Fourier Health

The models do the work and the humans check it, which is the arrangement that distinguishes an AI product from a services business with AI attached. Classification, labelling, structuring and summarisation of inbound records are all performed automatically, and the proprietary clinician network reviews and validates output rather than producing it. That is the opposite of the human scribe model this index grades down elsewhere, where a workforce drafts and software assists.

One qualification kept in view rather than applied against the grade: the company describes using best in class large language models rather than models it owns, so the differentiation sits in the pipeline, the specialty configuration, the attribution layer and the clinician network rather than in the underlying model.

Among the better described oversight models in this category, combining two things most competitors have only one of. A proprietary network of specialist clinicians reviews and validates summaries, which is a human gate staffed by people with the training to catch a clinical error rather than a formatting one, and the company states this layer also feeds continued improvement of its models.

Separately, attribution traces every summarised element back to its source location in the original records, so a reviewer can verify any assertion without leaving the product. Held at B because the parameters are undisclosed: it is not stated whether every summary is reviewed or a sample, at what rate, what reviewers find when they look, or what happens when a summary and its source disagree. No confidence signal, threshold or abstention behaviour is described. Ask for the review rate and the error rate it produces, since a clinician network that never reports finding anything is not an oversight layer.

One strong output level property and no model level disclosure. The strong part is attribution: every summarised element can be traced to where it appears in the original medical records, which the company frames explicitly as accountability and audit proofing rather than as a convenience feature, and that is the right framing for a category whose failure mode is silent.

Against that, no model or model family is named beyond a reference to best in class large language models, no accuracy figure exists, no omission measure is published, and no evaluation methodology or model card was located. Ask which models the pipeline uses and whether patient content reaches a third party inference provider.

The clinician network is simultaneously the strongest oversight feature on this record and its largest exposure, and both halves belong in one assessment rather than being weighed separately. A proprietary network of specialist clinicians reads patient records in order to validate summaries, which means identifiable clinical content reaches people outside the treating organisation and outside the vendor's own employees.

This index recorded the same duality in an ambient documentation vendor whose human review workforce delivered the best oversight in its lane and the widest exposure, and the guidance is the same here: evaluate the two as one decision, because the quality benefit and the disclosure are produced by the same mechanism and a buyer cannot have one without the other. What is missing is the description that would let a buyer size it.

Nothing states where the clinician network is located, how it is contracted, what each reviewer can see, whether access is scoped to the case in hand, or what is logged. And the company states the review layer supports continued improvement of its models, which implies reviewed content informs training without bounding it in any way. No retention period, de identification posture or model provider was located. Ask where the reviewers sit, what they can see, what is retained, and whether reviewed content trains the models.

Commercial validation without a published measure of benefit, which this index grades C. An 8.4 million dollar seed round in October 2025 led by Yosemite with several named participating funds is real external diligence, the founding team carries verifiable operating history at Redox and InfiniteMD, and a listing on the AVIA marketplace means an outside party assessed the product for its catalogue. Deployment is described as dozens of provider organisations and not one is named.

No case study, no quantified outcome, no accuracy data and no independent evaluation was located. For a company founded in late 2023 that is unsurprising, but it means nothing published establishes how well the summaries perform.

The clinician network is simultaneously the strongest oversight feature on this record and its largest privacy exposure, and both halves belong in the assessment. A proprietary network of specialist clinicians reads patient records in order to validate summaries, which means identifiable clinical content reaches people outside the treating organisation and outside the vendor's own employees.

This index recorded the same duality for Speke, where a human review workforce delivered the best oversight in its lane and the widest PHI surface, and told buyers to evaluate the two as one decision. The same applies here. Additionally, no retention period, de identification posture or training use commitment was located, and the company states the review layer supports continued improvement of its models, which implies reviewed content informs training without bounding it. Establish where the clinician network is located, how it is contracted, what it can see, and what is retained.

No compliance statement or business associate agreement terms were located, and the chain question the earlier assessment raised is now the substance of this axis rather than a detail.

The vendor describes a proprietary network of specialist clinicians who review and validate summaries. So named individuals outside the customer organisation read patient records as a matter of ordinary operation, and the agreement has to reach them.

One feature of that arrangement deserves stating because it differs from every other human review model this index has recorded. A specialist clinician reviewing another organisation's patient records is not treating that patient. They have no relationship with them, no duty of care to them, and the patient has no way of knowing they exist. Their access is quality assurance rather than treatment, which is a different permitted purpose with different constraints, and it cannot be justified on the basis that a clinician is looking at a chart.

So the questions are specific. Are these reviewers employees, contractors, or a separate entity, and if contracted, are they workforce under the agreement or subcontractors requiring their own. Where are they located. Are they licensed, and does that matter for the use case. Does every record pass through review or a sample, and can a customer decline. Is their access logged in a way the customer can audit.

The document types widen the exposure. Referral packets, faxes and handwritten notes arrive containing whatever the sending organisation included, which is frequently more than the receiving clinician needs.

Ask for the agreement, the reviewer status and location, and the access log.

No attestation and no trust centre were located. The company raised seed funding in late 2025 and was founded in late 2023, so the absence is unremarkable for its stage, and the earlier assessment was right that it should nonetheless be the first item requested.

The reason is the review network. Where a vendor's own architecture routes patient records to external specialist clinicians, the controls that matter are ones no product description reveals: how reviewer accounts are provisioned and revoked, whether access is scoped to assigned work or open across the estate, whether it is time bound, how personnel are screened, what device and network conditions apply, and what happens when a reviewer leaves the network.

A reviewer working from their own equipment, on records belonging to an organisation they have never dealt with, is a materially different access pattern from an employee inside a managed environment. Nothing published describes which applies.

That argument is the same one this index has now applied to several vendors with people in the loop, and it points the same way each time: a vendor whose model depends on staff access carries a heavier disclosure burden than one running a fully automated pipeline, not a lighter one. Human review genuinely improves output. It also enlarges the surface an examination exists to describe.

The founders' background in health data infrastructure suggests the vendor will understand exactly what is being asked, which makes a direct request more likely to produce a useful answer than a questionnaire.

Ask what external testing has been performed, the reviewer access model, and the roadmap to an attestation.

No clearance, device authorisation or exemption analysis was located, and the earlier assessment's scoping read holds: the product summarises and triages rather than diagnosing or recommending treatment.

Two design features strengthen that position and both deserve crediting, because they address the conditions the conditional exemption for decision support actually turns on.

The first is attribution. The vendor states that a provider can see where each piece of summarised data is attributed to in the original medical records, and that every output is cited to source. That is the professional's ability to review the basis, built as a product feature. It is the fourth record in this lane to offer it and the most explicit.

The second is the approval gate. The vendor describes summaries as configured per use case so that a provider simply needs to approve, which places a human decision between output and use.

The point the earlier assessment raised remains the live one and is not resolved by either. Pre visit readiness assessment and document triage shape what a clinician sees and in what order. A summary that is use case specific is by construction a selection, and selection is a judgement about relevance made before the clinician's judgement begins. Attribution lets a clinician verify what is present; it does not reveal what was omitted.

So the question to put is about exclusion rather than inclusion. Ask what determines that a finding is not relevant to a given use case, whether a clinician can see the full record from the summary view, and what happens to a finding the system judged immaterial.

One further note: clinical appeals is a named use case, which puts some output into a payer facing dispute rather than the chart.

The grade describes disclosure and incentive structure, and the credits are real. Attribution to source, a clinician verification layer and the company's own framing around veracity and auditability are the right instincts for this category, and more than most competitors at this stage offer.

Against that, the commercial framing places the work inside the revenue cycle explicitly: the company describes the process as reimbursable, and describes the benefit as improving clinical decision making while increasing revenue through better risk capture and coding accuracy. Risk capture is the same gradient this index tracks across the category, and a summarisation layer whose value is partly expressed in coding accuracy is operating on the documentation that determines payment.

Nothing here suggests inflation, and the counterweight of clinician verification is genuine, but the incentive should be visible to a buyer. Separately, no fairness, subgroup or demographic performance disclosure was located.

Two mechanisms sit on this record and the framing of the first is what makes it count. Every summarised element can be traced to where it appears in the original medical records, and the company frames that explicitly as accountability and audit proofing rather than as a convenience feature.

That framing matters because it identifies the right problem: in a category whose failure mode is silent, a traceable element is one a reviewer can challenge and an untraceable one is a claim nobody will ever test. A vendor that describes attribution as an audit property has understood what it is for. The second is the review layer, since a network of specialist clinicians validates summaries, which places domain trained humans between the model and the clinician relying on it.

Held at C because nothing is measured. No model or model family is named beyond a general reference to leading language models, no accuracy figure exists, no omission measure is published, and no evaluation methodology, model card or warranty, indemnity or remediation commitment was located.

Omission remains the number that matters, because a reviewer checking traceable elements is checking what is present and has no way to see what was never summarised, and a specialist reviewer improves that but does not solve it. Ask for the omission rate on clinically significant content, what proportion of summaries a clinician reviews, and which models the pipeline uses.

Unusually strong on standards and thin on named systems, which is the reverse of most records here and reflects the founding team's background building a major interoperability platform. Named protocols rather than vague claims: SMART on FHIR and HL7 version 2 for integration, bidirectional exchange through HL7 FHIR APIs, health information exchange connectivity for retrieving records the organisation does not hold, and user interface embedding inside systems such as Epic.

Delivery is genuinely flexible, covering EHR embedding, inbox push, API and a standalone web application, with output as structured fields or as a PDF. The company also describes aligning with existing ambient documentation workflows rather than competing with them, which is a thoughtful position. Held at B because Epic is the only EHR named and no customer integration is evidenced, so breadth is asserted through protocol support rather than demonstrated.

No hosting model, cloud provider, region or residency commitment was located. The delivery routes described, embedding in record systems, inbox, interface and web application, describe how output reaches a user rather than where processing happens, and that reading stands.

One architectural choice is now clearer and it is genuinely useful. The vendor states that a customer can bring their own model, use the vendor's, or combine both. Model portability of that kind is rare in this category and it is the right default: it lets a health system that has already negotiated terms with a model provider keep that relationship rather than inherit the vendor's.

It also means this axis has no single answer. Where a customer supplies the model, the residency and retention questions for inference belong to that customer's arrangement, not this vendor's. Where the vendor's model is used, they belong here and nothing is published about them. Establish which configuration applies to you before asking anything else, because the answers differ.

The reviewer question remains the one this row cannot avoid. Records flow to an external network of specialist clinicians, so the question is not only where the software runs but where the people are. A reviewer outside the United States would place protected health information offshore regardless of hosting, and that is a permitted arrangement that must be papered rather than a defect.

The ingestion surface is also unusually broad, covering faxes, handwritten notes and referral packets running to thousands of pages, so a large volume arrives before any summary exists.

Ask for the hosting region, the reviewer locations, the subprocessor list, and which model configuration applies.

No price, tier or pricing mechanism was located.

One commercial claim needs unpacking rather than crediting. The company describes its process as reimbursable because it sits within the revenue cycle. That is not a statement about what the product costs, and a buyer should establish precisely what it means, since a claim that a vendor's work is billable under existing codes is a materially different proposition from a claim that it improves revenue. The two are easily conflated in a sales conversation and they carry different compliance exposure.

Configurable rather than broad, which is the more useful property for summarisation. Summaries are shaped to specialty and use case, with oncology, geriatrics and rheumatology named as examples, and the same underlying records produce different outputs depending on what the summary is for.

Named workflows span referral management, pre visit readiness assessment, document labelling and triage, and record intake automation, which covers the front door of a practice rather than only the clinical encounter. Graded B rather than A because the specialty configuration is described as a capability rather than demonstrated through instrument level behaviour in any named specialty, and because no care setting beyond ambulatory provider organisations is addressed.