W

Wisedocs

The machine learning is the mechanism the buyer pays for. The product reads unstructured medical and claims files and produces sorted, indexed, deduplicated chronologies and cited summaries with red flags; the human validation layer checks the model's output rather than producing its own, and the vendor's founding product in 2023 was the AI medical summaries platform. Remove the models and there is a document store and a review queue.

Grounding is real and documented with links to source, and a benchmark exists, but no accuracy figure for the product itself is published. Every summary is said to link to its original document, with citations for each fact, and each summary is validated by a person before release. The home page says the company released the Medical Long Context Reasoning benchmark to measure how models reason across long, fragmented medical records, plus an independently run edition with Artificial Analysis, which is a testing regime.

The benchmark page was not opened, so whether it reports Wisedocs' own product accuracy is not established, and it may say more. The FAQ answers the accuracy question in qualitative terms.

What runs alone, what a person must validate, where the review sits and what constrains it are all published. Modes: sorting, indexing, deduplication and co-mingled claimant detection run automatically; every medical summary passes through human-in-the-loop validation by trained expert reviewers before release. Review surface: users are notified when verified documents are ready for final review, every fact is cited to its source page, and activity logs track actions.

Constraints: review templates turn the customer's case strategy into the focus of the review, and the enterprise page states explainable models and audit trails. Route back: the customer's own final review sits after the expert validation. Nothing states a confidence threshold below which a summary is held back.

Figures appear without a named customer on the pages read. The home page states sixty to eighty percent faster first touch and up to three times lower manual review cost, the medical summaries page states up to seventy percent faster record reviews, and a careers listing states an eighty five percent time reduction and a fifty five percent cost reduction, all unattributed. The vendor said its customer base more than doubled at its 2024 raise. A customer stories section exists in the navigation and was not opened, so it may say more.

Substantive published commitments on segregation and deployment, with an adverse training position stated and privilege not named. Segregation at the level an in-house or firm buyer requires: role-based access, SAML single sign-on, SCIM provisioning, activity logs and data access safeguards, with private cloud, on-premise and hybrid deployment available. Training: the enterprise page states anonymized training methods and the vendor states models trained on more than a hundred million documents, which is a published position that customer content trains after anonymization, with no agreement located either way.

Third-party providers: regional model hosting is offered without naming the providers. Retention and deletion: not located, and no customer agreement is published. Nothing addresses privilege or work product for a product whose legal buyers use it to build causation and liability arguments.

No advice line or supervision statement is published. The pages for defense lawyers, claims legal teams and malpractice litigators sell the product to attorneys and paralegals as producing defensible documentation ready for audit, and the design with a person in the loop is described in terms of clinical accuracy. No page read says that summaries are not legal or medical advice, who should rely on them, or how the product supports a supervising lawyer's duties, and no customer agreement is published that could carry such a statement.

A published governance statement with real mechanisms and a testing program, short of a named owner or bias findings. The enterprise page's AI Governance section states human-in-the-loop validation, audit trails, explainable models and anonymized training methods as the basis for ethical, accurate and accountable AI, and the company has published the Medical Long Context Reasoning benchmark with an independently run edition through Artificial Analysis, which is a public testing regime for the class of models it uses.

No ISO 42001 or equivalent certification, no accountable owner and no statement about uneven output across record types is published on the surfaces read.

Some of the ground is covered, and the rest is not published. On access control, the enterprise page lists role based access, SAML single sign on, SCIM provisioning, full activity logs and data access safeguards, with SOC 2 Type II and HIPAA compliance stated. Not located: a retention period, a deletion commitment, a subprocessor list or an incident notification practice. No customer agreement, DPA or trust center was found on the pages read, and the privacy policy was not opened and may say more.

No liability position is published on the pages that could be read. Product, role, enterprise and use case pages describe the defensibility of outputs without any indemnity, cap, warranty or insurance position, and no terms of service or customer agreement surfaced in search or in the material read. The site footer was not inventoried, so whether an agreement exists was not established, and one may say more. This records what could be located on the date, not a finding that no position exists.

Integrations are referred to without documentation an implementer could use on the surfaces read. The enterprise page describes disconnected systems as the problem the platform solves and offers SAML and SCIM identity integration, and the platform accepts drag-and-drop upload of PDFs, faxes and images; no claims system, case management platform or document management integration is named with what moves and in which direction, and no integrations page was opened.

Deployment options are published with tenancy, residency and model-processing location. The enterprise page states that Wisedocs is offered as multi-tenant SaaS, private cloud, on-premise or hybrid deployment, that regional model hosting is supported to meet data residency requirements, and that high-availability architecture underlies the enterprise tier; that answers who shares infrastructure, where data sits and where the models run. Specific regions are not enumerated.

Certification is real and dated, short of a report reachable without asking. The vendor announced SOC 2 Type II compliance on 13 August 2024 with zero findings across security, availability, processing integrity, confidentiality and privacy, and states HIPAA compliance across its role pages. No auditor, coverage period, report route or trust center is stated on the surfaces read.

The supply chain is partly disclosed. The vendor states its models are domain-trained on more than a hundred million documents, that regional model hosting is available so inference can be placed by residency requirement, and that its MLCR benchmark evaluates frontier closed- and open-weight models, which implies third-party models in the stack; no provider or model is named and no change-notification commitment is stated.

No pricing information is published at any level. The site offers a demo and a call to discuss claims workflows, and no unit of charge, tier, figure or pricing page appears in the material read. The footer was not inventoried, so a pricing page there was not ruled out and may say more.

Segment and coverage are described with substance and the boundaries are partly stated. The vendor names workers' compensation, auto and casualty, disability, liability, healthcare, legal defense, government and IME and QME claims as supported review types, role pages for defense lawyers, claims legal teams, claims adjusters and evaluators, a malpractice use case with files exceeding a hundred thousand pages, and North American operation.

The legal pages are written for the defense side, which is a stated orientation; no jurisdiction or record type is named as unsupported.