Knowtex

Founded by AI researchers to build this product and nothing else. Notably the founders spent a year working as human scribes alongside physicians while building it, which is domain grounding rather than model work but explains why the output targets notes, codes and orders together rather than notes alone.

The product generates diagnostic and billing codes and orders in real time from the conversation, and the company's chief executive describes order entry as the natural extension of ambient documentation and the focus of the 2026 roadmap.

That places this product among the systems that act rather than only propose. Generating an order is not the same act as generating a description of one: an order is executed against a patient, and the question of what stands between generation and execution becomes the whole of the oversight model.

Nothing published answers it. There is no confidence threshold, no order level accuracy rate, no abstention behavior, and no description of a review gate. Oversight is implied by the workflow rather than described as a control.

The reasoning offered publicly deserves examining rather than accepting, because it is an argument by analogy standing in for a measurement. The chief executive's position is that a system accurate enough to generate a note is accurate enough to propose the orders implied by it. That does not follow. A note is reviewed in full by the person who signs it, and an error in prose is visible to a clinician reading their own patient's story. An order is a discrete instruction that may be approved in a batch, and an error in it is a wrong medication or a wrong test rather than a wrong sentence. The consequences of error differ, the review behavior differs, and an accuracy figure for one does not transfer to the other.

No accuracy figure, model card, named models or evaluation methodology is published. Specialty specific tailoring is claimed as an architectural property but not described. A competitive federal evaluation implies performance was measured by an external party, which makes the absence of any published performance figure more conspicuous rather than less.

Nothing public names a party, and there is good reason to believe a full enumeration exists in reviewable form. A federal authorization to operate requires documented system boundaries and a component inventory, and the accompanying privacy assessment addresses permitted use, disclosure and retention directly.

A vendor that has completed that process has answered these questions to a regulator's satisfaction and has published none of the answers, which makes this a disclosure choice rather than an absence of governance, and the material is very likely obtainable on request. What is public is bare: no foundation model provider, model class or version, no hosting arrangement, no subprocessor list, and no statement on whether an external model service processes encounters.

The content involved raises the stakes of that silence. Veterans' records carry service connected conditions, mental health and substance use content at elevated prevalence, and information bearing on benefits entitlement, all of which attract protections beyond the ordinary health privacy rule. Nothing published says which parties handle it.

Strong outside validation, though not clinical outcome evidence. Knowtex was one of 15 finalists in the ambient dictation track of the 2024 Department of Veterans Affairs AI Tech Sprint, a structured competitive evaluation run by the government rather than the vendor, which Knowtex describes as a top three placing. When the VA's planned pilot award to Nuance did not go ahead, Knowtex was selected alongside Abridge, as Fierce Healthcare reported, and deployed from October 2025 across the largest integrated health system in the United States under a contract Knowtex reports at $15 million.

In September 2026 Knowtex told Nextgov it had been selected for the VA's five year enterprise ambient scribe contract, a multi vendor award with a $775.72 million ceiling; the VA had not confirmed its place at the time of reporting. An earlier pilot ran at the University of Rochester Medical Center from February 2023.

No outcome data, accuracy result or peer reviewed study is published. What is evidenced is that a competent outside party assessed the product favorably and kept buying it, not what it achieved clinically.

Audio and transcript retention, deidentification and training use are not addressed in any published statement.

The answers very likely already exist in contract form. A federal health deployment brings obligations beyond the commercial privacy baseline, including federal records requirements, system of records notices governing how a federal agency may hold and use personal information, and the privacy impact assessment that accompanies an authorization to operate. Those documents address retention, permitted use and disclosure directly. A vendor that has been through that process has answered these questions to a regulator's satisfaction and has published none of it.

One question is specific to this deployment. Veterans' health records include service connected conditions, mental health and substance use content at elevated prevalence, and information bearing on benefits entitlement. Content of that kind carries protections beyond the ordinary health privacy rule, and Knowtex does not say what it retains of that content or for how long.

A second follows from a detail in the contracting notice. The system is designed so a provider can initiate a recording without manually entering patient identifying information, meaning the association between an encounter and a patient is established by the system from record context rather than by the clinician. That is a real workflow improvement, and it moves an attribution decision into software. Nothing published says what confirms the recording is attached to the right patient.

The training question is unanswered in either direction and matters more where federal data is involved.

No business associate agreement posture is published.

The federal deployment changes what the agreement question means.

A federal health agency is a covered entity and its vendors are business associates, so an agreement is required. But the agreement is not the whole of the contractual position in that setting. Federal deployment brings its own instruments: the contract itself with its data clauses, a privacy impact assessment, the security assessment supporting an authorization to operate, and the rules governing how a federal agency may hold personal information. Those are the documents that actually constrain what this vendor may do with veterans' encounters, and they are more prescriptive than a commercial agreement.

Terms negotiated for a federal customer do not transfer, and nothing says what a commercial agreement with this vendor contains, whether the retention and use limitations match the federal ones, or whether processing for commercial customers happens in the same environment or a separate one.

The vendor also reports deployment across a large number of commercial health systems alongside the federal contract, so both postures exist in parallel.

There is no named or dated attestation and no trust center, so a counterparty has nothing published to check. The substance behind that gap is unusually favorable.

The vendor holds a federal contract to deploy across the veterans health system, awarded following a competitive evaluation challenge and reported at $15 million, with rollout from October 2025 across a system comprising one hundred and seventy medical centers and more than a thousand outpatient sites. Reporting of the award notes that the selected companies then had to clear additional hurdles before integration, and a senior official described the privacy and consent work involved as substantial.

Deployment at that scale inside a federal health system requires an authorization to operate and the security assessment that produces it. That process examines control implementation against a federal control baseline, is granted by a named authorizing official, and is a materially higher bar than a commercial attestation. It is very unlikely this vendor does not hold one.

So this is a publication gap rather than a control gap, and a consequential one. A vendor holding federal authorization and not saying so is leaving its strongest evidence unused, and every commercial buyer evaluating it is asking for a lesser assurance than the one that already exists.

No clearance is claimed and none is required for documentation, but the platform reaches past documentation.

The company describes its platform as covering the full care journey: pre visit insight and chart summarization, in visit ambient documentation, and post visit specialty specific coding, order automation and clinical decision support. Decision support is named as a current platform capability rather than a roadmap item, and order automation sits alongside it.

Those are three different regulatory conversations. A note the clinician signs attracts no pathway. An order is an action taken on a patient rather than a description of one. Decision support is the category where the exemption depends on a professional being able to review the basis of a recommendation independently, and where intended purpose rather than technical sophistication sets the threshold.

No device claim is made and no clinical efficacy claim is attached to any of it. What is missing is a scoping statement: nothing published describes what the decision support does, what it is grounded in, or what the clinician is shown before acting, so which side of a conditional exemption it sits on is left open.

One piece of evidence comes from the deploying institution rather than the company. A senior official at the federal health system publicly described substantial policy work on informed consent and the protection of veterans' privacy preceding rollout. That is notable because the largest integrated health system in the country treated consent, rather than device classification, as the hard problem in adopting ambient documentation. It is institutional confirmation that consent is a live regulatory question for ambient scribes, from a body with no commercial interest in saying so.

Knowtex publishes no fairness statement, subgroup analysis or accent and dialect performance disclosure.

This gap carries particular weight here, and it is closer to a missed opportunity than an omission.

The veterans health system serves more than nine million people and is the largest integrated health system in the country. Its population differs from the commercial average in ways that bear directly on speech recognition and on documentation quality: an older and predominantly male cohort, high prevalence of hearing loss, traumatic brain injury, post traumatic stress and substance use, distinctive regional distribution including substantial rural enrollment, and clinical vocabulary specific to service connected conditions.

Every one of those characteristics is a plausible source of differential performance, and the deployment makes them measurable. The system has extensive research infrastructure, standardized records across all sites, and a mandate to evaluate what it adopts. A national deployment inside that system also generates the sample size that makes subgroup evidence meaningful rather than anecdotal.

So the question here is not whether the vendor could publish evidence but whether the evaluation is being done and by whom. Deployment evaluations conducted by the health system may exist without the vendor holding or publishing them.

Substantial accountability apparatus surrounds this product, but none of it is published or necessarily portable. The federal deployment brings obligations that do not apply to a commercial scribe: federal records requirements, notices governing how an agency may hold and use personal information, and a privacy impact assessment accompanying an authorization to operate.

For patients inside that system, federal privacy law also gives an individual a route to seek amendment of a record held about them, which is a statutory mechanism rather than a vendor courtesy. A competitive federal evaluation additionally implies that performance was measured by an external party. The problem is that none of it travels and none of it is visible.

No figure from that evaluation is published, no privacy documentation is public, and a commercial buyer inherits none of the federal apparatus, because a federal certification does not transfer. One design detail moves a decision into software without a described control.

A provider can initiate a recording without manually entering patient identifying information, so the association between an encounter and a patient is established by the system from record context rather than by the clinician. Attaching a note to the wrong patient is among the most consequential documentation errors possible, and nothing published describes what confirms the match or what happens when it is wrong.

Positioned as EHR agnostic and deeply integrated into workflow, accessible from mobile, tablet or computer with no dedicated hardware, and designed to fit team based care. The VA deployment is meaningful evidence that integration works at scale in a complex federal record environment. No named commercial EHR integrations, certification program listings or write back architecture are published, so breadth outside the federal deployment is asserted rather than demonstrated.

No hosting region, residency configuration or subprocessor detail is published.

The federal deployment implies a government cloud posture, and that is unconfirmed. Federal health workloads are normally required to run in an authorized environment, physically and logically separated from commercial infrastructure, operated by screened personnel and subject to the boundary defined in the authorization. If this vendor operates such an environment, its answer for federal work is strong and documented.

What does not follow is that a commercial customer receives the same thing. Running a federal enclave is expensive, and vendors commonly operate two environments with different controls, different personnel access and different subprocessors. A federal deployment is not evidence of the commercial posture.

Nothing says which environment serves commercial customers, what boundary and controls apply there, or which of the federal protections carry over.

The model provider question is unanswered and matters here. The vendor is backed by a major cloud provider among its investors, which suggests but does not establish where workloads run, and nothing states whether an external model service processes encounters or what it retains. In a federal enclave that question has a documented answer; commercially it does not.

No rate card, and two public numbers that are easy to misread. Knowtex announced its VA pilot contract at $15 million. That figure comes from the vendor's own announcement rather than a published award notice, and it covers a federal pilot whose scope, support and pricing differ from a private health system agreement, so it says little about commercial pricing. For comparison, the federal award notice for Abridge's pilot under the same program shows $5.4 million.

The second number is the September 2026 enterprise contract, which carries a public ceiling of $775.72 million over five years shared across every vendor that can compete for task orders. A ceiling is the most the VA may spend across all of them, not a price or a commitment to any one vendor.

Nothing published says how Knowtex prices commercial customers, whether by clinician, encounter or module.

Specialty specific by design rather than one model behind templates, with oncology workflows named explicitly, and the VA deployment spans 170 medical centers and 1,193 outpatient clinics covering effectively every specialty and care setting in a national system. That breadth currently rests on a single customer relationship rather than a distributed commercial base, and no specialty count or language coverage is published.