Guardant Health

The company markets an artificial intelligence layer called the Smart Platform and a branded capability called InfinityAI, and machine learning genuinely drives the signal detection and classification that lets a blood test call cancer from circulating tumour DNA. The epigenomic integration in the 2026 panel is described as delivering a several fold sensitivity increase, which is a model driven gain rather than a chemistry one.

What holds this at B is a disclosure the company makes itself. Its own product footnote states that being powered by InfinityAI refers to technology used in the DEVELOPMENT of product features, and that those features are reported as professional service and are not approved by the regulator. That is a meaningfully different claim from artificial intelligence operating at the point of result. The assets a buyer actually engages are a certified laboratory, a proprietary panel and an approved assay, and the approvals attach to the assay rather than to any model.

Low autonomy, and here that is the correct design rather than a shortfall. The product is an ordered diagnostic test that returns a report to a clinician, who decides what follows. There is no autonomous action on a patient and no workflow the system completes on its own.

The oversight question that does bite is what happens after a result. The company states that a positive Shield result should be followed by colonoscopy and that Medicare covers that follow up, which is a published care pathway rather than an implied one. It also states plainly that the test can produce false positives and false negatives and that a negative result does not rule out disease. Naming both error directions in patient facing material is better candour than most screening products offer.

The assay is documented to a standard almost nothing else in this index reaches, because regulatory approval forces it. Panel composition, genomic footprint, the shift to combined genomic and epigenomic profiling, and performance characteristics all pass through a public approval process, and the laboratory developed test is separately and explicitly labelled as not approved.

The artificial intelligence layer is the opaque part. No architecture, training data, validation methodology or standalone performance is published for InfinityAI or the Smart Platform, and the company's own footnote says the capability is not a standalone product. So the buyer can inspect the test in detail and cannot inspect the model at all. Graded B because the approved half is genuinely transparent, with the split recorded rather than averaged away.

Graded on an honest basis and flagged for re verification. The company operates certified clinical laboratories and handles protected health information under United States health privacy law, and no specific published stewardship position was located in this pass on specimen retention, secondary research use of genomic data, or whether patient data contributes to model development, nor was any hosting arrangement or sub processor list found.

That last question matters more here than in a typical record and the reason is the nature of the material. Genomic sequence is identifying in a way that does not expire and cannot be revoked, and it implicates biological relatives who never consented to anything, so a permission granted once by one patient reaches people who were never asked.

The company states that its platform compounds insight from more than a million patients tested, which is a description of a data asset built from patients, and what those patients were told about that, and what they can decline, is the thing to establish. Absence of a retrieved statement is not proof none exists, and this grade should be re verified against the published privacy notice and the consent forms before it is quoted. Ask for specimen and sequence retention, the secondary research position, whether test data contributes to model development, and what the consent form covers.

Among the strongest evidence positions in the index, and the reason is that the validation comes from bodies with no commercial interest in the result. Shield was approved as the first blood test for primary colorectal cancer screening on the strength of a prospective trial published in the New England Journal of Medicine, and in 2026 it was added to the American Cancer Society colorectal cancer screening guidelines. An independent guideline body adopting a test is a different order of evidence from a vendor publishing its own accuracy.

On the therapy selection side, each of the 25 companion diagnostic indications is tethered to a therapeutic trial, including the phase 3 BREAKWATER trial supporting the BRAF V600E colorectal indication. Guardant Reveal holds Medicare coverage for recurrence monitoring in colorectal cancer through the molecular diagnostics services programme, which required its own evidence review. More than one million tests have been performed.

Graded on an honest basis rather than on retrieved detail, and flagged for re verification. The company operates certified clinical laboratories and handles protected health information under United States health privacy law, but no specific published stewardship position was located in this pass on specimen retention, secondary research use of genomic data, or whether patient data contributes to model development.

That last question matters more here than in a typical record. Genomic sequence is identifying in a way that does not expire and cannot be revoked, it implicates biological relatives who never consented to anything, and the company states that its platform compounds insight from more than one million patients tested, which is a description of a data asset built from patients. What patients are told about that, and what they can decline, is the thing to establish. Re verify against the published privacy notice and consent forms before this grade is quoted.

The usual framing does not map cleanly onto this vendor and the note says so rather than forcing it. A clinical laboratory ordering and reporting its own tests is generally a covered entity in its own right, not a business associate of the ordering practice, so the business associate agreement that this axis normally turns on is not the instrument that governs the relationship. The obligations are direct rather than contracted downward.

No published compliance statement, notice of privacy practices summary or agreement posture was located in this pass. Graded C on that basis and flagged for re verification. A buyer should be asking about data use permissions in the test requisition and consent, which is where the real terms sit for a laboratory, rather than looking for a vendor agreement that may not exist.

A dedicated search for a trust centre, SOC 2 report, HITRUST certification or ISO 27001 certification returned nothing attributable to this company. For an organisation of this size holding genomic data on more than a million people, the absence of a discoverable published security posture is notable in itself.

One avenue was not exhausted and should be before this grade is relied on. As a company listed on a United States exchange it is required to disclose its cybersecurity risk management, strategy and governance in its annual filing, and that filing is public and authoritative. Regulatory filings beat marketing pages on this kind of question. Flagged for re verification against the annual report rather than left at a low grade on a failed marketing search, because low security grades assigned that way have a poor track record in this index.

The deepest regulatory position in this category and the record separates its parts, because the company sells products on three different routes at once and buyers conflate them.

Approved: Guardant360 CDx, the first liquid biopsy approved for comprehensive genomic profiling, carrying 25 companion diagnostic indications; Guardant360 Liquid CDx, approved May 2026 with the earlier seven indications transferred; Shield, approved July 2024 as the first blood test for primary colorectal cancer screening, with a higher throughput workflow approved in 2026. Not approved: Guardant360 Tissue, a laboratory developed test run under clinical laboratory certification, which the company labels as such. Also not approved: the InfinityAI features, which the company states are reported as professional service.

That last carve out is the finding. The artificial intelligence sits deliberately outside the approved device, so the regulator has reviewed the assay and not the model layered on top of it.

No subgroup performance, model monitoring policy or bias evaluation appears in the marketing or product material reviewed. The approval process for a screening test does require demographic performance data in the public labelling summary, so evidence of this kind very likely exists in documents not retrieved in this pass, which is why the grade is C rather than lower. Re verify against the summary of safety and effectiveness.

The substantive concern is specific to population screening. Colorectal cancer incidence, stage at presentation and outcomes differ materially across groups in the United States, and a blood test that is more or less sensitive in one group changes who gets sent for a colonoscopy and who is reassured. A test adopted into a national screening guideline is applied to entire populations rather than to selected patients, which raises the cost of any performance gap that is not published.

The buyer can inspect the test in detail and cannot inspect the model at all, and the record keeps that split rather than averaging it away. On the assay side the documentation reaches a standard almost nothing else in this index does, because regulatory approval forces it: panel composition, genomic footprint, the shift to combined genomic and epigenomic profiling, and performance characteristics all pass through a public approval process.

One further practice deserves specific credit as scope discipline. The laboratory developed test is separately and explicitly labelled as not approved, so a clinician ordering it knows which regulatory status attaches to which product rather than inheriting the approved product's credibility across the whole catalogue. Vendors in this index frequently let a single clearance colour an entire portfolio; distinguishing them in your own material is the opposite behaviour.

The artificial intelligence layer has none of that. No architecture, training data, validation methodology or standalone performance is published for the analytical capabilities, and the company's own footnote states the capability is not a standalone product, which acknowledges the position without filling it. No warranty, indemnity or remediation commitment was located. Ask what the analytical layer contributes to a reported result, whether a clinician can see a result without it, and for standalone performance if it influences reporting at all.

Graded on an honest basis. The delivery model is a send out laboratory test with results returned through an ordering portal, so the integration question is about electronic ordering and structured result return into the record rather than about an application running inside the health record.

No specific named electronic health record integration, structured result interface or standards level detail was located in this pass. For a test used at this volume and reimbursed at this scale, ordering integration almost certainly exists in some form, so this is a gap in what was retrieved rather than an established absence. Flagged for re verification.

The concept translates awkwardly and the note records that rather than inventing a fit. Nothing is deployed at the customer. Specimens are shipped to the company's own certified laboratories, sequencing and analysis run there, and a report comes back, so the data resides with the vendor by design and there is no on premise, private cloud or customer controlled option available to want.

The company's clinical laboratory is identified in its own product material as being in Redwood City, California, which places the analysis and the resulting genomic data in the United States. No published retention schedule for specimens, sequence data or derived results was located, and for genomic material that is the substantive residency question rather than server location.

The strongest commercial disclosure in the index, and as with Artrya the reason is structural: a listed company selling a reimbursed test has to publish what a private vendor can withhold.

The actual numbers are public. Shield reimbursement moved from 920 dollars to 1,495 dollars per test on 1 April 2025 under Advanced Diagnostic Laboratory Test status. The pricing MECHANISM is published too, which is rarer than the number: during a nine month period the company collects private payer rates, and the median of those sets the Medicare payment from 1 January 2026 to 31 December 2027. Patient exposure is stated rather than left to be discovered, with zero out of pocket for eligible Medicare fee for service patients and under 50 dollars for a majority of insured patients including those qualifying for assistance, and coverage named for Veterans Health Administration community care and TRICARE.

What is not published is a list price for the therapy selection tests, which reach buyers through payer contracts rather than a rate card.

Deep rather than broad. Everything sits inside oncology, but within it the coverage runs the length of the disease: screening an average risk population who do not have cancer, monitoring for recurrence after curative intent treatment, and selecting therapy in advanced solid tumours across tumour types.

Three quite different buyers follow from that. Primary care and screening programmes order Shield, oncologists order Guardant360 and Reveal, and biopharmaceutical companies buy into the platform for drug development and trial work. Held at B rather than A because it is one specialty, however completely covered, and because the screening indication that carries the guideline endorsement is currently colorectal cancer alone.