Clarify Health

Machine learning generates the outputs customers buy, on a business whose competitive moat is the dataset underneath.

The intelligence is real and specific. Predictions are the deliverable rather than a feature: outcomes forecasts, cost and utilisation projections, network performance assessments and referral pathway analysis, produced at a stated volume of more than 20 billion predictions across 300 million mapped lives. The most recent suite is described as using machine learning and natural language processing together to span cost, quality and utilisation in one application, and natural language processing implies unstructured content is being read rather than only claims fields aggregated.

What qualifies the grade is where the defensibility sits. The company's own lead claim is the industry's largest and most robust dataset, and much of what an analytics customer values here is coverage, linkage and benchmarking against a population that no competitor can assemble. A substantial share of that value would exist with conventional statistical methods applied to the same data, because the scarce asset is the data rather than the technique.

That is a different shape from the vendors graded higher in this index, where a model produces something that could not otherwise exist at all.

Graded B on the same basis as other platform vendors here: genuine inference doing consequential work, inside a product that is more than the inference. Ask what the models contribute over benchmarking, and what accuracy they achieve.

Analysis reaching human decision makers, with no calibration published for decisions that carry real consequence.

The operating model is sound in shape. Outputs are insights, benchmarks and predictions delivered through self service applications to analysts, executives and clinical leaders who decide. Nothing acts autonomously, and the self service framing means a human is always interrogating the analysis rather than receiving a verdict.

What gives this axis weight here is what the decisions do. Provider network optimisation determines which clinicians and facilities a health plan contracts with. Referral pathway analysis influences where patients are sent. Cost and utilisation assessment shapes which providers are flagged as outliers. These are consequential decisions about institutions and, through them, about patients' access, taken at population scale on the basis of model output.

No accuracy figure, confidence interval, operating point or validation description is published for any of it. A network model that ranks a provider poorly may be correct, or may be reflecting case mix it failed to adjust for, and nothing published lets a customer distinguish those or tells them how much uncertainty attaches to a ranking.

Risk adjustment is the specific unanswered question. Comparing provider cost and quality fairly requires adjusting for the patients each provider actually treats, and the methodology used is described only as precise.

Ask for the risk adjustment methodology, the confidence attached to provider rankings, and how customers are advised to act on marginal differences.

Scale and data sources are quantified; method and performance are not.

What is disclosed is genuinely informative about shape. The dataset is quantified at more than 300 million mapped lives with more than 20 billion predictions generated, and source categories are named rather than left vague, covering health plan claims, pharmacy data, government sources and negotiated pricing rates. Techniques are identified as predictive analytics, machine learning and natural language processing, and the architecture is described as modular stackable building blocks on a common platform. Leadership has also explained the approach by analogy to the statistical methods popularised in professional sports, which conveys the intent honestly.

What is absent is anything measurable. No accuracy, discrimination, calibration or error figure is published for any model, no model card exists, no validation methodology is described, and no operating point is given for predictions that drive network and referral decisions.

One phrase recurs and carries no content. Precise methodologies is used repeatedly to describe the analytic approach, and precision is a claim that ordinarily arrives with a number. None accompanies it.

The prediction count is a volume measure rather than a quality one. Twenty billion predictions says how much the platform produces and nothing about how often it is right.

One pre emptive note: further scale figures cannot move this grade. Only published model performance and a validation methodology will.

Ask for accuracy and calibration on the flagship prediction types, and the validation approach behind the precision claim.

For a company whose principal asset is licensed data, the data supply chain is the supply chain, and it is undisclosed.

Source categories are named at a general level, covering health plan claims, pharmacy data, government sources and negotiated pricing rates. No supplier, aggregator, clearinghouse or data partner is identified, no licensing terms are described, and nothing states which government programmes contribute or under what agreements.

That matters more than the usual infrastructure question because the dataset is the differentiator. A buyer relying on insight derived from 300 million linked lives has no way to assess whether the underlying licences permit the use being made, whether contributor agreements restrict downstream application, or whether the data supply is durable. Claims data licensing in this market runs through a small number of aggregators and pricing transparency sources, and dependency on any of them is a business risk that customers inherit without visibility.

Coverage composition is equally undisclosed and shapes every output. Which payers, geographies and lines of business are represented determines where the models are strong and where they are blind, and no breakdown is published.

The conventional layers are also absent: no cloud or hosting provider, no sub processor register, no machine learning framework, and no statement on whether customer contributed data enters the shared corpus.

Ask for the data sources by supplier and licence, coverage composition by payer and geography, and whether customer data enters the shared dataset.

Consistent independent customer recognition, and no measured outcome behind any of it.

The third party assessment is genuine and comes from two distinct sources. A major healthcare analyst firm ranked the company first for data analytics platforms serving payers, and a large software marketplace named the platform top in its relationship index for healthcare analytics in Winter 2025, alongside high performer and momentum leader placements. Both derive from customer reviews rather than vendor submissions, so they measure delivered satisfaction, and being rated first on quality of support, ease of doing business and likelihood to recommend tells a prospective buyer something real about the experience of working with this company.

What those rankings do not measure is whether the predictions are accurate or whether using them improves anything.

No peer reviewed publication was located. No customer is named anywhere in public material, which is unusual for a company of this size and funding. No quantified outcome exists at any site, and no accuracy or validation figure is published for any model.

The headline claim deserves scrutiny rather than acceptance. A figure of 2.6 billion dollars in customer verified return appears throughout company material. Customer verified is an unusual construction implying an assurance process, and no such process is described, no methodology is published, no time period is stated and no breakdown by customer or use case is offered. A number that large with that little behind it should not be treated as evidence.

Ask for named references, published model accuracy, and the methodology behind the return figure.

A dedicated pass located no encryption statement, no retention schedule, no access control model, no deletion process, no re identification risk assessment and no data ownership position.

The scale is what makes this the record's central gap rather than a routine omission. Longitudinal journeys across 300 million lives, built by linking claims, pharmacy, government and pricing sources, constitutes one of the largest assembled healthcare datasets held by any vendor in this index. It is also the company's principal asset, which means the incentives around retaining and reusing it run in one direction only.

The linkage itself deserves published treatment and receives none. Constructing a patient journey requires matching records across sources to the same individual, and that matching is exactly the operation that raises re identification concern, because a linked longitudinal record is far more distinctive than any single claim. No description of the linkage method, the identifiers used, or the safeguards applied was located.

The multi tenant commercial structure adds a second unaddressed question. The same dataset supports competing health plans and pharmaceutical manufacturers simultaneously, and nothing states what one customer's contributed data does for another customer's insights, or whether contributors can restrict downstream use.

Nothing addresses whether the models are trained on customer contributed data or on the licensed corpus.

One pre emptive note: further analyst rankings cannot move this grade. Only a published account of dataset governance, linkage safeguards and retention will.

Ask how the dataset is governed, how linkage manages re identification risk, and what contributors can restrict.

A dedicated pass located no health privacy position of any kind, and the dataset makes that the most consequential omission in this record.

The platform is built on longitudinal patient journeys spanning more than 300 million unique lives, assembled from health plan claims, pharmacy data, government sources and negotiated pricing rates. That is close to the entire insured population of the United States, held as linked longitudinal records rather than aggregate statistics.

Nothing published explains how that dataset exists lawfully. Whether the records are de identified under a recognised standard, whether a limited data set arrangement applies, what data use agreements govern the government sourced component, whether health plan contributors permit secondary commercial use, and how re identification risk is managed across a linkage of that scale are all unaddressed. Linkage is the crux: joining claims, pharmacy and pricing data into a single patient journey increases re identification risk precisely because the combination is more distinctive than any source alone.

No business associate agreement template, execution requirement or subcontractor position was located either.

One question follows for any customer. A health plan contributing its own membership data to a platform that also serves competing plans and pharmaceutical manufacturers needs to know what its data becomes and who benefits from it, and no published material addresses that.

Ask for the legal basis of the dataset, the de identification standard applied, what contributors permit, and the agreement template.

A dedicated pass located no security page, no external attestation, no trust centre, no penetration testing statement and no vulnerability disclosure policy. No security credential of any kind was found.

The absence is difficult to reconcile with the customer base. Health plans conducting vendor security reviews before contributing membership data, and pharmaceutical companies with their own vendor assurance regimes, will have examined this company's controls repeatedly. A posture therefore exists and is simply not published, which leaves a prospective buyer unable to begin diligence without a sales conversation and is a straightforward gap between what the company can evidently satisfy and what it shows.

The asset makes the omission serious rather than procedural. A linked longitudinal dataset covering close to the entire insured United States population is among the highest value healthcare data concentrations in private hands, and the attack surface question is not whether it is targeted but how it is defended.

The analyst and marketplace recognitions do not substitute. Both measure customer satisfaction and support quality, neither examines security controls, and a top relationship ranking says nothing about encryption, access management or incident response.

One pre emptive note: further awards or customer recognition cannot move this grade. Only an external attestation, or security documentation available under agreement, will.

Ask whether an information security attestation is held, what documentation is available under agreement, and how access to the core dataset is controlled internally.

Device regulation is not the applicable regime, and two regimes that do apply go unaddressed.

Enterprise analytics delivered to business and clinical leadership sits outside device regulation. The output informs contracting, network design, population management and commercial strategy rather than diagnosing or treating an individual, and no clearance is claimed or evidently required. That much is a clean position.

The regimes that do apply concern data rather than devices. A platform assembling claims, pharmacy and government sourced records across 300 million lives operates under health privacy law's de identification and limited data set provisions, under data use agreements governing any federal data component, and increasingly under state health data and consumer privacy statutes that treat linked health records distinctly. None of that posture is described anywhere.

A second regime is emerging directly around this product's function. Payer use of algorithms to influence coverage, network participation and utilisation decisions is attracting regulatory and legislative attention, with transparency and fairness requirements taking shape. A vendor supplying network optimisation and utilisation analytics to health plans sits squarely in that path, and nothing addresses it.

One further consideration follows from the life sciences business. Analytics supporting pharmaceutical commercialisation touches promotional and market access activity that carries its own compliance framework.

Ask for the de identification standard and data use agreement position, and the compliance posture on emerging payer algorithm oversight.

A dedicated pass located no fairness testing, no subgroup performance, no calibration data, no drift monitoring, no governance framework and no external audit.

The bias exposure here is structural and is arguably the most consequential in this lane, because the outputs shape institutional decisions rather than individual ones. Models trained on historical claims learn the patterns that claims encode, and claims encode access, coverage and referral habits as much as clinical need. A network optimisation model that identifies high performing providers from cost and utilisation history will tend to favour those serving healthier, better insured populations, and to penalise safety net providers whose patients arrive sicker and cost more for reasons outside the provider's control.

The consequence is different in kind from a clinical model's error. A provider ranked poorly may lose network participation, which removes access for the community that provider serves. The harm lands on patients who were never scored.

Risk adjustment is the mechanism that would mitigate this and its adequacy is unexamined publicly. The company describes precise methodologies without publishing what is adjusted for, whether social and economic factors are included, or how performance holds across provider types and patient populations.

The dataset composition compounds it. A corpus assembled from commercial claims and pharmacy sources will represent insured populations unevenly, and coverage gaps become model blind spots.

No accountable owner for model behaviour is named.

Ask what the risk adjustment accounts for, how provider rankings perform across safety net and academic settings, and how dataset coverage varies by payer type and geography.

Nothing allocates responsibility, alongside one of the largest unsupported value claims in this index.

A dedicated pass located no service level agreement, no accuracy warranty, no performance guarantee, no indemnity and no remediation position.

The asymmetry is the finding. Company material prominently claims 2.6 billion dollars in customer verified return, a figure repeated across releases with no methodology, verification process, time period or breakdown disclosed. A vendor willing to publish a benefit number of that magnitude publishes nothing about what happens when the analysis is wrong.

The harm profile is institutional and slow rather than clinical and immediate, which makes recourse harder to pursue and easier to omit. A health plan that excludes a provider from its network on the strength of a cost ranking that failed to adjust properly for case mix has taken a decision affecting that provider's viability and its patients' access, and the error may never be attributed to the analysis. A health system that restructures referral pathways on a flawed prediction absorbs the cost quietly.

No model accuracy figure is published, so a customer cannot characterise the residual risk it accepts when acting on a ranking.

One pre emptive note: further rankings, awards or return claims cannot move this grade. Only contractual terms, or published model performance with a stated error profile, will.

Ask what the agreement warrants on analytic accuracy, the methodology behind the return claim, and what recourse exists where a decision rests on a flawed prediction.

A dedicated pass located no electronic health record integration, no interface standard, no marketplace or validated listing, no published application programming interface and no named source or destination system.

The inbound picture is characterised by category rather than by system. Data arrives from health plans, pharmacies, government sources and pricing datasets, which describes what kind of data is used and names none of the systems or suppliers it comes from.

The outbound direction is the more consequential gap and is entirely undescribed. Insights are delivered through cloud based self service applications, which means the platform is a destination rather than something that reaches into the systems where work happens. Nothing states whether a referral pathway insight can surface in a scheduling or record system, whether a network finding can flow into contracting workflow, or whether a population health opportunity reaches care management tooling.

That matters because the company's own framing is actionability. Insight delivered to an analyst who must then persuade an operational team to act is a different product from insight delivered into the operational system itself, and only the former is evidenced.

A product supporting referral network optimisation would benefit particularly from reaching the point where referrals are actually made.

Graded D consistent with the treatment of an absent named integration across this index. Ask which systems are integrated inbound and outbound, through what standards, and whether insights reach operational workflow.

A dedicated pass located no hosting provider, no region, no residency commitment, no tenancy model and no continuity position.

The only architectural statement is that business applications are cloud based, which establishes a hosted model without describing it.

The tenancy question is unusually pointed for this vendor and unaddressed. The platform holds a single linked dataset spanning 300 million lives and serves competing health plans, competing health systems and pharmaceutical manufacturers from it simultaneously. Those customers are commercial rivals whose contributed data, queries and derived insights are sensitive to one another, and how tenants are separated, what one customer's activity reveals to the platform, and whether derived analytics flow across the tenant boundary are all questions a security review would raise immediately.

Residency matters less here than for a clinical monitoring product and is still unstated, and any government sourced data component ordinarily carries its own location and handling requirements.

Continuity is a lower order concern given analytics workloads are not real time, and no availability position is published even so.

One question follows from the scale. Whether customers can obtain their own data or derived models on exit, and what happens to contributed data if a contract ends, is a practical matter for any organisation contributing membership data and is nowhere addressed.

Ask which provider hosts the platform and where, how tenants are segregated, what governs government sourced data, and what happens to contributed data at contract end.

Cost is absent from every published surface. A dedicated pass located no pricing page, no unit of charge, no range, no tiering, no implementation fee position and no minimum commitment.

One piece of structural disclosure exists and is useful without being a price. The offering is described as modular stackable building blocks that customers combine to fit specific goals, which tells a buyer that scope is composable and that the entry point can be narrow. It also implies that cost grows with the number of modules adopted, and nothing indicates how.

The unit question is otherwise wide open. A platform serving health systems, physician groups, ambulatory surgical centres, health plans, technology companies and pharmaceutical manufacturers cannot plausibly price all of them the same way, and charging could follow covered lives, attributed patients, modules licensed, seats, data volume or an enterprise licence. A pharmaceutical customer buying commercialisation insight and a health plan buying network optimisation are different businesses buying from the same platform.

The return argument is published without its cost side, which is the specific asymmetry here. A claimed 2.6 billion dollars of customer return is prominent, and no figure anywhere lets a prospective customer estimate what it would pay to pursue a share of it.

Ask for the unit of charge by customer type, how modules price incrementally, the data onboarding cost, and the contract term.

Coverage spans the whole commercial ecosystem rather than one segment, which is unusual and is enabled by the underlying data asset.

Buyer types named include health systems, individual physician groups, ambulatory surgical centres, health plans, technology and services organisations, and life sciences companies. Those are genuinely different customers with different questions: a health system asks about referral leakage and network position, a payer asks about provider cost variation and network design, and a pharmaceutical manufacturer asks about treatment pathways and market access. Serving all three from one platform is possible only because a longitudinal claims dataset spanning 300 million lives answers all of their questions from the same underlying records.

Use case coverage is correspondingly wide and enumerated rather than gestured at, covering growth, network optimisation, care delivery improvement, population health management, value based care performance, cost containment, quality assessment, referral pathway analysis and therapy commercialisation.

The population scale is itself a coverage claim of a kind, since 300 million lives approaches the entire insured United States population and means most customers will find their own patients represented.

What is not evidenced is depth. No customer is named, nothing states how the base divides across those buyer types, and nothing indicates which modules carry real adoption against which are available. A platform this broad will have concentration somewhere, and the concentration is undisclosed.

Ask how the customer base divides by segment, and which modules carry the deepest adoption.