Sample Healthcare

The reasoning and extraction engine is the product. There is no system of record, services organisation or prior software platform underneath, and every described capability, document processing, insight extraction from medical records, correspondence automation and the copilots themselves, is model output. The company's stated endpoint of full automation only makes sense for a company whose product is the model.

This vendor states its autonomy trajectory openly and that candour deserves credit before the grade is explained. The company describes a progression: it starts by making a team roughly ten times faster, its systems then learn from that team, and eventually they fully automate, with the customer staying in control. Almost every other vendor in this index presents human oversight as a permanent design property.

This one presents it as a stage on the way somewhere else, and says plainly that the system learns from the people whose work it is heading toward absorbing. Graded C because nothing governs the transition. No accuracy bar, confidence threshold or performance gate is published for when a workflow graduates from assisted to autonomous, no abstention behaviour is described, and no statement identifies which workflows are considered safe to automate fully and which are not.

The audit trail and lineage capability is real and valuable, but it is forensic: it lets an organisation reconstruct what a system did after the fact, it does not prevent the system doing it. Ask what the criteria for full automation are, who signs off on each graduation, and whether appeals and clinical adjacent workflows are treated differently from correspondence.

One strong infrastructure property and no model level disclosure. Versioned lineage graphs, end to end visibility across datasets, decisions and hand offs, and audit trails tracing any value back to its source are genuine engineering artefacts rather than interface features, and they are more developed than most of this category offers.

Against that, no model or model family is named, no accuracy figure exists, no evaluation methodology was located, and the precision claim is expressed as unparalleled, which is unfalsifiable as written. The phrase your data, your models, your intelligence suggests customers may bring or control their own models, which would be a significant disclosure if true and is nowhere explained. Ask what it means.

The company says its systems learn from the customer's team over time, and separately that models are tailored to each organisation. Read together those suggest learning is confined to a customer's own model rather than pooled across customers, which would be the favourable reading and the one most buyers would want, and it is not stated, so a buyer is relying on an inference the vendor could confirm in a sentence.

Establish whether corrections and interventions by one customer's staff can influence outputs seen by another, and whether any shared base model is updated from customer material. The question this arrangement raises most sharply is ownership.

A model shaped by a customer's own team over months is an asset built from their staff's work and their patients' records, and no published term says who owns it, whether it is deleted, returned or retained by the vendor at termination, or whether the customer can take the accumulated tuning with them. That is a commercial lock in question and a data question at once, since a retained model carries forward what was learned from records the customer no longer wants the vendor to hold.

The material involved is broad, since the platform reads medical records, correspondence, intake documents and appeal files, so the training question covers clinical content rather than metadata. Ask for retention, the isolation boundary, and the disposition of a tailored model on termination.

Outcomes claimed, none verifiable. Customers are described only by type: publicly traded companies, national provider groups, a top diagnostics provider and a major health system. No organisation is named, no case study exists, no funding was located and no independent evaluation was found.

The published figures do not reconcile cleanly either, with 85 percent of tasks automated, 2 to 5 times speedups and roughly 10 times faster teams offered as separate claims with no baseline, denominator or measurement method. One customer outcome sentence on the company's own site is garbled to the point of being unreadable, and parts of the site still carry unreplaced placeholder text, which is weak evidence that the public material has been reviewed carefully.

No retention period, de identification posture or explicit training use commitment was located.

The statement the earlier assessment isolated is the right thing to press, and the second pass makes its likely meaning clearer without confirming it. The company says its systems learn from the customer's team over time, and separately that models are tailored to each organisation. Read together those suggest learning is confined to a customer's own model rather than pooled across customers, which would be the favourable reading and the one most buyers would want.

It is not stated, and the difference matters enough to confirm in writing. Establish whether corrections and interventions by one customer's staff can influence outputs seen by another, whether any shared base model is updated from customer material, and what happens to a tailored model at termination: whether it is deleted, returned, or retained by the vendor.

That last question is the one this arrangement raises most sharply. A model shaped by a customer's own team over months is an asset built from their work and their patients' records, and no published term says who owns it or what becomes of it.

The material involved is broad. The platform reads medical records, correspondence, intake documents and appeal files, so the training question covers clinical content rather than metadata.

Ask for the retention schedule, the isolation boundary between customers, and the disposition of a tailored model on termination.

A precision that matters. The company references HIPAA in a specific and narrow context, stating that its audit trails satisfy HIPAA, SOC 2 and FDA record keeping requirements in real time. That is a claim about the auditability of its logs, not a statement that the company is HIPAA compliant, and it is certainly not a business associate agreement commitment.

No compliance statement and no BAA terms were located anywhere. Graded C rather than Not Rated because a HIPAA adjacent claim exists, and it is a narrower one than a reader skimming the page would take it to be.

A new variant for the compliance language watchlist. The site says its audit trails satisfy SOC 2 record keeping requirements. Satisfying a standard's record keeping requirements is not the same as holding an attestation against that standard, and the two read almost identically at a glance.

No SOC 2 report, ISO 27001 certification, HITRUST certification, trust centre or security page was located. Graded C because a real capability is described, comprehensive logging, while the credential a security review actually asks for is absent.

Ask directly whether a SOC 2 report exists, of which type, and over what period. This joins SOC 2 aligned, SOC 2 certified data centres and certified providers at all stages on the watchlist.

No clearance or device authorisation exists and none should. This is administrative document processing spanning patient intake, eligibility, medical record reading, revenue cycle work and insurance appeals. It produces no clinical output and no device framework attaches.

The stray reference the earlier assessment noticed is best read as loose marketing, since nothing in the product description touches a regulated device function. If it signals anything, it would be customers with regulated record keeping obligations of a different kind, and that is worth confirming.

What governs is the regime that applies across this corner of the index: unfair claims practices regulation, state insurance oversight, and evidentiary standards where output supports an appeal.

One distinction deserves stating precisely, because it is easy to get backwards. The platform requests and checks prior authorisations through named intermediaries, so it operates on the requesting side of that transaction rather than the deciding side. The wave of state legislation now restricting automated systems in utilisation review is aimed at payers using software to deny or limit care. It does not reach a provider using software to assemble and submit a request. A vendor on this side of the transaction sits outside those statutes, and a buyer should understand that rather than assume the opposite.

What does apply is accuracy. A submitted authorisation or appeal assembled by software becomes part of a record a payer relies on.

Ask what review precedes submission, and who is accountable for a submitted error.

The grade describes disclosure and the credit is genuine. Stating publicly that the endpoint is full automation, and that the systems learn from the team in the meantime, is more honest than the standard permanent human in the loop framing that several competitors offer without qualification. Auditability is treated as infrastructure.

Against that, no fairness, subgroup or demographic performance disclosure of any kind was located, no governance framework or responsible AI documentation exists, no criteria are published for what may be automated, and the accuracy claim is an unfalsifiable superlative.

The workflows named, prior authorisation and insurance appeals, sit on the payer facing side where this index holds that automating an approval is low risk while automating a denial or an appeal outcome is not, and nothing addresses that distinction.

One infrastructure property here is a genuine engineering artefact rather than an interface feature, and it is the right one for this axis. Versioned lineage graphs, end to end visibility across datasets, decisions and hand offs, and audit trails tracing any value back to its source together mean a disputed output can be reconstructed: a reviewer can establish which data produced it, which version of the system was running, and where a hand off occurred.

Most vendors in this index offer citation at best, which shows where a statement came from and not what the system did with it. Lineage is the stronger property because it survives the question a citation cannot answer, namely what changed between the source and the output. Held at C because nothing is measured and one claim is unfalsifiable as written.

No model or model family is named, no accuracy figure exists, no evaluation methodology was located, no warranty, indemnity or remediation commitment attaches, and the precision claim is expressed as unparalleled, which asserts a comparative superiority with no comparator and no number.

One phrase deserves a direct question rather than an inference: the framing that this is your data, your models, your intelligence suggests customers may bring or control their own models, which would be a significant disclosure if true and is nowhere explained. Ask what it means, and for accuracy figures with the test set behind them.

Integration with EMR and revenue cycle systems is claimed alongside deployment in days, which is the right shape for a document workflow product that has to reach both clinical and financial systems. Nothing is named: no EHR, no revenue cycle platform, no integration standard, no marketplace listing and no customer deployment. An organisation cannot establish whether its own stack is supported without asking, and the deploy in days claim is unverifiable without knowing what it integrates with.

No hosting model, cloud provider, region, residency commitment or customer hosted option was located, and the claim the earlier assessment flagged remains unexplained in architectural terms. A statement that a customer's data, models and intelligence remain under their control implies something meaningful and is not a description of where anything runs.

The second pass supplies a partial reading. The vendor states that models are tailored to each organisation, so the control claim most likely refers to per customer model specificity rather than to customer controlled infrastructure. Those are different things and a buyer should not accept the first as the second.

What the second pass does establish is that this axis is not a single question here. The platform names connections to customer data warehouses, prior authorisation intermediaries, a customer relationship platform and a clinical fax network, and describes connecting to any tool, data lake or payer portal. So the product is a hub, and information moves outward through several routes to parties that are not this vendor.

That reframes the residency question usefully. Where the platform hosts matters, and so does what leaves through each connector, under whose agreement, and to which jurisdiction. A prior authorisation intermediary and a fax network each hold protected health information in their own right.

Ask for the hosting region, the full connector and subprocessor list, what data each connector transmits, and whether the control claim means anything about infrastructure.

No price, tier or pricing mechanism was located. The commercial arguments offered, defensible return on every workflow and multiples of speed improvement, describe claimed value rather than cost.

Coverage is described by workflow rather than by care setting or specialty, and the span claimed is wide: patient intake at one end through to insurance appeals at the other, with correspondence processing and prior authorisation named in between. Buyer types include a diagnostics provider and a health system, which are genuinely different organisations with different document problems.

Graded C because no clinical setting, specialty or instrument level behaviour was located, and because the breadth is asserted across a very wide span with only two unnamed customer anecdotes supporting it.