314e Dexit

The model drives the core module, and the product now carries value without it. DextractLM, the company's own model for healthcare documents, classifies more than 150 document types, extracts patient and payer fields, matches the patient record and routes each document, and that automation does not exist without it.

But Dexit is now sold first as a HIPAA cloud fax service (virtual numbers, porting, delivery tracking, retries, audit logs, routing rules) with eligibility and claim status checks and EHR write-back alongside, and a buyer moving off on-premise fax servers would still have a product with the model removed.

The review route is published, and the page then promises it makes the output correct. Documents whose classification confidence falls below a set threshold are queued for human review before routing, while high-confidence documents, described as the majority, are classified, extracted, matched to a patient and written to the record with no one reading them. The FAQ then says this "ensures nothing is misclassified or misfiled", and the page adds "Nobody manually read it.

Nobody manually sorted it." That promise is where the risk sits: a document filed confidently to the wrong patient's chart never reaches the review queue, and a buyer told nothing is misfiled will not audit for it. The threshold itself, the share of documents it catches and any reconciliation reporting are not published.

The approach and the model are named without version or evaluation. DextractLM is described as a proprietary model built for healthcare documents, trained on healthcare documents and refined by customer corrections, classifying more than 150 document types in under ten seconds and extracting patient, payer, diagnosis code and date fields; earlier company material named a classification technique it calls a Fusion Strategy and separate extraction and classification figures. No model card, training data description, version history or test set is published, and a model that retrains on corrections changes without a published schedule.

A structurally short chain: the model is the company's own and, it says, self hosted. The product page presents DextractLM as a proprietary model built for healthcare documents, and the company stated in September 2025 that Dexit runs on self hosted models rather than an outside inference service. The security policy places all core infrastructure and data storage in the United States in Tier 4 data centers.

The company publishes no subprocessor list (the privacy notice lists categories of third parties for the website and marketing, not the product), does not say where the self hosted models run, and states no retention or training terms for the chain. Eligibility and claim status checks run through the company's own Veritable product.

Vendor reported figures with no method, and no named customer on the product page. The page claims 95 percent or better model accuracy, processing in under ten seconds against more than eight minutes per document manually, and go-live in under four weeks; earlier company material gave over 95 percent extraction precision and 97 percent classification accuracy. None carries a test set, document count or customer. The page says Dexit is trusted by hospitals, RCM companies and physician practices without naming one; 314e's company-wide customer and success story pages may name Dexit sites.

Categorical statements are published about where the model runs and how it learns, without a retention schedule. The company stated in September 2025 that Dexit runs on self hosted models, so document content is not sent to an outside model provider's interface, and its security policy keeps core infrastructure and data storage in the United States with encrypted backups.

The product page adds that every correction a customer's team makes feeds back into the model, which answers the training question in the permissive direction and leaves open whether one customer's corrections improve a model shared with others. No retention period for faxes, documents and extracted fields, and no de-identification statement, is published.

A business associate agreement comes with the product, with the agreement itself not published. The Dexit page lists "BAA included" among its commercial terms, next to no auto-renewal and predictable costs, so it is clear before contact that one is standard. The company's security policy is written to the HIPAA Security Rule and covers the provision of its products and services, and together with the privacy notice it describes itself as the company's commitments to customers.

The agreement's text, the tier it applies at and whether it binds subcontractors are not available before contracting. The roles are simple: hospitals, practices and RCM companies handling their own or their clients' patient documents make 314e a business associate or subcontractor.

The company presents a SOC 2 Type II as its own. Its security policy says 314e completed a SOC 2 Type II audit, by an auditor it does not name working with the Vanta platform, covering security, availability, processing integrity, confidentiality and privacy across its company-wide data handling, and the Dexit page carries a Type II badge. Its policies are described as based on the NIST Cybersecurity Framework and HITRUST, which is alignment rather than a HITRUST certification.

The policy also describes background checks, least-privilege access with multi-factor authentication, encryption in transit (at rest where sensitivity requires), logging, incident response and breach notification. The report, its period and its system boundary are not published, and there is no trust center offering them; a companywide audit across six products leaves open whether Dexit sits inside the tested boundary.

No device claim is made and none applies: classifying, extracting and filing inbound documents, and checking eligibility, are administrative actions. The compliance surface is records governance, since a document filed to the wrong patient's chart is a record integrity failure with clinical consequences, and HIPAA and accreditation standards require accurate records. The company states no position on that exposure beyond its human review of low-confidence documents.

Nothing is published on how the model's behavior is governed or tested. The product page gives a single figure, 95 percent or better model accuracy, with no test set, document mix or method, says every correction a customer's team makes feeds back into the model, and describes a confidence threshold for human review without stating it. No evaluation, error analysis, change control for a model that learns from corrections, or performance breakdown is published.

The exposure is technical rather than demographic: classification and extraction degrade with scan and fax quality, handwriting and form layout, so documents from smaller or older senders may be misfiled more often, and a single aggregate figure hides that.

A review route exists and nothing stands behind the output. Low-confidence documents are queued for human review, fax transmissions carry an audit log, and the RCM dashboard reports extraction accuracy, throughput and cost per document by client and week, so a customer can measure how the product performs on its own documents.

No terms of service are linked from the site, and no accuracy warranty, remediation commitment or error rate with method and denominator is published; the 95 percent figure has neither. The harm case is a document filed to the wrong patient's record, which the person it belongs to cannot see.

Named record systems with one directional writing. The product page says Dexit writes structured data from each inbound document directly to the correct patient record and matches documents to existing records, naming Epic, MEDITECH, Cerner, athenahealth, eClinicalWorks, NextGen, Allscripts and PointClickCare, and for practices AdvancedMD and Practice Fusion, through what it calls standard integration patterns with no middleware; documents also arrive by email, scan, API and forms. No integration method, marketplace listing or named production site is shown, and the writing runs one way into the record.

A single cloud service with residency committed. Dexit is sold as cloud fax and document processing, and the company's security policy states that all core infrastructure and data storage are in the United States, in Tier 4 data centers at dispersed locations, with encrypted backups at separate US sites; the company has said the models are self hosted.

Tenant isolation, the provider that operates the data centers and the processing path for support and engineering access are not described; the company has offices in Yardley, Dublin (California) and Bengaluru, and nothing says whether staff outside the United States can reach customer documents.

No price, tier or unit is published, and the commercial terms are partly stated: the Dexit page promises no auto-renewal, no surprise price increases, a business associate agreement included and costs that stay predictable as volume grows, with cloud fax live within days and full document automation in under four weeks. The RCM dashboard reports cost per document per client, which suggests a per-document unit, but no rate is shown and every route leads to a demo.

Whether the unit is the page, the document, the fax number or the facility is left open, as is whether cloud fax and document processing are priced separately and whether eligibility checks carry their own charge.

Three settings are described with a workflow each and no named reference in any. Hospitals are sold the move off on-premise fax servers with documents filed to the patient record; RCM companies are sold per-client volume without added headcount and a dashboard of accuracy and cost per document; physician practices are sold intake with insurance verified before scheduling, naming ambulatory record systems.

The document types (prior authorizations, EOBs, referrals, lab results, discharge summaries) are administrative, and PointClickCare in the integration list points to post-acute use as well. Breadth is claimed but no customer is named on the product page.