Shift Technology

Nothing survives the removal test, and the company has never been anything else. It states a 100 percent focus on insurance from its founding in 2014, and every product it sells is a model output rather than a workflow with models attached: fraud detection across claims, applications and organised rings, underwriting fraud caught before binding, subrogation opportunity assessment, payment integrity, and claims automation from first notice of loss.

There is no policy administration system, no billing engine and no system of record underneath to survive if the models were removed. The technical lineage supports the claim rather than merely asserting it, with an early adoption of large language models in 2020, more than 2.6 billion policies and claims analysed, and a current description combining generative, agentic and predictive approaches. The agentic layer extends model dependence rather than diluting it, with agents drafting demand packages, guiding negotiation and estimating medical treatment duration.

The stated posture is consistently assistive and a customer quote demonstrates it in practice, with one area where the agents go further than that framing suggests. On the assistive side the language is uniform: artificial intelligence empowering human experts, agents that recommend relevant next actions for claims handlers, and a published customer statement describing the value as consistently identifying suspicious activity at first notice of loss and assigning the claim to the appropriate expert for investigation.

Fraud detection surfaces cases for special investigation units rather than disposing of them, which leaves the adjudication with a person. The area to examine is the newer agentic suite. Agents that draft an initial demand package, guide a negotiation, estimate medical treatment duration and develop a treatment plan are producing substantive work product rather than surfacing a signal, and claims automation includes straight through processing by design. Nothing published states what an agent may complete without review or what a handler must approve.

Real published technical work about the field, and no published measurement of its own models. The technical work is genuine and rare: an ongoing research series includes an edition evaluating reasoning models against performance, cost and latency across insurance use cases, which is comparative model evaluation methodology put into the public domain rather than a marketing paper. Explainability is a stated design property and a customer corroborates it.

Scale gives some confidence in the training base, at more than 2.6 billion policies and claims analysed. What is missing is the product's own numbers. Across two passes no detection rate, false positive rate, precision or recall measure, validation methodology, sample or observation period was located for any detection product, and the headline figure of more than 5 billion dollars in fraud savings is an aggregate across the customer base with no method attached. For fraud detection the false positive rate is the number that matters most, because each one is a legitimate claimant investigated.

Among the best evidenced records in this index, corroborated from several directions rather than one. Customers are named with named executives on the record: a global insurer renewed a five year strategic agreement in March 2026 extending a collaboration begun in 2016 across 15 countries, quoting its own chief transformation officer for European markets, and a second large European mutual signed as a strategic partner for fraud and risk.

Scale is stated precisely at more than 115 insurance customers across 25 countries with more than 5 billion dollars in reported fraud savings. Independent recognition is current and specific, including a luminary designation from a named analyst house for insurance fraud detection.

The strongest single item is structural: the dominant property and casualty core platform vendor named this company its strategic partner for insurance decisioning and separately invested in it, which is validation by the party best placed to build the capability itself. A national industry body selected it to build a country wide motor fraud platform.

Explainability claimed by the vendor and corroborated by a customer, plus published technical research, which together put this above the norm. The vendor describes its insurance grade artificial intelligence as accurate, explainable and secure, which alone would be marketing.

What lifts it is that the customer says the same thing independently: in announcing a five year renewal, a global insurer's own executive framed the relationship around combining technology with human expertise, and the vendor's chief executive characterised the value as explainable, insurance specialised artificial intelligence. A customer signing a five year commitment and describing the technology in those terms is meaningful corroboration.

The company also publishes an ongoing technical research series, including an edition evaluating reasoning models on performance, cost and latency across insurance use cases, and threat research on document fraud and synthetic media in the generative era. Absent across two passes: model card, published evaluation of its own models, red team result, incident disclosure and acceptable use boundary.

Nothing is published on a privacy or trust surface, and what can be established about the regimes this company operates under comes from an unusual place. A senior security role advertised by the company names the intersection it expects that person to manage: European data protection law, the international privacy information management standard, a health information security framework, and the French health data hosting certification, alongside financial services regimes.

That indicates a genuine and broad privacy infrastructure including health data, which matters here because the platform's agents assess whether medical billing applies to a claim, estimate treatment duration and develop treatment plans, meaning medical records pass through it. But a recruitment advertisement is not a privacy posture, and across two passes no published privacy policy content, data processing description, retention schedule or subprocessor list was located. The claimant whose medical file is analysed is not the customer and has no described route to any of it.

The most detailed public account of this company's security and compliance posture is a job advertisement, which is worth stating plainly because it is both informative and the wrong source. A senior security role posted by the company describes maintaining compliance with core standards naming a service organisation control framework and the international information security management standard, while managing an intersection that includes the artificial intelligence management standard, the privacy information management standard, a health information security framework, the French health data hosting certification, the European operational resilience regime and New York state financial services rules.

That is a serious and specific compliance surface. It is also unverifiable: across two passes no trust centre, certification page, downloadable certificate, attestation report, penetration test summary or subprocessor list was located. A related observation belongs on the record rather than in the grade.

The company was recruiting a chief information security officer in 2026 while serving more than 115 insurers across 25 countries, and whether that is a first appointment or a replacement is a fair question for a buyer to ask.

An unregulated software vendor whose regulatory surface is visible mainly through a job advertisement, which is the finding rather than an aside. A senior security role posted by the company names the regimes it expects that hire to translate into engineering controls: the European artificial intelligence regulation, the international standard for artificial intelligence management systems, European data protection law, a health information security framework, the French health data hosting certification, the European operational resilience regime for financial entities, and New York state financial services rules.

That is a precise and demanding list and indicates the company knows exactly where it sits. What is absent is any of it stated as a position. Across two passes no published statement on the artificial intelligence regulation was located, despite this being a French company selling automated decisioning into insurance underwriting and claims across 25 countries, an application the regulation treats as high risk, and no authorisation, registration or supervisory outcome was found.

Explainability is claimed and corroborated, and no fairness evidence exists behind it. The exposure is broad because of where these determinations land. A fraud flag routes an individual claimant to a special investigation unit, which means delay, scrutiny and potentially denial for a person who is never told a model selected them. Underwriting fraud detection can prevent someone obtaining cover at all.

And the medical agents reach further still, assessing whether billing applies to a claim, estimating treatment duration and developing a treatment plan, which shapes what an injured claimant actually receives. Insurance fraud detection carries documented disparate impact concerns, and detection built partly on claim characteristics correlating with geography, vehicle age, provider choice or repair network can distribute suspicion unevenly while appearing neutral.

Across two passes no fairness testing, disparate impact analysis, model card, or conformity work under the European artificial intelligence regulation was located. Explainability partly answers this, since an explained flag can be reviewed, but review is not measurement.

No commercial instrument is published. Across two passes no terms of service, master agreement, warranty, indemnity, liability cap or service level was located, and nothing states what an insurer is owed if detection fails or misfires. The uncovered exposure has two distinct shapes here. For the customer, a missed organised fraud ring is a direct loss and a false positive rate that is too high is an operational cost, and neither is addressed. For the third party it is sharper.

A claimant wrongly routed to a special investigation unit experiences delay, intrusive scrutiny and possible denial at the point they are least able to absorb it, and someone whose treatment duration is estimated low or whose medical billing is judged inapplicable may receive less care than they need. Those people have no relationship with this vendor, are not told a model made the assessment, and have no published route to see or contest it.

The foundation model dependency is named openly, which most vendors in this index avoid. Through the cloud provider's own published case study the company discloses that it runs on that provider's hosted large language model service and vision service within its artificial intelligence platform, uses machine learning based optical character recognition for document extraction, and adopts the provider's latest models soon after release so customers get improved detection quickly.

It separately discloses that it began using large language models in 2020. A buyer therefore knows whose foundation models read their claim files and policy documents, and can assess that provider's own terms and data handling directly. Two qualifications belong on the record. No model family or version is named, so the specific model behind a given decision is not identifiable. And the disclosure sits on the cloud provider's marketing surface rather than the vendor's own, which is where a buyer would look and where such pages are withdrawn without notice.

The strongest distribution position available in this market, and it goes beyond a partnership. The dominant property and casualty core platform vendor named this company its strategic partner for insurance decisioning, granted it premier technology partner status, lists an accelerator for it on its marketplace, and separately made a direct strategic investment in it.

That combination means the product reaches insurers through the system of record they already run, with integration built by the core vendor's own programme rather than assembled by each customer, and it means the party best placed to build this capability itself chose to partner and invest instead.

Around that sit further named integrations: a major risk data provider's claims product combined with this platform's decisioning, a global services firm partnering on detection and deterrence, and a services partner co building a national industry platform. Delivery runs on a named hyperscaler's artificial intelligence services. No public developer documentation was located, which is the one gap.

Cloud delivery on a named provider, with the placement questions unanswered. The platform runs on a major hyperscaler's artificial intelligence services, disclosed in that provider's own case study, and the vendor describes a software as a service platform. Beyond that, across two passes no region list, tenancy description, residency commitment or customer hosted option was located. The question is unusually live for this company.

It serves a global insurer across 15 countries and more than 115 customers across 25, several of which sit under regimes that constrain where insurance and health data may be processed, and the platform handles medical records through its billing and treatment agents.

A senior security role advertised by the company names the French health data hosting certification among the regimes it manages, which implies a residency capability exists for at least one sensitive category, but a recruitment advertisement is not a residency commitment and nothing published states what is available to whom.

No price, unit or tier is published, and two passes across the company's site, its product and resource pages, its news archive and third party directories produced nothing on how any product is charged. The published entry route is a demonstration request. Two structural questions follow from the shape of the offering and neither is answerable publicly.

The estate now spans detection, claims automation, underwriting and subrogation as separate product families, and nothing indicates whether they are licensed individually or as a suite, which matters because an insurer wanting only fraud detection is a very different buyer from one automating its whole claims operation.

And distribution through the core platform vendor's marketplace raises the question of whether an accelerator bought there prices differently from a direct agreement, with no published answer. One indirect signal exists: an independent review positions the platform as suited to medium and large insurers and suggests smaller teams may find simpler tools more practical.

Breadth on every dimension this axis measures, and evidenced rather than claimed. Institutionally, more than 115 insurance customers across 25 countries, spanning a global multinational insurer operating the platform across 15 of its own markets, a large European mutual, a national industry body deploying it as shared infrastructure, and a long established American motor mutual.

Geographically the company operates from France, the United States, Mexico, Brazil and Japan, covering Europe, North America, Latin America and Asia. By line of business it covers automotive, health, property and casualty and life, across both personal and commercial books.

By function it reaches the whole claims and underwriting lifecycle rather than one step, from application fraud before binding through first notice of loss, claims handling, payment integrity, subrogation recovery and compliance risk. The 100 percent insurance focus means none of that breadth is diluted by adjacent industries.