Wisedocs

Domain trained models are the product, not a layer over a services business. The company states training on more than 100 million real world claims and medical documents, and the entire pitch is explicitly to displace business process outsourcing rather than to sell it: the machine sorts, deduplicates, indexes, extracts and summarises at volume, and human reviewers verify the output rather than produce it. That is the opposite arrangement from the services businesses this index has rejected, where a large human workforce does the work and AI is named in the marketing.

Among the better described oversight models reviewed, with one internal tension left unresolved. The company describes clinician and expert review of summaries as a structural step rather than an option, notifies users when verified documents are ready for final review, states plainly that the platform augments claims professionals rather than replacing them and that claims teams still make the final decisions, and links every summary, insight and risk signal to source with page level citations so any assertion can be checked against the underlying page.

Deduplication and co mingled record detection add a further integrity layer. Held at B because the sampling and thresholds are undisclosed and because two of the company's own claims sit in tension: every summary is described as passing human in the loop validation, while a headline outcome is that up to 70 percent of claims processing is automated. Ask which is true, what proportion of output a human actually reads, whether review is universal or sampled, and what the reviewers find when they look.

One genuine disclosure and one clear gap. The corpus is described, more than 100 million real world documents, which is more than most vendors say about what their models learned from, and page level citation makes each output traceable to its source page. Against that, no model or model family is named, no methodology is published, and there is no accuracy figure anywhere.

The company states only that the platform is highly accurate and that accuracy improves further through continuous model refinement and feedback. Highly accurate with no number attached is unfalsifiable, and it belongs on the same watchlist as absolute accuracy and up to 100 percent capture claims elsewhere in this index. For a product whose output is used to justify contested decisions, an unquantified accuracy claim is a weak foundation.

One control here is genuinely novel and this index has not seen it elsewhere. The platform detects co mingled records, identifying pages in a file that belong to a different claimant, which prevents both a wrong decision and an unauthorised disclosure at the same time: without it, one person's medical records travel into another person's claim file and are read by parties with no basis to see them, and nobody in the workflow is looking for it.

That is a privacy control arising from a quality feature and it deserves crediting on this axis. Two things hold the grade. The company states that accuracy improves through continuous model refinement and feedback, which implies customer documents inform model improvement, and no no training commitment, retention period or de identification posture was located to bound that. And the data subject position here is unlike anywhere else in this index.

The people whose medical records are processed are claimants: they have no relationship with this vendor, did not select it, are frequently the opposing party to the customer, and in most cases will never know the platform read their file. That does not make the processing improper, and it removes every practical check that exists in a treatment relationship, where a patient can ask, complain, or take their care elsewhere. It therefore raises the bar on what the vendor should publish rather than lowering it. Ask for the training position, retention on claimant records, and what happens to a file once a claim closes.

Deployment scale asserted without an independent measure, which this index grades C. The customer base is described only by type: leading United States property and casualty and disability carriers, a Department of Health and Human Services claims programme, one of the largest state workers compensation insurers. A federal claims programme is a meaningful procurement signal, but none of these organisations is named, so none can be verified or contacted.

Reported outcomes, 60 to 80 percent faster first touch, up to 70 percent of claims processing automated, up to a threefold reduction in manual review cost and case review up to 70 or 80 percent faster depending on the page, are entirely vendor reported, carry no stated denominator or baseline, and vary across the company's own surfaces. No independent study, audit or third party evaluation was located.

A compliance programme is claimed, covering HIPAA and SOC 2, and the co mingled record detection is a real privacy control that this index has not seen elsewhere: identifying pages belonging to a different claimant prevents both a wrong decision and an unauthorised disclosure.

Two things hold the grade at C. The company states that accuracy improves through continuous model refinement and feedback, which implies customer documents inform model improvement, and no no training commitment, retention period or de identification posture was located to bound that. And the data subject position here is unlike anywhere else in this index. The people whose medical records are processed are claimants.

They have no relationship with this vendor, did not select it, are frequently the opposing party to the customer, and in most cases will never know the platform read their file. That does not make the processing improper, but it removes every practical check that exists in a treatment relationship, and it raises the bar on what the vendor should publish rather than lowering it.

HIPAA compliance is explicitly claimed alongside SOC 2, with no business associate agreement terms published, which is the standard middle rung on this axis. One nuance specific to this market is worth understanding rather than treating as a gap.

Much of the work sits outside the HIPAA covered entity framework: workers compensation is subject to a specific HIPAA exception, property and casualty insurers and law firms are generally not covered entities, and records obtained through litigation discovery or an authorisation are governed by court rules and state privacy law rather than by HIPAA. Claiming HIPAA compliance in that setting is a credit rather than a requirement, but a buyer should establish which regime actually governs its own files, because the answer will often not be HIPAA.

SOC 2 Type II is claimed for the company itself rather than for a hosting provider or a supplier, and the report type is specified rather than left ambiguous, which answers the standing question this index asks of every SOC 2 claim. Held at B because no trust centre, security page, status page or scope statement was located, so the claim cannot be verified without asking, and because SOC 2 is an attestation rather than a certification. Ask to see the report and its scope under NDA, and confirm the observation period.

No clearance or device authorisation exists and none should. The earlier assessment's scoping is right and worth restating as the model for this corner of the index: the output informs a claims, coverage or litigation decision rather than a diagnosis or treatment decision, so the software as a medical device analysis that applies to clinician facing summarisation does not apply here.

What governs instead is a different and substantial body of law. Unfair claims practices regulation and state insurance department oversight reach how carriers evaluate and settle claims. Discovery and evidentiary rules reach anything produced for litigation. And a growing set of state statutes now governs the use of automated systems in coverage and utilisation decisions specifically.

The second pass adds that the vendor has articulated the right test for itself, which is unusual. It frames its own category around defensibility rather than accuracy alone, and names the components: page level citations tracing every assertion to a source document, human review before output reaches an adjuster, and an executed agreement before protected health information is processed. Those are evidentiary properties rather than clinical ones, and they are exactly what a record has to have to survive challenge in a claims dispute or a courtroom.

A vendor serving government agencies alongside carriers and law firms is operating where that standard is tested.

The useful question is not about a device pathway. Ask how the vendor supports a customer challenged on an automated coverage decision, and what its outputs are designed to withstand.

The grade describes disclosure and the structural position of the product, not any allegation of misconduct, and the credits belong first. The company states explicitly that it augments rather than replaces claims professionals and that humans make the final decision, it builds auditability in through page level citation, and defensibility is the stated design goal. Now the structural point a buyer and a regulator should both see. This product sits on one side of an adversarial process.

Its customers are carriers, third party administrators, defence firms and independent evaluators, and its marketing language is about spotting inconsistencies, catching red flags, flagging contradictions, assessing liability faster and building stronger arguments. The person whose records are summarised is generally the claimant on the other side.

This index already holds the principle that on the payer side automating an approval is low risk while automating a denial is not, and the same asymmetry applies here: a summary that omits a supporting treatment record disadvantages someone who cannot see the tool, was never told it was used and has no route to contest its output.

Nothing published addresses that asymmetry, and no fairness, subgroup or demographic performance disclosure of any kind was located, which matters in workers compensation and disability where outcome disparities are well documented.

One genuine disclosure and one clear gap. The corpus is described at a stated scale of real world documents, which is more than most vendors say about what their models learned from, and page level citation makes each output traceable to the source page, which is the right granularity for a document review product: a reviewer disputing a summary can go to the page rather than to the file.

Against that, no model or model family is named, no methodology is published, and no accuracy figure appears anywhere. The company states only that the platform is highly accurate and that accuracy improves further through continuous refinement and feedback.

Highly accurate with no number attached is unfalsifiable and belongs on the same watchlist as the absolute accuracy claims recorded elsewhere in this index, and the continuous improvement clause compounds it, because a moving system cannot be characterised by any figure even if one were offered.

The stakes are what make it weak rather than merely unsatisfying: the output is used to justify contested decisions about benefits and claims, where the affected person is on the other side of the dispute and a missed page or a mischaracterised record shifts a determination they cannot audit. Ask for an accuracy figure with its method and denominator, the miss rate on clinically relevant pages, and what a claimant or their representative can obtain about how their file was summarised.

Graded against what this product is rather than penalised for a mismatch. There is no EHR integration of any kind and none should be expected, because the buyer is an insurer, a third party administrator or a law firm and does not operate one.

What exists instead is genuine format agnostic ingestion, handling PDFs, faxes, images, low quality scans and handwritten notes, with automatic sorting, splitting, labelling and deduplication, plus an open API for pushing structured summaries into existing claims and legal systems. That is real interoperability for its market. It is graded C rather than higher because the API is the only named integration surface, no named claims or case management systems are listed, and no standards based exchange is described.

A second pass reaches the vendor's enterprise material and overturns the earlier absence comprehensively. This is the strongest disclosure on this axis encountered anywhere in the recent grading, and it answers both halves of the question rather than one.

On deployment model, four options are offered and named: multi tenant software as a service, private cloud, on premise, and hybrid. Offering an on premise path at all is rare in this category, and it matters for the buyers this product serves, since some carriers and government agencies cannot place claimant records in a shared environment.

On residency, the vendor states that data can be stored in the United States or Canada according to a customer's regulatory needs, and separately that regional model hosting is supported. That second point is the one most vendors miss. Committing to store data in a region means little if inference runs elsewhere, and this vendor addresses both.

Supporting controls are named rather than gestured at, with transport and storage encryption standards specified, access controls, and a regular programme of penetration testing and vulnerability scanning.

Two things hold it short of the top grade. No subprocessor list was located, which matters for a platform that necessarily involves storage, processing and possibly model services. And the residency statement extends to other regions as required, which is commercially sensible and leaves the set undefined.

Ask for the subprocessor list and for the residency commitment in contract language.

No price, tier or pricing mechanism is published and every route ends at a demo booking. The only commercial framing offered is a claimed threefold reduction in manual review cost, which describes displaced spend rather than what the platform charges.

Broad across claim types and reviewer roles rather than across clinical specialties, which is the right measure for this product. Named coverage spans workers compensation, disability, auto and general liability, and personal injury litigation, and the roles served include claims adjusters, insurance defence and claims legal teams, independent medical evaluators and qualified medical evaluators, peer review firms, third party administrators, carriers and public sector claims programmes.

Graded B rather than A because the depth is horizontal across claim types rather than demonstrated through instrument level work in any of them, and because the vendor itself notes the value scales with document volume, so smaller or occasional files see materially less benefit.