Tempus

Graded on the indexed products only. Next is an AI clinical decision support system, Hub runs on an agentic architecture, and David is a generative AI clinical assistant. AI is the mechanism of all three. The broader sequencing and pharma services business is out of scope for this record.

Hub's prior authorization agent gathers guidelines, payer policy, and patient information and outputs support documents for further use by the care team, a human in the loop posture. Northwestern's David deployment includes real time monitoring of deployed algorithms and agents. Oversight is described but not formalized in a published governance framework.

Architecture is described at a marketing level: agentic workflow platform (Agent Builder), multimodal patient record, generative AI assistant. No model cards, named models, or technical documentation retrieved.

This is the most candid record in the index and one of the least favourable, and both halves belong in the same sentence. Public reporting obligations force this business to describe its data practices more plainly than a private peer would, and it does: the company states that the success of its business depends on continued access to, and the ability to monetise, de identified patient data, and that it relies on the statutory rights available in both its covered entity and business associate capacities to do so.

The de identification method is described with some specificity, covering removal, modification or masking of direct identifiers. A vendor that tells you it will monetise data derived from your patients has told you the truth, and this index should credit the telling while grading the substance. The substance is an unbounded permission.

De identified data may be shared with the company's own researchers and developers and with third parties inside or outside the United States, and once de identified it falls outside the privacy notice entirely and may be used for any lawful purpose. The data is also substantially genomic, which is the category where de identification is most contested, because a genome retains identifying structure that identifier stripping does not remove.

Consolidated class action litigation is pending concerning genetic data obtained through an acquisition, with allegations of use for model development without consent required by state genetic privacy law; those allegations are undetermined and are recorded because the practice is central to the business model. Establish exactly which permissions your agreement grants over derived data.

Vendor reports a 662 patient prospective study across six sites for Next, with six new indications added in May 2026. Prospective multi site evidence is rare in this index; held back from A because the publication venue and peer review status of the study were not verified.

The disclosure here is unusually full, and the grade reflects what it discloses rather than any gap. Public company reporting obligations force this business to describe its data practices more plainly than a private peer would, and the description is candid: the company states that the success of its business depends on continued access to, and the ability to monetise, de identified patient data, and that it relies on the statutory rights available to it in both its covered entity and business associate capacities to do so.

The de identification method is described with some specificity, covering the removal, modification or masking of direct identifiers including names, specific geographic locations, social security numbers and full dates of birth, and the company states the process is designed to meet and exceed the applicable standard.

Two published terms carry the grade down. De identified data may be used and shared with the company's own researchers and developers and with third parties inside or outside the United States, and once de identified it falls outside the privacy notice entirely and may be used for any lawful purpose. That is an unbounded secondary use permission. And the data in question is substantially genomic, which is the category where de identification is most contested, because a genome retains identifying structure that ordinary identifier stripping does not remove.

The company is also currently defending consolidated class action litigation in the Northern District of Illinois concerning genetic data obtained through an acquisition, in which plaintiffs allege that data was used for artificial intelligence model development and disclosed to third parties without the consent required by state genetic privacy law. Those are allegations that have not been determined, and they are recorded here because the practice they concern is central to the business model rather than incidental to it. Buyers and contributing institutions should establish exactly which permissions their own agreements grant over derived and de identified data.

This is the most complete HIPAA disclosure in the category, and it is complete because Tempus occupies both regulated roles and explains which applies when.

The software terms of use state that Tempus is a covered entity, define the term against the regulation itself, and state that Tempus operates in that capacity when a clinician orders tests and receives results. Separately, the same terms carry business associate agreement provisions that apply automatically where Tempus processes protected health information outside its covered entity role, unless a separate agreement has been executed. So the instrument is published rather than merely named, and it is published alongside an explanation of when each role governs. The company also publishes a Notice of Privacy Practices, which is a covered entity obligation rather than a marketing document, and it names the affiliates it covers.

Securities filings set out the same dual position in operational terms: data arrives in a covered entity capacity when the company performs sequencing on behalf of a patient, and in a business associate capacity when it provides other services to a provider, in which case it relies on those providers to have obtained the necessary patient permissions.

One caution belongs on this axis without changing the grade. HIPAA is not the whole of health privacy law, and a strong HIPAA position does not settle every question about the same data. State genetic privacy statutes impose written consent requirements that HIPAA does not, and the company is currently defending consolidated class action claims under one of them. Buyers should establish which agreements govern their own data flows and should not treat HIPAA coverage as answering state law questions.

No trust centre, security page or third party attestation specific to this company was located, and no SOC 2, HITRUST or ISO 27001 claim was retrieved from its own material.

This is a statement of what was found rather than a conclusion that none exists, and the scope limitation is worth being explicit about: the company name collides closely with several unrelated technology vendors and audit firms, which pollutes search results for this particular question. It should be retested directly against the company's own domain on refresh.

The absence is nonetheless notable at this scale. This is a publicly listed company operating clinical laboratories, holding genomic and clinical records at population scale, and selling into health systems and pharmaceutical companies whose vendor review processes normally require an attestation before contracting. Comparable disclosure would be expected and was not found.

Buyers should ask which attestations are held and at what scope, and should ask specifically whether the assessment boundary covers the laboratory operations, the data platform and the acquired subsidiaries, since this company has grown substantially by acquisition and a certification obtained for one entity does not extend to another.

The company operates its own clinical laboratories, and its published terms use laboratory certification and accreditation as the governing standard for whether a given output may be used clinically. That is the laboratory developed test route rather than the device clearance route, and it is a legitimate and distinct regulatory position.

What earns the grade is that the boundary between clinical and non clinical output is drawn explicitly and contractually rather than left to inference. The terms state that research use only data is not provided or validated for clinical use under the laboratory certification and accreditation requirements, is not intended for use in diagnosing or treating patients or otherwise informing their care, and is provided as is without warranty. A customer therefore knows which outputs carry laboratory validation behind them and which do not. That is the same structural disclosure the index has found valuable elsewhere, where a vendor states plainly that a given configuration of its product is not the validated one.

Held at B rather than A because the position was assessed at the level of the contractual framework, not per product. This company sells many distinct assays and applications and has acquired several businesses with their own regulatory histories, and the specific clearance, approval or authorisation status of individual products was not enumerated in this pass. Buyers should ask for the regulatory status of each specific test or application they intend to use, and should not assume it transfers across the portfolio or across acquired entities.

No responsible artificial intelligence statement, bias assessment, subgroup performance analysis or model governance framework was located.

The absence is more conspicuous here than for most vendors, for two reasons. The company's central asset is a large multimodal clinical and molecular dataset used to develop and train models, so questions about who is represented in that dataset bear directly on how its models behave. And the company is publicly defending litigation about how part of that dataset was assembled, which places the governance of the data supply itself in issue.

There is also a specific and underappreciated tension worth putting to the vendor. The published de identification process removes or masks specific geographic locations. That is correct practice for protecting identity, but geography is one of the strongest available proxies for population composition, so the same step that protects the individual also removes the information needed to check whether a cohort represents the patients a model will be applied to. A company operating at this scale is better placed than almost anyone to resolve that tension, by publishing population composition and subgroup performance at an aggregate level that does not reintroduce identifiability.

Buyers should ask how the training populations are composed, how model performance varies across them, and what governance applies to the development of models built on data acquired through acquisition.

Two passes located no accuracy figure, no evaluation methodology, no model card, no published limitations and no warranty, indemnity or remediation commitment for any capability. Architecture is described at a marketing level, covering an agentic workflow builder, a multimodal patient record and a generative assistant, with no model named and no technical documentation retrieved. The contrast within the same company is what makes this notable.

On the data side, reporting obligations produce candid and specific disclosure of practices that are commercially sensitive and unflattering. On the artificial intelligence side, where no equivalent obligation exists, the disclosure stops at product names. That is the same pattern this index has recorded elsewhere: an organisation demonstrably capable of rigorous public reporting applies it where an external requirement exists and not otherwise.

The consequence side deserves naming because the outputs reach oncology decisions. A capability that surfaces therapy options or trial matches for a cancer patient produces suggestions a clinician acts on under time pressure, and no published error characteristic exists for any of them, nor any statement of which capabilities are decision support and which are informational. Ask which models underlie each capability, for accuracy on the specific outputs you would act on, for the limitations documentation, and for what the vendor commits to when a recommendation is wrong.

David is deployed directly into the EHR at Northwestern Medicine, underpinned by a multimodal patient record; Next delivers near real time alerts to clinicians in workflow. Depth demonstrated at a flagship site; breadth of EHR coverage beyond that deployment not documented.

There is no deployment model to assess in the ordinary sense, and that is itself the finding. The primary delivery mechanism is a laboratory service: a specimen is sent to the company's laboratories, and results return to the clinician through a software application used for ordering tests and receiving reports. The customer does not host anything, and the data lands in the vendor's environment by design rather than by configuration.

No hosting arrangement, cloud provider, infrastructure description, implementation timeline or uptime commitment was located.

On residency the published position runs the other way from a commitment. The privacy notice states that de identified data may be shared with third parties inside or outside the United States, which tells a buyer that derived data is expected to cross borders rather than that it will be confined to a region. No jurisdictional guarantee of any kind was found, and securities filings describe expansion into additional geographies as a growth objective, so the question is likely to become more rather than less relevant.

Buyers with residency obligations, and in particular institutions operating under state or non United States data protection regimes, should establish in writing where identifiable data is processed and stored, and separately where derived and de identified data may travel.

Public listing forces a level of commercial disclosure that no private vendor in this category matches, and the index has previously credited exactly this effect where a listing obliges a company to publish what it would otherwise keep private.

What a buyer can therefore see is substantial. Segment structure separates the testing business from the data and applications business, so the relative weight of each is visible. Revenue, margin and customer concentration are reported. And the risk disclosures state the commercial dependency plainly: the business depends on continued access to, and the ability to monetise, de identified patient data, and on obtaining the permissions necessary to use that data commercially. A prospective customer or data contributing institution can read, in the company's own regulatory filings, why its data is commercially valuable to the counterparty. Very few vendors in healthcare put that in writing anywhere.

Held at B rather than A because none of this is a price. No list price, fee schedule or pricing mechanism was located for any individual test, for the software, or for data licensing, and the pricing basis for data and applications agreements is not described even in outline. Financial statement transparency tells a buyer how the company earns money in aggregate; it does not tell them what they will pay.

Buyers should ask for the pricing mechanism for each component separately, and institutions considering contributing data should ask how that arrangement is valued, since the filings establish that the data has independent commercial worth to the counterparty.

Coverage is genuinely broad and it has been assembled deliberately. The company's origin and centre of gravity is oncology, and it has extended through acquisition into hereditary and germline genetic testing, pharmacogenomics and digital pathology, with securities filings describing expansion into additional disease areas and geographies as an explicit growth objective. That is a wider span of the diagnostic pathway than any other vendor assessed in this category.

Held at B on two grounds. Breadth acquired is not the same as breadth integrated, and nothing published describes how these businesses relate to one another operationally, whether the datasets are combined, or whether a customer contracting for one gains anything from the others. Second, the coverage described is of test menu and disease area rather than of care setting: nothing distinguishes how these services work for an academic cancer centre, a community oncology practice, a rural hospital or an international customer, and access to specialist genomic testing varies enormously across exactly those settings.

Buyers should ask which entity within the group will actually be their counterparty, and should ask for evidence of performance in a setting resembling their own rather than in the flagship institutions most likely to be cited.