Tennr

Tennr sells the reading of inbound healthcare documents: faxes, scanned forms, portal submissions and handwritten orders arrive, RaeLM, the company's own vision language model, extracts the clinical and administrative data, and the platform evaluates it against payer criteria to say what is covered, what is missing and what should happen next.

Remove the model and nothing remains to sell. What is left is a fax inbox, which is the thing customers already had and were drowning in: the case studies describe staff spending 10 to 25 minutes per order retyping faxes before the platform arrived. Every capability on the product page depends on the extraction underneath it, including triage that ranks orders by clinical and operational context, agentic communications that chase missing documentation, and the autopilot that completes high confidence work without a person.

The company's positioning is consistent with that reading rather than decorative. It argues that general purpose language models cannot read a smudged fax, a checkbox or a handwritten order against payer criteria, and that a purpose built model trained on healthcare documents is required, and it says outright that artificial intelligence is not the point, the completed patient and the paid claim are.

Autonomy is marketed as a named capability and the boundary around it is asserted rather than described.

What is published: the platform completes document work end to end, triage surfaces the orders that need attention using clinical and operational context, workflow orchestration is described as adapting to a customer's staffing models and escalation paths, and Autopilot is described as completing high confidence work automatically while maintaining built in quality control.

What is missing is every number and rule behind those words. Nothing states what confidence threshold sends a document down the automatic path, who sets it, what the quality control actually does, whether a sample of automatic work is reviewed and by whom, what happens to a document the model cannot read, or whether a person ever sees an order that was handled without them. The phrase high confidence implies a threshold exists; the buyer cannot see it or move it from anything published.

The decisive figure is the share of documents completed with no human touch, and it is not published. That number would show how much autonomy is actually running, and how much review staffing remains.

The model is named and its training corpus is quantified three ways: a reported 100 million anonymized healthcare documents, 2.3 billion data fields and 8,000 sets of payer criteria, with the architecture described as a vision language model, RaeLM, plus an orchestration engine over it. Quantifying a corpus on more than one measure is more informative than a headline number, because document count speaks to volume, field count to the density of extraction and criteria sets to payer breadth. Naming the architecture as vision language also tells a buyer which failure mode to test for, which is misreading rather than misreasoning.

The privacy policy adds the provenance: de identified data is used for research and development, to refine the company's algorithms and machine learning applications, and to improve the service. So a buyer can see both what the model is and where a substantial part of its training material comes from.

Nothing deeper is published: no model card, no benchmark methodology, no evaluation results, no accuracy figure by field type or by source quality, and no versioning or update practice, which matters on a model retrained from customer documents because a buyer has no way to learn that the extraction behind its orders has changed.

The chain is structurally short at the layer that matters, and undisclosed everywhere else.

The model is the company's own. RaeLM is presented as a proprietary vision language model built by the founding team out of their own language model research and trained on the company's healthcare document corpus, with the explicit argument that general purpose models cannot do this work. So the component that reads a patient's referral is not a third party service, and there is no quiet swap of an external model behind the extraction.

Everything around it is unnamed. No hosting provider appears, no fax or telephony carrier is named for an ingest surface spanning more than fifty source types, no subprocessor register is published, and the privacy policy establishes only that vendors and service providers exist and that the company uses HIPAA compliant products and services for health data storage and processing. Google Analytics is the single named third party and it belongs to the website rather than the product. The outbound voice capability added in 2026 is the other open question, since speech recognition and synthesis are rarely built in house alongside a document model.

What the policy does disclose is the direction of data: de identified customer documents feed the company's own research, development and model refinement.

Named customers with before and after figures, on top of consistently reported scale.

The customer stories carry the detail. Home Medix reports referral processing dropping from 25 minutes to under 5, quadrupled intake volume and 32 days of staff time saved over six months. NEB Medical reports processing per fax falling from 10 to 15 minutes to under 5, with throughput rising from 50 to 170 faxes a day per user, clearing a two to three week backlog. Further stories name FloMed at four times faster patient processing, Rothman Orthopaedics at twice the processing without added headcount, and Synergy Orthopedics clearing a six week backlog and billing 33 percent more month on month. Customer named quotations describe handling 260 faxes in a single day.

Scale is stated consistently: more than 10 million documents a month across the customer base, with the company reporting more than 150 healthcare organizations and a referral network of around 320,000 referring partners.

None of it is independent, and the method is thin: the comparisons are per order handling times rather than audited outcomes, most carry no period, and there is no peer reviewed or third party evaluation. The figure that matters most, the share of documents completed with no human touch, is not published.

The company publishes a plain statement on secondary use: de identified data is used for research and development of new products, to refine its algorithms and machine learning applications, and to improve the service, and it may be disclosed in that form. Read alongside the claim that RaeLM was trained on more than 100 million anonymized healthcare documents, that tells a buyer where a substantial part of the training corpus comes from.

What is not stated is which de identification standard applies. The privacy policy paraphrases the concept rather than naming the enumerated identifier method or an expert determination, and the two leave different residual risk.

The modality makes that gap harder to close. The product reads scanned faxes, photographed forms and handwritten orders, so identifiers sit in handwriting, letterheads, fax transmission headers, signatures and stamps rather than in labeled fields. De identification here is an image and text extraction problem before it is a policy one, and a redaction process validated on structured records says nothing about a smudged fax header.

Retention is the other gap: the policy commits only to keeping data as long as necessary, with no schedule by data type and no deletion process for documents once an order is complete.

The role is stated and the instrument is not published. The privacy policy carries a section on personal information processed on behalf of healthcare provider customers, says the company processes health data pursuant to its agreements with those customers, and states that it applies HIPAA privacy and security standards to all health data it collects. It names the categories it handles: patient intake forms, referral forms, insurance forms and related documentation.

No business associate agreement, template or scope statement is published, so the posture is visible and the instrument is not, and nothing states whether the agreement binds the subcontractors that carry documents in and out.

One scope question follows from the network product. Referral status visibility is offered to referring providers, receiving providers and patients, which means information moves between organizations that are separate covered entities and are not party to each other's agreements with the vendor. Nothing says what each participant can see about a referral originating elsewhere, under whose authority that disclosure is made, or what the patient facing view exposes.

Controls are asserted and nothing independent stands behind them. The site's only legal or assurance document is the privacy policy, reached from the footer, and it describes encryption in transit and at rest to industry best practice, a combination of technical, administrative and physical safeguards, and internal access control policies. It is candid that no transmission is ever fully secure and says the customer uses the service at its own risk.

No attestation appears anywhere: no SOC 2 report, no HITRUST certification, no ISO 27001, no trust center, no penetration testing statement, no vulnerability disclosure policy and no security documentation offered under agreement.

One phrase is doing more work than it looks. The company says it uses HIPAA compliant products and services for health data storage and processing, which describes the posture of its infrastructure suppliers rather than an examination of its own. A vendor can run on compliant infrastructure and still have no independent assessment of how it configures and operates it, and that gap is where incidents happen.

At this volume, more than 10 million documents a month for over 150 organizations, including handwritten clinical orders, larger customers have very likely asked for an attestation; none is published.

No device pathway applies and none is claimed. Reading inbound documents, extracting data and checking it against payer criteria is administrative work with no diagnostic or treatment decision surface, and the company says so in its own terms, arguing that the point is a completed patient and a claim that is not denied.

Other regimes reach the functions sold, and no position is published on any of them. Prior authorization work sits under the federal interoperability and prior authorization requirements and under the state laws now conditioning artificial intelligence involvement in coverage processes. The outbound voice capability reaches the telephone consent framework, where a synthetic voice calling a patient requires prior express consent, and nothing published says whether the agent identifies itself or how consent is checked before it dials.

Evaluating a referral against payer criteria to predict a denial is not a coverage determination. It does shape what gets submitted and in what form, which sits closer to the coverage process than pure document handling. Nothing says whether any workflow can end with a referral abandoned rather than corrected, which would be a practical denial reached without a payer involved.

Governance appears as a product phrase and nothing sits behind it. The autopilot is described as quality controlled with built in quality control, which is a commitment in language rather than an evaluation: no framework is named, no monitoring metric is published, no error or overturn rate is reported, and no bias or fairness work appears anywhere.

Two mechanisms make this concrete for this product rather than abstract.

Extraction accuracy on scanned and handwritten material is not uniform. It varies with fax quality, form design, handwriting legibility and whether a practice sends clean digital output or a photographed page, and those correlate with how well resourced the sending practice is. An accuracy gradient by document quality is therefore an accuracy gradient by where the referral came from, and the patients affected are those referred from the least resourced settings. Nothing published reports accuracy by source quality or by sending organization type.

The criteria evaluation is the second. Flagging which orders are likely to be denied, and triaging which cases move first, shapes what proceeds smoothly and what stalls before any payer has decided anything. Nothing states whether flag rates or triage outcomes are monitored by payer, plan type or referral source, or who reviews the criteria mapping when a payer changes its rules.

The product is built so that a person catches most mistakes, and nothing stands behind the ones that get through.

The mechanisms are real. Autopilot is described as completing only high confidence work automatically, with the rest surfaced to staff, triage ranks what needs attention first, the platform flags what documentation is missing before submission, and the network gives referring providers, receiving providers and patients a shared view of where a referral stands, which is itself a way for the affected party to see that something has stalled. So a person reviews anything the model is unsure about before it goes further.

What does not exist is any commitment. No accuracy warranty, no service level, no indemnity, no remediation obligation and no error rate appears anywhere, and the privacy policy states plainly that the customer uses the service at its own risk. The characteristic failure of a document extraction product is not an obvious error but a plausible value in the right field that nobody checks, and a downstream system treats it as read: a wrong quantity on an equipment order, a misread date of service, a diagnosis code that came off a handwritten line. Nothing published says what happens then, who carries the cost of a denial that followed, or whether an extracted value is marked as machine derived when it reaches a person.

Integration is claimed in both directions and no system is named anywhere a buyer can check.

What is claimed is substantial. Orders arrive through more than fifty kinds of electronic fax, phone, email and portal source, and the company says it turns them into structured records directly in the customer's record system, retrieves missing documentation automatically from health information exchanges and connected record systems, and writes changes back wherever charts and orders need to reflect them. That is the right design for this product, because the value collapses if a structured order has to be rekeyed at the end.

What is missing is every particular. No record system, practice management system or pharmacy system is named, no interface standard is described, no marketplace or validated listing is cited, and no customer story ties the platform to a named system. The health information exchange claim is the same shape: exchanges are a real source of clinical records, and none is named.

That leaves the strength on the intake side, where the breadth is real and the sources are the ones this work actually arrives on, and the verification on the write back side absent.

Nothing published says where the platform runs or where documents rest. No hosting provider, region, tenancy model, residency commitment, backup or continuity position appears on the site, and the privacy policy, which is the only legal document published, says only that the company uses HIPAA compliant products and services for health data storage and processing. That establishes that third parties hold the data without naming any of them or saying where they are.

The ingest surface makes the omission consequential. Documents arrive from more than fifty kinds of electronic fax, phone, email and portal source, and fax and telephony carriage is usually operated by someone other than the vendor, so clinical orders pass through infrastructure the customer has never assessed before they reach the platform at all. Nothing describes that path or how long a document persists at each hop.

Retention compounds it: the policy commits only to holding data as long as necessary, with no schedule and no deletion process for source documents once an order is complete. For a product whose corpus is source documents, where they live and how long they stay is the question.

No price is published and the shape of the offer is discoverable. The product is sold as an enterprise agreement through a demo request, with a downloadable product catalog and separate pages for each buyer segment, so a prospect can see which capabilities a project would involve before contacting sales: document intake and extraction, eligibility verification, prior authorization, order processing, communications and the autopilot layer on top.

What no page answers is the unit of charge, and on a document processing product the candidates diverge sharply. Per document or per referral, per user, per location, per order completed, or a platform fee would each produce a different bill at the volumes the customer stories describe, where one site moves from 50 to 170 faxes a day per user. Whether the autopilot is priced separately from the reading, and whether the network and voice capabilities carry their own charges, is also unstated.

The customer evidence makes the question concrete rather than academic: the stories are told as headcount avoided, so a buyer comparing the platform to hiring intake staff needs a number on this side of the comparison. No implementation fee, contract term or minimum appears.

Coverage is enumerated by function and by buyer, with named customers behind part of it.

Six buyer segments have their own pages: durable medical equipment suppliers and manufacturers, infusion providers, fertility clinics, physician groups, hospitals and health systems, and specialty pharmacies. The functions are named rather than implied: referral and document intake from fax, phone, email and portals, patient intake and registration, eligibility and benefits verification, prior authorization, order processing through to delivery, retrieval of missing records from exchanges and referring providers, outbound communications by phone and message, and a referral network giving referring providers, receiving providers and patients shared status.

Validation sits behind the equipment and specialty practice end, which is where the customer stories are: two durable medical equipment suppliers, two orthopedic groups, a sleep management business and a complex rehabilitation pharmacy, each with its own before and after account.

The span reaches beyond where the evidence sits. Hospitals, health systems and specialty pharmacy each have a page and none has a named deployment, and the referral criteria work is inherently payer and service line specific, so nothing shows what is production proven outside equipment and specialty practice.