Sayvant

Ambient generation is the whole product and the company exists only to do it. Note for completeness rather than as a deduction: the underlying generative capability runs on Microsoft Azure OpenAI Service, so the differentiated work is the emergency medicine specific reasoning, note structure and clinical content layered on top rather than a foundation model built in house.

Draft and review, with the platform generating charts that providers review before signing, and vendor materials describing preservation of clinician control as a design goal. The HITLAB evaluation assessed error prevention and system visibility among its criteria, an external look at the interaction design. No acceptance rate, edit burden figure or confidence threshold is published, and the oversight question is sharper here given what the system is designed to construct.

Names its own infrastructure: Sayvant states it is built on Microsoft Azure OpenAI Service, so the underlying model family and its known behavior can be reasoned about rather than treated as a black box of proprietary AI. It also describes its design philosophy explicitly rather than in adjectives, stating that most ambient tools capture what was said while Sayvant generates what the note needs to reflect. No accuracy figure, model card, hallucination rate or evaluation methodology is published.

The model layer is named: Microsoft Azure OpenAI Service, stated by the vendor's own executives in a published customer account of the build. Naming it means known model behavior, published limitations and the provider's own data commitments can be weighed, rather than the system being treated as proprietary artificial intelligence.

A second party is named in the chain. Since June 2026 the platform queries the BMJ clinical intelligence knowledge graph in real time during an encounter to surface ranked differential diagnoses, interaction warnings and protocol prompts. That is a third party receiving encounter derived context during care, not merely a reference source consulted later, and what passes to it is not described.

The naming stops at those two. No subprocessor list is published, no model version or update policy is stated, and the retention and training terms that govern content once it reaches the model service are not given. The customer agreement adds one relevant term rather than an answer: de identified data under the health privacy rule is stated not to be the subscriber's data, and the vendor may use subscriber data in aggregated form.

The setting makes the retention question sharper. This is an acute care product supporting a clinician running several patients concurrently across multiple sessions, so the volume and sensitivity of content moving through the chain is high.

An ambitious study design, limited by who ran it rather than how it was built.

Published June 2026, it analyzed over 250,000 encounters across 50 emergency departments against a historical baseline of more than 20 million encounters spanning three years, across academic medical centers, community hospitals, rural departments and freestanding facilities. It takes patient throughput and professional fee capture as primary outcomes rather than charting time, which is a better question than charting time alone. Reported results are a 2.5 percent increase in professional fee charges per encounter for physicians and 3.7 percent for advanced practice providers, with time to provider falling 5.9 and 9.5 percent respectively.

The design is not the limitation. The study was conducted and published by the vendor, released through a press distribution service rather than a journal, not peer reviewed, and with outcome measures chosen by the party selling the product. A charge capture increase is also the outcome that most needs scrutiny, since the mechanism the vendor describes, closing the gap between care delivered and what the note reflects, is indistinguishable from the outside from more intensive coding.

Independent work exists alongside it: a HITLAB heuristic usability evaluation with named academic authors, and a KLAS Emerging Company Spotlight reporting a 94.5 performance score from interviews with 14 people across 10 provider organizations. Both are real third party assessments and neither measures documentation accuracy.

One published term limits the evidence base. The customer agreement bars subscribers from disclosing comparative results of the platform's operation against other products. Independent comparative evidence is therefore contractually constrained as well as scarce.

Audio and transcript retention, deidentification and training use go unaddressed in anything published. The security page is substantive on controls and silent on stewardship, and those are separate matters: how well data is protected says nothing about how much of it is kept or what else it is used for.

Two features of the acute care setting make the omissions more pointed here than for a clinic product.

The environment is one the vendor itself identifies as uniquely hard for ambient capture. An emergency department is loud, and a clinician may be running several patients at once, returning to each across multiple sessions that accumulate into one chart. The vendor supports that concurrency deliberately. It also means several patients' recorded speech is open simultaneously under one clinician's session, which raises an attribution question of the same kind seen in institutional care: what prevents content from one encounter reaching another patient's chart, and what does the clinician confirm about which chart they are adding to.

And emergency departments treat patients who cannot consent. Unconscious, intoxicated, confused or acutely distressed patients are recorded in the ordinary course of the work, and psychiatric presentations and intoxication are documented in the same note. Nothing published addresses how consent is handled where it cannot be obtained, or whether retention differs for those encounters.

A third party account states the platform runs on a named cloud and model provider. The vendor's own material does not confirm it, but it makes the subprocessor retention question concrete.

Compliance with the administrative and technical controls required by the health privacy and health technology statutes is stated, which describes safeguards rather than the contractual relationship with the customer.

The governing customer agreement is published in full and settles the question rather than leaving it open. The master subscription agreement, published openly and last updated September 2024, governs access to the platform for every subscriber. It contains no business associate provisions: no statement of business associate status, no reference to a business associate agreement, no permitted-use or safeguard terms for protected health information, and no breach notification obligation. Its single mention of the health privacy rule is a clause stating that de-identified data under that rule does not constitute the subscriber's data.

So the position is a compliance claim without the underlying document, on a platform the vendor states is live across more than a hundred hospitals. Agreements plainly exist and are negotiated privately through order documentation; what a prospective buyer can read before contracting is an agreement that does not mention the relationship.

The dates raise one currency question. The agreement predates the product's published expansion. It was last updated in September 2024; since then the vendor has extended from emergency documentation into hospital medicine, case management and clinical documentation improvement, added admission-level and critical-care advisories, and integrated a third-party clinical knowledge graph that surfaces differential diagnoses inside the encounter. An agreement written for emergency department documentation may not anticipate any of that.

A second follows from the partnership model. The vendor describes working with revenue cycle and documentation improvement partners to deliver recommendations at the point of care. Where a partner's system contributes to or receives content derived from an encounter, nothing published says whether that partner is a subprocessor under this vendor's agreement or a separate counterparty.

This is a substantive security disclosure, built around a stated attestation.

The vendor publishes a dedicated privacy and security page stating an AICPA SOC 2 Type 2 certification alongside regular internal and third party audits, continuous automated security testing including static and dynamic analysis and vulnerability scanning, and third party penetration testing.

Two things lift this above a certification claim. Continuous testing and independent penetration testing are ongoing activities rather than a point in time examination, and naming them separately from the attestation shows the vendor understands they answer different questions. And the vendor states its security, privacy and risk programs align with the United States national framework for AI risk management alongside the cybersecurity framework. That first reference matters, because the AI risk framework addresses the failure modes an attestation does not reach: model behavior, evaluation, monitoring for drift and the governance of systems whose output changes over time. A SOC 2 report examines whether controls operated; it does not ask whether a model still performs as it did at deployment.

Access control is also concrete rather than described. The platform integrates with hospital identity providers for single sign on, so authentication runs on the institution's own credentials and deprovisioning follows the institution's process.

What is missing is specificity: the report date and audit period, a scope statement, and evidence for the alignment claim rather than assertion, since alignment with a framework is a self assessment unless someone has examined it.

No clearance claimed and none required for documentation. The regulatory context is billing rather than device, with the 2023 revisions to medical decision making documentation requirements as the stated driver.

Two described capabilities push this product away from plain documentation.

The first is inferred reasoning. The company's chief executive describes the product as going beyond transcription to write out the actual thinking of the physician, including when the physician did not articulate it explicitly, naming differential diagnosis and why a course of treatment was chosen or rejected, and connects this to charge capture accuracy. A product that generates the written justification alongside the billing code it selects removes the justification's function as an independent check. This goes further: the system infers the clinical reasoning itself, records it as the physician's, and the stated value is that documented reasoning supports the billing level, because decision making complexity is what determines that level.

The consequence sits with the clinician rather than the vendor. A physician signing that note is attesting to a differential they may not have formed and to alternatives they may not have weighed. If the reasoning is accurate it is a real improvement, since unarticulated thinking is real and goes undocumented constantly. If it is inferred incorrectly the record contains a fabricated account of the physician's judgment, attributed to them, in the document a payer audit or a malpractice claim will examine.

The second is newer and the vendor markets it as such. Since June 2026 the platform surfaces ranked differential diagnoses, drug disease interaction warnings and protocol prompts specific to the presentation, drawn from an integrated clinical knowledge graph, alongside admission level advisories and critical care prompts. The vendor's own announcement frames this as ambient documentation evolving into clinical decision support rather than as documentation.

The marketing describes decision support, and no device position is published alongside it.

Sayvant states its billing incentive plainly. Its clinical AI lead describes an emergency department note as a structured argument for medical necessity and says that where most ambient tools capture what was said, Sayvant generates what the note needs to reflect. Its own headline study takes professional fee capture as a primary outcome, and its chief executive summarizes the result as charges up and wait times down.

A documentation system explicitly optimized to construct the justification for billing level, and marketed on the resulting revenue lift, carries an upcoding gradient, because the tool and the clinician now share an interest in a more thoroughly justified chart. Saying this out loud is at least candid. No fairness, subgroup or accent disclosure is published, despite support for more than 30 languages.

Mechanisms exist. The platform is built around review before signature, with the clinician accepting or declining suggested language rather than receiving a finished note. The agreement defines a unique electronic identification attached to every action taken on the platform, which is an audit trail. And recommendations drawn from the integrated clinical knowledge graph are described as traceable to the underlying evidence.

Nothing stands behind the output. The agreement warrants only that the platform will substantially conform to its own documentation, disclaims all other warranties, caps aggregate liability at twelve months of subscription fees, and bars any claim brought more than a year after it accrues. Indemnification runs one way on the clinical question: the subscriber indemnifies the vendor against claims arising from the provision of medical care by its own clinicians, while the vendor's indemnity covers intellectual property only. No accuracy figure, hallucination rate or limitation is published.

And there is no route for the harmed party. The patient is not a party to the agreement and no mechanism is described for them.

Two features of the acute setting create exposures with no described control. The first is attribution under concurrency: a clinician may run several patients at once, returning to each across sessions that accumulate into one chart, and the vendor supports that deliberately. Several patients' recorded speech is open simultaneously under one clinician's session, and nothing published describes what prevents content from one encounter reaching another patient's chart. Documentation attached to the wrong patient in an emergency department is among the most consequential errors available.

The second is consent where it cannot be given. Emergency departments treat patients who are unconscious, intoxicated, confused or acutely distressed, and those encounters are recorded in the ordinary course of the work, with psychiatric presentation and intoxication documented in the same note. Nothing published addresses how consent is handled where it cannot be obtained, whether retention differs, or what a patient is told later. The recorded person in that situation is often unable to object at the time and has no route to raise it later.

Integration is documented, and for a product live at seventy sites that is the first question that matters.

The vendor has a standards based integration with the major record system built on the open application interface specification for health data, recognized through that vendor's own partner showroom program, and states integrations with three further major record systems. Access runs through hospital identity providers by single sign on.

That matters because of what it is not. Some competing products reach the chart by driving the user interface, through browser extensions or robotic process automation, which means their actions inherit a clinician's session and appear in the audit log as that clinician typing. A standards based integration is permissioned, carries its own identity, and is constrained by the access controls the institution already operates. It is also reviewable: an institution can see what scopes were granted.

Two limits remain. The depth of write back is not described per system, so whether the note lands in structured fields or as narrative text is unstated and is likely to differ across the four. And integration is described as configured at each clinical site, which suggests per site work rather than a uniform capability, so what a given hospital gets may not match what the partner listing implies.

Partial. The hosting substrate is disclosed, running on Microsoft Azure OpenAI Service, which lets a buyer inherit Azure's regional and compliance story as a starting point. Nothing further is published on data residency configuration, regional deployment or subprocessors.

No published rate card. Sold to emergency department groups and hospital systems. The business case Sayvant itself advances is revenue rather than cost: its published outcomes are charge capture and throughput, so the return argument rests on professional fee capture increasing rather than on documentation labor falling, which is a materially different justification to take to a compliance committee.

Deliberately narrow, and the narrowness is the product rather than a limitation. Coverage runs across the acute care spectrum from freestanding emergency departments to Level 1 trauma centers, with urgent care alongside, and support for more than 30 languages. Emergency medicine imposes its own constraints, including encounter pace, disposition decisions and a documentation standard tied to medical decision making complexity. The product is built for emergency departments and is not designed for scheduled outpatient visits.