Galen Health

The product is AI agents, described by the company as teammates, performing the work themselves rather than presenting someone else's output. Case preparation, trial identification, live retrieval during a meeting and maintenance of a decision knowledge base are all generative and retrieval tasks, and there is no underlying platform, system of record or services business the AI is attached to. Graded on what the product is; note that the rest of this record is thin and a later refresh may revise this if the company turns out to be assembling third party components.

No oversight mechanism was located in a second pass: no confidence signal, no threshold, no verification step, no statement of what a clinician must check. The reasoning in the earlier assessment is the right frame and the second pass adds something that sharpens it.

That reasoning first, because it is the important part. A tumour board is where a cancer treatment recommendation is agreed by a group. Content entering a group decision carries a risk that content read by one clinician does not: no individual owns verifying it, and an error can be absorbed into a consensus nobody separately checked. Diffusion of responsibility is a well described property of group decisions, and it is precisely what a board is otherwise designed to counteract.

The vendor's own framing works against that. The product is marketed as an AI teammate, a digital teammate that is part of your team, engaged with rather than operated. That language is deliberate and it is not harmless in this setting. A teammate contributes to a consensus; a tool produces output someone checks. Framing the system as the former invites a board to treat its contributions the way it treats a colleague's, which is with the presumption of a professional standing behind them. Nobody is standing behind these.

So the questions are concrete. What is the board expected to verify before relying on a generated case presentation. Who is named as accountable for it. Is generated content visually distinguished from clinician entered content in the presentation. And what happens when a live retrieved insight is wrong during the discussion.

Nothing was located. No model or model family named, no accuracy figure of any kind, no methodology, no evaluation, no model card, and no description of how case preparations or trial matches are produced or verified.

The absence is unremarkable for a company at this stage and it sits awkwardly with what the product does, which is the point worth making rather than the grade.

Trial matching is a function where the field has established that disclosure is possible. An academic cancer centre has built and published a matching system that uses language models to interpret eligibility criteria against a full clinical record, describing its approach to unstructured note processing and patient summarisation. That work is in the open. A commercial vendor performing the same function can reasonably be asked what it does differently and how it knows the result is right, and the existence of a published peer means the answer is not commercially impossible to give.

The verification question is the one that matters most and it is separate from accuracy figures. Ask what checks a generated case presentation against the underlying record, whether trial matches are traced to the eligibility criteria they were judged against, and what the system does when the record is incomplete on a criterion rather than clearly failing it. Silent treatment of missing data as absent disease is a known failure mode in matching and it produces confident wrong answers.

A young company will not have benchmarks. It should be able to describe its method and its verification step.

Ask which models are used, whether an external provider processes clinical content, and how matches are verified.

Nothing identifies any party in the chain: no model or model family, no hosting arrangement and no sub processor list was located, and no retention, training or de identification position was found. The feature that matters is larger than a records question.

The vendor describes creating a database of care decisions and maintaining a knowledge base of care discussions to learn from and optimise clinical decision making, and separately describes unifying oncology workflows across multiple hospitals with a single layer that standardises care. Read together those describe a cross institutional learning loop: decisions reached at one board accumulate, and the accumulated set informs what is surfaced elsewhere.

That is a coherent product idea and it is the thing to examine hardest, because tumour board decisions are not generalisable facts. They are consensus judgements made with local context, covering which trials are open here, what this surgical team can offer, what this patient wants and what this payer will cover. Standardising across institutions using accumulated decisions imports one institution's constraints into another's reasoning, and neither board sees that happening.

So the questions are ownership, separation and consent: who owns the accumulated store, whether it is segregated or pooled, whether each institution was told its deliberations would inform recommendations elsewhere, and what happens to a contribution at termination. The content is cancer diagnoses, prognoses and the reasoning behind treatment choices, including choices not to treat.

Claims exist and none of them is verifiable. The company states it is trusted by leading cancer centres without naming one, and asserts hours saved weekly and hospital cost reduction in the millions with no baseline, denominator, method or customer attached. No named deployment, no case study, no funding disclosure and no independent evaluation was located. Graded C rather than Not Rated because deployment and benefit are asserted; nothing supports either.

No retention period, training use statement or de identification posture was located. The feature the earlier assessment identified is confirmed and is larger than it first appeared.

The vendor describes creating a database of care decisions and maintaining a knowledge base of care discussions to learn from and optimise clinical decision making, and separately describes unifying oncology workflows across multiple hospitals and clinics with a single layer that standardises care.

Read together those are not a records feature. They describe a cross institutional learning loop: decisions reached at one board accumulate, and the accumulated set informs what is surfaced elsewhere. That is a coherent product idea and it is also the thing to examine hardest, because tumour board decisions are not generalisable facts. They are consensus judgements made with local context: which trials are open here, what this surgical team can offer, what this patient wants, what this payer will cover. Standardising across institutions using accumulated decisions imports one institution's constraints into another's reasoning, and neither board sees that happening.

So the questions are ownership, separation and consent. Who owns the accumulated store. Is it segregated per customer or pooled. If pooled, was each institution told that its deliberations would inform recommendations elsewhere, and did it agree. What happens to a customer's contribution when the relationship ends, and can it be withdrawn.

The underlying content is unusually sensitive: cancer diagnoses, prognoses, and the reasoning behind treatment choices, including choices not to treat.

Ask for the retention schedule, the separation model, and the training position in contract language.

The company describes itself as HIPAA ready rather than HIPAA compliant, and the distinction is worth holding onto. Ready is not a defined term and asserts capability rather than a state: it typically means the software could be operated compliantly given the right configuration and agreements, which places the obligation on the customer. Compliant, with a business associate agreement in place, is a commitment the vendor makes.

This index already tracks hedged compliance language such as HIPAA compliant infrastructure and SOC 2 aligned, and HIPAA ready belongs on the same watchlist.

Graded C rather than Not Rated because a claim was located and it is a weak one. Ask whether a BAA is offered and executed as standard.

No attestation and no trust centre were located in a second pass.

The published claim is a hedge and should be read as one. The platform is described as ready for the health privacy rule. Ready is not compliant, and neither is an attestation. It describes a state of preparedness the vendor has assessed itself as having reached, which is the weakest formulation in a field where this index already treats compliance claims as self assessment rather than evidence. There is in any case no certification for that rule.

The customer profile raises the expectation rather than lowering it. The vendor states it is trusted by leading cancer centres, and organisations of that kind run substantial security reviews before granting a vendor access to oncology records. If those reviews have been passed, evidence of them exists and would be far more useful to a prospective buyer than the current claim.

What an examination would need to cover is unusually broad for a young company, because of the accumulation feature. The platform maintains a store of care decisions across institutions, so it holds not only current case material but a growing corpus of clinical reasoning drawn from multiple organisations. Separation between customers is the control that matters most there, and it is commercial as well as regulatory: cancer centres compete for referrals and for trial placement.

Ask which report is held or scheduled, what its scope covers, whether any cancer centre's own assessment can be shared, and how tenant separation is implemented for the decisions store.

No clearance, device authorisation or pathway statement was located, and the scoping question the earlier assessment raised remains live and unaddressed by the vendor.

The capabilities that put it there are confirmed: preparing case presentations for a treatment decision forum, identifying clinical trials for individual patients, and supplying real time insights about a patient during the discussion in which the treatment decision is made.

Trial identification is the sharpest of the three, because a trial match is a statement that a specific patient may be eligible for a specific intervention. Presented into a board considering treatment options, it functions as an option on the list. A false positive sends a patient down an evaluation that ends in screen failure; a false negative means an option was never discussed, and nobody in the room knows it was omitted.

One disambiguation belongs on this record and is not a criticism of either company. A separate and unrelated business markets pathology products under a similar name which do carry regulatory clearances for cancer detection and grading. A buyer searching for this vendor will encounter those, and the two must not be conflated. Nothing located suggests this company claims or holds any clearance.

The published position is a compliance hedge rather than a statement: the platform is described as ready for the health privacy rule. Ready is not compliant and is not a regulatory posture.

Ask where the company believes the case preparation and trial matching functions fall, and to see the analysis written down.

No governance framework, fairness statement, subgroup performance disclosure or responsible AI documentation was located.

The observation the earlier assessment made is not a theoretical concern in this function, and the second pass confirms that the field itself treats it as a live problem. Trial matching is widely recognised as a place where access patterns produce uneven representation, and companies working in this exact space describe expanding access to representative and underserved populations as an explicit objective of their partnerships. So the equity question is one the segment has already named. A product surfacing trial options into a treatment discussion inherits it whether or not it addresses it.

The mechanism is worth stating because it is not obvious. A trial matching system reflects the trials it knows about and the criteria it can evaluate. Trials concentrate at academic centres, eligibility criteria encode exclusions that correlate with comorbidity and access, and a patient whose record is thinner because they have had less care will match less well. None of that requires bias in the model. It follows from the data and the criteria, and it produces the same result: patients who already have better access get more options surfaced.

That is measurable. Ask what proportion of patients receive at least one match, broken down by the characteristics the institution already tracks.

A peer benchmark exists: an academic cancer centre has published a trial matching system with described methodology. A vendor in this space can reasonably be asked for comparable transparency.

Two passes located no model named, no accuracy figure, no methodology, no evaluation and no description of how case preparations or trial matches are produced or verified, and no warranty, indemnity or remediation commitment. The stage explains part of this and the comparison explains why it still matters.

Trial matching is a function where the field has established that disclosure is possible: an academic cancer centre has built and published a matching system using language models to interpret eligibility criteria against a full clinical record, describing its approach to unstructured note processing and patient summarisation.

That work is in the open, so a commercial vendor performing the same function can reasonably be asked what it does differently and how it knows the result is right, and the existence of a published peer means the answer is not commercially impossible to give. The verification question matters more than accuracy figures here and is separate from them.

Ask what checks a generated case presentation against the underlying record, whether trial matches are traced to the eligibility criteria they were judged against, and what the system does when the record is incomplete on a criterion rather than clearly failing it. Silent treatment of missing data as absent disease is a known failure mode in matching and it produces confident wrong answers that look exactly like correct ones. A young company will not have benchmarks; it should be able to describe its method and its verification step.

No electronic health record vendor, integration mechanism, standard or data source is named anywhere, and a second pass did not change that.

The earlier assessment was right that this is the first thing to establish, and its reasoning is worth keeping in the buyer's hands: a tumour board product that cannot reach the source records automatically is a meeting tool rather than a case preparation tool. The entire value proposition, saving hours of preparation weekly, depends on records arriving without someone assembling them.

The multi site claim makes the gap wider rather than narrower. The vendor describes unifying oncology workflows across multiple hospitals and clinics with a single layer. Multi site oncology care is exactly where records are most fragmented: a patient may be diagnosed at one organisation, imaged at another, operated on at a third, and presented at a board hosted by a fourth. A product claiming to unify that must be reaching several record systems belonging to several organisations, and nothing describes how.

So the questions are basic and none has a published answer. Which record systems are supported, by what mechanism, and is it a permissioned interface or something else. Does the product reach records at organisations other than the customer, and under what authority. What identity do its reads carry. And what happens for a case where a participating site is not connected.

A buyer should establish this before evaluating anything else on this record, because the answer determines what the product actually is.

No hosting model, cloud provider, region, residency commitment or customer hosted option was located.

Two features of this product make the answer matter more than the general case.

The first is the live component. The platform retrieves insights in real time during a tumour board discussion, which means a query about a named cancer patient leaves the institution and returns while a treatment decision is being made. That is a different pattern from batch processing overnight, and it means the path is exercised repeatedly during a meeting attended by people from several departments and potentially several organisations.

The second is the decisions store. Because the platform accumulates care decisions rather than discarding material after use, residency governs a growing corpus rather than a transient pipeline. Where that corpus sits, and whether it sits in one place while the institutions contributing to it sit in several jurisdictions, is a question a buyer should settle in contract rather than accept as currently configured.

Nothing establishes whether an external model service processes clinical content, which for a product generating case presentations and matching trials is the first subprocessor question. Oncology case material is among the most sensitive content in any record, and a patient's diagnosis, stage and prognosis reaching a third party service is a disclosure a buyer needs to have authorised.

Ask for the hosting region, whether it can be fixed by contract, the subprocessor list, the model provider, and where the decisions store is held.

No price, tier or pricing mechanism was located. The only commercial statement available is a claim that the product cuts hospital costs by millions, which describes asserted value rather than what it charges.

Narrow by design and the narrowness is the point. This is an oncology product addressing one specific forum, the multidisciplinary tumor board, and it covers that forum end to end: preparation before the meeting, retrieval and insight during it, trial identification alongside it, and a decision knowledge base after it. It is positioned to span multiple hospitals and clinics under one layer, which matters because tumor boards are frequently run across sites.

Graded B rather than higher because no tumour type, staging system, guideline set or other instrument level behaviour was located, and no evidence establishes the product working in any named cancer centre.