314e Dexit

The model is the product. 314e built and named a proprietary extraction model, DextractLM, described as purpose built for healthcare documentation rather than adapted from a general purpose model, and the classification and extraction it performs is the entire value of Dexit. There is no underlying document repository, scanning estate or network that would retain worth without it.

Note the distinction this record depends on: the parent company is substantially a services and consulting business, but Dexit is a separately licensable product built around its own model, which is why it clears the services test that this index has applied to reject vendors whose AI turns out to be a thin layer over human delivery.

Dexit automatically identifies a document's type and associates it with the correct patient and encounter, which is a consequential automatic action: a document filed to the wrong chart is both a record integrity failure and a potential patient safety event, and it is the kind of error that stays invisible until someone goes looking for a result that is not there. The company publishes accuracy figures but describes no oversight architecture around them.

There is no stated confidence threshold, no exception queue behaviour, no description of what happens to the several percent of documents the model gets wrong, no human verification gate before filing, and no audit or reconciliation reporting offered back to the health information management team. Accuracy is not oversight, and a published rate without a described handling path for the residual is exactly the gap this axis measures.

Materially more disclosure than the other document intake products assessed here. 314e names its model, DextractLM, states that it is purpose built for healthcare rather than general purpose, states that Dexit runs on self hosted models rather than third party inference APIs, names the classification technique it calls a Fusion Strategy, and publishes precision figures for both extraction and classification.

Naming the model and the hosting posture answers two questions most vendors in this shape leave open. Held at B because the disclosure stops at the names: there is no model card, no training data description, no test set, and Fusion Strategy is a product name rather than a described method, so an evaluator cannot assess how classification actually works or where it degrades.

One substantive architectural answer, which is more than most products of this shape offer. The company states that the product uses purpose built, self hosted models, and for a system ingesting complete clinical documents that removes the largest and least visible exposure in generative document processing: document content is not being sent to a third party model provider.

Most vendors in this segment either say nothing or describe controls over a pipeline whose final hop they never mention, so naming the hosting posture answers the question a security review would reach last and care about most. Held below the top grade on an ambiguity that matters and should be resolved rather than assumed.

Self hosted is doing a great deal of work in that sentence, and it can mean hosted inside the customer's own environment, where the customer's controls apply and no data leaves, or hosted on the vendor's infrastructure, where the third party provider is removed and the vendor becomes the sole custodian instead. Those are materially different positions and only the first eliminates transfer.

Nothing else is established either: no retention for processed documents and extracted entities, no statement on whether customer documents further train or tune the named model, and no sub processor list. Ask where the self hosted models actually run, retention on documents and extractions, and the training position in writing.

Two specific accuracy figures are published, over 95 percent precision in entity extraction and 97 percent document classification accuracy, and publishing numbers at all puts 314e ahead of most competitors in this shape. But neither figure carries a derivation. No test set is described, no denominator or document count is given, no breakdown by document type or source quality is offered, and no independent or customer validation was located.

This index's standing lesson from Pieces applies directly: a quantified error rate is only as good as its published derivation, and publishing a number invites the methodology question that never publishing one avoids. No named customer, reference site or case study was located, and the throughput claims of ten times faster than manual and up to three times efficiency have no stated baseline.

One substantive architectural answer, which is more than most of this shape offers. 314e states that Dexit uses purpose built, self hosted models. For a product ingesting complete clinical documents, self hosted inference means document content is not being sent to a third party model provider's API, which removes the largest and least visible exposure in generative document processing.

That is the same class of answer that earned OmniMD credit elsewhere in this index for stating plainly that no patient data routes through third party consumer AI services. Not established: what retention applies to processed documents and extracted entities, whether customer documents are used to further train or tune DextractLM, and whether self hosted means hosted in the customer's environment or in 314e's.

A second pass overturns the earlier absence, and the way this vendor handles it is worth noting because almost nobody in this index does it.

The product page states that a business associate agreement is included, and states it as a commercial term alongside no automatic renewal, no unexpected price increases and predictable costs as usage grows. So the agreement is presented as standard rather than as something to be negotiated, and a prospective buyer learns before contact that one comes with the product.

That is a small disclosure with real practical value. This index has recorded many vendors where an agreement plainly exists because enterprise customers demand one, but nothing published tells a smaller buyer whether they will get the same. Placing it in the commercial terms answers that question for everyone.

The supporting posture is described rather than asserted. The product is a document management system handling inbound clinical correspondence, and its material describes role based access control governing who may reach protected health information, extending to control over individual inbound channels, with encryption and audit trails covering access and modification.

What is not published is the agreement itself or its terms, and no subprocessor disclosure was located. For a product that writes structured data into a record system and routes documents onward, the subprocessor position is the natural next question.

Ask for the agreement text, the subprocessor list, and whether terms differ between hospital, revenue cycle and practice customers.

A second pass overturns the earlier absence. The parent company completed a SOC 2 Type II audit, announced with the certification body named and the compliance automation platform used alongside it, and maintains a security policy page describing the result. The report covers the trust services criteria rather than security alone.

Type II is the report that matters, since it tests whether controls operated effectively across a period rather than whether they were designed suitably at a moment. This index has repeatedly found that distinction blurred; here it is stated correctly.

One scoping question follows from how the attestation is described, and it is the right thing to ask rather than assume. The audit is presented as company wide, and the company sells six distinct products spanning training, document management, data archiving, integration, eligibility verification and marketing. A company wide attestation covering that range either has a broad system boundary or a narrower one than the phrasing suggests. Establish whether this product specifically sits inside the audited boundary, and ask for the report and its system description rather than the announcement.

Supporting controls are described concretely on the product's own material: role based access control extending to which users may reach particular inbound channels, encryption in transit and at rest, and audit trails covering access and modification.

No trust centre with a self serve report request path was located, and no additional certification such as the healthcare specific framework was found.

Ask for the report, its system boundary, and the current audit period.

No device authorisation and none apparently required, since document classification and filing are administrative rather than clinical determinations. One point worth recording without overstating it: associating a document with the wrong patient record is a record integrity failure with clinical consequences, and health information management is a regulated function subject to record accuracy obligations under HIPAA and to accreditation standards, so the relevant compliance surface here is records governance rather than device regulation. No position on it is published.

No fairness, subgroup, error distribution or performance variation disclosure was located. The exposure in document processing is technical rather than demographic and it is unreported here as it is across this shape: classification and extraction accuracy vary with scan quality, handwriting, form layout and the sending organisation's equipment, so documents arriving from smaller, older or under resourced practices may be processed systematically worse than those from large well equipped senders. A single aggregate accuracy figure conceals exactly that distribution. No per source quality reporting, error distribution or exception analytics is described.

Materially more disclosure than the other document intake products assessed here, and the part that counts is a number. The company names its model, states that it is purpose built for healthcare rather than general purpose, states that the product runs on self hosted models rather than third party inference interfaces, names its classification technique, and publishes precision figures for both extraction and classification.

Publishing precision for both stages separately is the useful form, because extraction and classification fail differently: a misclassified document goes to the wrong queue and is usually noticed, while a mis extracted value enters a record looking exactly like a correct one. A vendor reporting them apart has understood that. Held below the top grade because the disclosure stops at the names.

There is no model card, no training data description and no test set behind the published precision, so a reader cannot tell what the figures were measured on or whether the corpus resembled their own document mix, and the classification technique is a product name rather than a described method, so an evaluator cannot assess how it works or where it degrades. No warranty, indemnity or remediation commitment attaches. Ask what the precision figures were measured against, performance by document type and image quality band, and what the system does with a document it cannot classify confidently.

Described as integrating with multiple platforms and data sources, with no specific electronic health record named for Dexit anywhere located, no integration method described and no reference deployment cited. The parent company has substantial EHR consulting and integration expertise and sells adjacent interoperability products including a FHIR native archive, which makes the capability plausible, but capability at company level is not the same as a documented integration for this product and this axis grades the latter.

Better disclosed than most of this shape because the company states a hosting posture rather than leaving it implicit: Dexit runs on self hosted models. For an organisation with data residency obligations or a policy against sending clinical documents to external inference services, that is the material fact and it is stated up front.

The remaining ambiguity is what self hosted means in practice, specifically whether models run inside the customer's own environment or inside 314e's infrastructure, and no region, tenancy model or residency commitment is published either way.

No price, tier or pricing unit is published. The company directs prospects to a product page to sign up, which implies a more self serve motion than the enterprise demo gate common in this category, but no rate appears in public materials. The unit question matters for document processing more than for seat based products: per document, per page, per user and per facility produce very different economics for an organisation handling high inbound volume, and nothing indicates which applies.

Scoped to hospital and health system health information management, addressing the 50 to 500 document types a health system handles. Breadth of document type is claimed but breadth of deployment is not evidenced: no named customer, no setting specific reference and no specialty capability is described, and there is no ambulatory, specialty or international coverage claim. The scope is coherent and narrow, and the absence of any named reference site is the limiting factor rather than the scope itself.