Greater Than

The removal test leaves raw position data. Machine learning built on deep learning layers decomposes driver behaviour into thousands of variables to find patterns the company states were previously invisible, producing individualised predictions of accident probability and emissions per trip in real time. The models are protected by seven patents and have been trained continuously on driving data since 2004, yielding more than seven billion individual driver profiles. Nothing about the proposition, from crash probability scoring to dynamic premium calculation, exists without them.

Scores are produced automatically and continuously, and the stated use in fleets keeps a person in the loop, with risk managers identifying the highest risk drivers so that training, coaching or other intervention can follow. On the insurance side the position is different, since the platform enables dynamic premiums reflecting the actual risk of the individual driver in real time, which is pricing that adjusts without a human deciding each time. Nothing published describes a confidence indication on a score, a threshold above which a person reviews before a premium moves, or what recourse exists inside the system when a score appears wrong.

This vendor engages model quality more substantively than most, and does so by attacking the industry standard rather than defending its own. Its published position is that harsh braking and acceleration events, the signals telematics has ranked drivers on for two decades, showed no relevant relationship to crashes in research the author conducted personally, which is a falsifiable claim about a competitor methodology and an unusual thing for a vendor to put in writing.

Behind it sit seven patents, 22 years of continuous training data, and selection into a European Union funded research programme, which is external technical scrutiny. What holds it below the top grade is the headline figure: an accuracy of 99.98 percent to factual risk and accident probability is quoted through third parties with no methodology, denominator or validation source attached, and a number that precise invites the question of what exactly was measured.

A global insurer and a major vehicle manufacturer's mobility arm are both named as clients, and a Nordic telematics group launched an entire insurance broker brand on these scores, covering more than 200,000 connected vehicles in one market, with its chief executive describing risk segmentation and pricing built directly on the models.

A leading policy administration platform partnered to distribute the crash probability and climate impact scores to its own insurer customers, whose executive is quoted on the arrangement. Training scale is exceptional and specific: billions of real world trips from more than 100 countries, equivalent to over 855,000 person years of driving. The company is publicly listed, 22 years old, and its technology was selected for a European Union funded research programme.

Pooling is the stated foundation of the product, with models trained on billions of trips across more than 100 countries and every new customer's driving contributing to a common corpus of over seven billion driver profiles. That is disclosed as the source of predictive strength rather than concealed, which is the position recorded for Upstart, and it means one insurer's book improves the scores sold to its direct competitors. Nothing states whether a customer can decline to contribute, what is retained when a contract ends, or how the data of a fleet's drivers is separated from the insurers who might later price them.

No data protection agreement, retention schedule, subprocessor list or consent description was located, and the payload is among the most granular personal behavioural data in this index. Continuous position and driving data is captured second by second and assembled into a persistent individual profile, so the platform holds where a person went, when, how fast and how carefully, accumulated over years.

European operation imposes a statutory floor and nothing published states how long a driver profile persists, whether it follows a person between insurers or fleet employers, or what a driver is told when their employer enrols the vehicle they drive.

No attestation, certification, trust centre or enumerated framework was located. A global insurer and a major policy administration platform have both completed vendor assessment on this company, so the assurance exists privately, and for a platform holding continuous location and behavioural records on individual drivers across many jurisdictions, publishing that control set would address the question every prospective insurer and works council will raise.

No supervisor, statute or instrument is named. The omission carries weight because motor insurance pricing is among the more prescriptively regulated activities in European financial services, with restrictions on permitted rating factors and on differentiation between customers, and because European rules on artificial intelligence now treat risk assessment and pricing in insurance as a category requiring specific governance. A platform whose entire purpose is individualised premium differentiation across those markets identifies none of it.

The company makes a genuine fairness argument and it is worth taking seriously. Conventional motor risk ranking rests on lagging indicators, licence violations and crash history, which price a driver on a past they cannot change, and the company states its own research found no relevant relationship between the conventional event signals and actual crashes.

Pricing current behaviour instead replaces history with something a driver controls, and the broker brand built on these scores markets exactly that as customer beneficial segmentation. The unaddressed exposure is that behaviour is not free of circumstance either: night shift workers, urban drivers, long commuters and those in older vehicles drive in measurably riskier conditions for economic reasons, so behavioural scoring can reproduce income patterns without any protected characteristic appearing as a criterion. Replacing lagging indicators with behavioural ones moves the fairness problem rather than resolving it, and no analysis across driver populations was located.

No commercial guarantee or indemnity was located, and what earns this grade is something almost nothing else in this index provides: the person being scored sees their own score. Drivers receive their risk profile and targeted guidance on avoiding crashes, one deployment sends regular performance reports directly to the customer, and the loyalty product rewards improvement, so the subject of the model is told what it concluded and given a route to change it.

That converts an opaque rating into a feedback loop. What is missing is challenge rather than visibility: nothing describes how a driver disputes a score they believe reflects the conditions they drive in rather than how they drive.

The data chain is unusually short and that is a genuine strength: driving data is collected first hand through the company's own device, application or connected vehicle feeds rather than licensed from brokers, and the models are its own and patented, so no external provider sits between the raw signal and the score. A named public cloud is disclosed as infrastructure for at least one product. What is not published is anything further, with no model provider identified, no subprocessor list and no statement of what mapping, weather or vehicle data supplements the driving signal.

Distribution reaches insurers through the systems they already run rather than requiring them to adopt another. A partnership with a major policy administration platform places the crash probability and climate impact scores inside that platform for its insurer customers to consume directly, which is the shortest path into underwriting and pricing workflows. One product is published on a leading cloud marketplace for procurement and deployment.

Data ingestion is equally flexible, spanning an on board diagnostics device, a smartphone application, and existing connected vehicle feeds, and one offering is stated to require no technology integration at all to be applied across an existing customer base, with driver onboarding handled by text message.

A named public cloud is identified as the platform behind at least one product, which is more than most vendors disclose, and no region selection, residency commitment or private deployment option was located. The question is material given operation across more than 100 countries and the nature of the payload, since continuous location data about identified drivers is treated as sensitive in several of the markets served and its processing location would ordinarily need stating.

No pricing, packaging or basis of charge was located. One product is listed on a major cloud marketplace, which implies a defined commercial unit exists, and no rate is published there or elsewhere. Nothing indicates whether charge falls per scored driver, per vehicle, per trip or as an enterprise licence, which matters given buyers range from a single fleet to a global insurer scoring an entire motor book.

Four buyer types are served through the same scores, spanning motor insurers, fleet operators, mobility and car sharing providers and vehicle manufacturers, and a broker brand has been built on the platform, which shows it supports both the underwriting and distribution ends of the chain.

Geographic reach is genuine, with training data drawn from more than 100 countries and the scoring described as geography independent, alongside named deployments in the Nordics, Australia and with a global insurer. The limit is line of business: this is motor risk exclusively, and everything the company does depends on a vehicle being driven.