes across the company's site, its resource library, its press coverage and third party profiles located no trust centre, no service organisation control report, no information security certification and no security page.

The grade rests on the inference standard used elsewhere in this index rather than on published evidence, and the supporting facts are reasonable: the named client list includes several of the largest carriers in Europe and Latin America, and organisations of that size do not connect a supplier to live claims and underwriting workflows, or route policyholder imagery through it, without a security assessment they have the expertise to conduct.

The controls therefore probably exist and are not published. The practical consequence is that a buyer can assess nothing before entering a confidentiality agreement, which is a heavier burden for a smaller vendor because there is less public corroboration to fall back on.

The company holds no insurance licence and does not claim one, the ordinary position for a technology supplier, and it operates as a processor inside carrier workflows across more than 25 jurisdictions. Its public regulatory presence is recognition rather than status, holding a European Innovation Council Seal of Excellence and having received regional public grants and investment from a Spanish state industrial technology body, all of which speak to innovation assessment rather than to supervision of its outputs.

Two passes located no statement on how the product supports a carrier's own regulatory obligations in any market, no position on European artificial intelligence regulation, and no description of how the same model is validated against differing national requirements when severity determinations feed settlement decisions in jurisdictions with materially different claims handling rules.

Two passes located no responsible artificial intelligence statement, no governance framework, no fairness testing, no model card and no bias evaluation. The precision worth recording is where the regulatory pressure does and does not fall.

Motor and household damage assessment sits outside the explicit high risk schedule of the European artificial intelligence regulation, which names risk assessment and pricing in life and health insurance rather than property lines, so the sharpest legal hook that applies to several peers does not apply here. That lowers the compliance exposure without lowering the fairness question.

Computer vision performance varies with image quality, lighting, device and capture conditions, and those correlate with the claimant's circumstances, so a system assessing damage from photographs taken by policyholders can systematically underperform for some populations. Nothing published indicates that this has been tested.

The affected party is the policyholder and nothing published gives them a route. A severity determination made from photographs sets what a claim is worth, a pre existing damage finding drawn from historical imagery can reduce or defeat a claim entirely, and an underwriting inspection can shape whether cover is offered and on what terms.

In each case the person affected has no relationship with this vendor, receives no notice that a model produced the assessment, cannot see what the model concluded or why, and has no published mechanism to obtain a human reassessment. Pre existing damage detection deserves particular attention, because a finding of that kind functions as an accusation as much as an assessment and is drawn from images captured for an unrelated earlier purpose. Between vendor and insurer nothing published allocates liability for a misclassified loss, a severity error or a false pre existing damage finding.

The models are almost certainly the company's own, and almost certainly is doing the work because nothing published says so. The business was built around proprietary computer vision from 2017, well before commercially available vision models made a licensing route practical, and 65 percent of staff are stated to work on the technology, both of which point to internal development.

What is absent is any statement of it: no architecture description, no account of how training data was assembled and labelled, no disclosure of whether any third party model or vision service is used for any part of the pipeline, and no version or change notification commitment. For a vendor whose entire value is model quality, publishing nothing about model provenance leaves a buyer unable to assess what it is actually buying or what would happen to performance if a dependency changed.

Integration reaches three parties rather than one, which is more than the product category requires and is the substance behind this grade. The insurer side connects into claims and underwriting workflows with the company stating it implements where the client needs it and at the pace required, indicating adoption by use case rather than a platform replacement. The policyholder side is a capture interface reached at the moment of loss or inspection.

The repair side is genuine and less common: workshops are helped to prepare estimates, and a repairer's estimate can be validated against the image evidence to support authorisation, which places the vendor inside the insurer to supplier settlement loop rather than only in the customer facing part of it.

The German deployment demonstrates the deepest integration published, connecting to local telematics providers so an incident is detected automatically and capture begins without anyone initiating it. No interface documentation, integration count or partner marketplace was located.

Two passes located no deployment documentation of any kind: no hosting provider, no region list, no processing location statement, no tenancy description, no residency option and no on premises path. The gap carries weight in proportion to the geography, since operations span more than 25 countries across the European Union, the United Kingdom, Latin America and South Africa, and image data captured from a policyholder in one of those markets is subject to that market's rules on where it may be processed and stored.

A buyer in Colombia, Mexico or South Africa has nothing published to check against, and a European buyer cannot establish whether processing stays inside the bloc. The historical image library used for pre existing damage checks compounds this, since it implies persistent storage of claim imagery whose location is undescribed.

Pricing is absent from every vendor surface, with a demo request as the only route, consistent with this lane. The commercial claim made is reduced operating cost for insurers, stated repeatedly across the company's own material and its investors' commentary, and never quantified with a figure, a baseline or a unit.

One structural detail bears on cost without describing it: the company states it implements where the client needs it and at the pace the client needs, which indicates modular adoption by use case rather than a platform commitment, and that shape usually implies per transaction or per inspection charging. Nothing published confirms it.

Two questions matter for a buyer here and neither is addressed: whether charging follows images processed, claims assessed or a subscription, and whether the underwriting inspection and claims assessment use cases are licensed separately.

Geographic reach is genuinely broad for a company of roughly 63 people, covering more than 25 countries with named deployments across Spain, Portugal, Italy, France, the Nordics, the United Kingdom, Mexico, Colombia, Argentina and South Africa, and the regional variation in use case indicates real local implementation rather than a single product shipped everywhere.

Buyer types extend past insurers to fleet operators, leasing companies and rental firms, and the workflow reaches repair workshops as participants rather than as subjects. What holds the grade at this band is line of business concentration.

The product covers motor and household property and nothing else, which excludes commercial lines, specialty, liability, life and health entirely, and motor is clearly the centre of gravity given the Spanish underwriting position, the German telematics integration and the fleet product. Depth in two personal lines is a narrower footprint than the country count suggests.