DeepScribe

Ambient capture and specialty tuned note generation is the product, extended by AI coding and pre visit intelligence. All model driven.

The oversight model is the category standard and is stated unambiguously, which is what earns the grade. DeepScribe generates documentation directly within the electronic health record for a clinician to review and sign off, so a licensed human sits between the model output and the medical record, and the signature is the moment of clinical and legal accountability.

The product is described as fully ambient, requiring no dictation and no wake words, which removes the clinician from the capture loop entirely and makes the review step the only human checkpoint in the workflow. That is worth stating plainly rather than treating as neutral. Third party review material also reports that the company previously promoted a human quality assurance review layer which has since been de emphasised.

That is recorded as reported rather than established, and is worth confirming with the vendor, because a move from human in the loop quality assurance to fully automated output is a genuine change in the oversight model. Held below a higher grade because there is no published accuracy, hallucination or omission rate, nothing stating what the system does when audio is poor or speakers overlap, and no indication to the reviewing clinician of which parts of a note are low confidence. A reviewer who is not told where the uncertainty sits is exposed to the automation bias that fluent, usually correct drafts reliably produce.

What is disclosed is the infrastructure layer, not the model. A February 2024 announcement names AWS SageMaker and AWS Bedrock as the systems supporting continuous training and monitoring, which tells a buyer the model platform and the cloud dependency.

That is a partial and real disclosure, but materially weaker than naming the model provider, because Bedrock is a hosting and access layer fronting many different foundation models: knowing a vendor uses it does not establish which model is producing clinical documentation. Abridge, in the same category, names OpenAI outright on a public subprocessor list. Everything else on this axis is marketing vocabulary without technical content.

The product is described as built on the largest database of natural patient conversations in the world, trained on more than two million patient encounters, and using composable AI, a vendor coined term that is nowhere defined. There is no model card, no architecture description, no version or update policy, no statement of what is proprietary versus a wrapped third party model, and no published accuracy, error or benchmark figure of any kind. That last absence decides the grade, because the clinical evidence on this record is strong: the outcomes are documented while the system producing them is not characterised at all.

The infrastructure layer is named and the model layer is not, and the distinction between those two is the reusable lesson from this record. A 2024 announcement names Amazon Web Services SageMaker and Bedrock as the systems supporting continuous training and monitoring.

That is a real disclosure and it tells a buyer where the workload runs, but Bedrock is a model access layer fronting many different foundation models from several providers, so knowing a vendor builds on it establishes the platform and not the model. Naming a model access layer is not naming a model, and a buyer reading a press release should not treat the two as equivalent. Nothing else fills the gap.

No foundation model provider, model class, version or update policy is published, no sub processor list was located, and the descriptive language around the technology is marketing rather than technical, with a vendor coined term for the architecture that is nowhere defined and a training corpus described by size rather than by composition or source. The training question compounds it.

The company's own announcement describes continuous training and monitoring as an ongoing activity, while nothing states whether customer encounters or transcripts are among the data used, so a buyer cannot tell whether their own content is in the loop that keeps the model current. Ask which foundation models are invoked through the access layer, and whether customer encounters feed the continuous training described.

The strongest independent rating in the ambient documentation category: a 98.8 KLAS Emerging Company Spotlight score with A+ marks across all six measured categories, plus 2025 KLAS Top 5 rankings across outcomes, patient experience, and clinician experience. KLAS is buyer survey based rather than vendor asserted.

Supported by an August 2025 study in Journal of Clinical Pathways reporting improvements in diagnosis capture and note quality in oncology, which is specialty specific evidence rather than a general claim.

The retention clause is the finding, and it is the weakest recorded in this index for a product that records patient encounters. The published privacy policy states that the company will keep personal data for as long as it remains necessary for the identified purpose or as required by law, and that this may extend beyond the termination of the relationship. There is no period, no ceiling, no deletion commitment and no separate treatment of encounter audio.

Set that against the alternatives now on record in the same category: Avo commits in writing that patient clinical data is processed in real time only and never stored, Nabla publishes a full deletion lifecycle including backup expiry, and Abridge publishes certificates of destruction as a catalogued control. A retention clause that outlives the contract is a materially different proposition. A second gap compounds it.

A February 2024 announcement describes continuous AI training and monitoring built on AWS SageMaker and Bedrock, so training is an ongoing activity by the company's own account, yet nothing states whether customer encounters or transcripts are among the data used. Genuine credits are on the record: de identification of patient data is claimed, alongside encryption in transit and at rest, multi factor authentication and limited user access. The audio retention period and the model training terms are both worth obtaining in writing.

Stronger than the security position, and the difference is worth separating because the two are often conflated. The security practices page carries a dedicated vendor management and business associates section describing the selection and assessment of vendors and business associates, due diligence, and contractual agreements to safeguard the security and confidentiality of protected health information when working with third parties.

HIPAA adherence is stated explicitly by name rather than as vague regulatory alignment, and a dedicated compliance contact address is published, which signals a compliance function that is reachable before a sales conversation. A business associate agreement demonstrably exists and is referenced in third party vendor assessment material, and the product is distributed through the AWS Marketplace, a channel with its own contractual baseline.

Held below a higher grade because the instrument itself is not public: no terms, tier, execution path or signing entity is published, so a buyer cannot establish the shape of the agreement before engaging. One internal tension is worth raising with the vendor: the HIPAA posture is asserted confidently while the retention clause in the same company's privacy policy is open ended and may outlive the relationship. Those are not contradictory in law, but together they mean the protection is contractual rather than architectural.

A dedicated security practices page exists, with a named compliance contact address, which is more than several competitors offer. But it states that the company holds relevant certifications without naming a single one. SOC 2 appears in a February 2024 company announcement describing data privacy protocols achieved through SOC 2, again without specifying the type, and the type is the substantive question: Type 1 tests control design at a point in time, while Type 2 tests whether controls operated effectively across a period.

Third party directories and competitor comparison pages assert SOC 2 Type II, but those are not vendor disclosure and are not credited here. No trust centre, ISO 27001, HITRUST, penetration testing statement or subprocessor list was located. What is substantively described is real: end to end encryption, multi factor authentication, limited user access, de identification of patient data, and a vendor and business associate due diligence process.

The contrast within the same category is instructive, since Abridge publishes a trust centre naming SOC 2 Type 2 with its scope, along with TX-RAMP, accessibility conformance and a full subprocessor list.

No FDA clearance, none claimed and none required for ambient documentation. As with other scribes, the base position is simple: the system drafts, a licensed clinician reviews and signs, and the signature is the control point, so the clinical decision support exclusion debate that dominates records like UpToDate, OpenEvidence and Glass Health does not arise for the core product. What does arise is scope, and here it is not a forecast but a shipped feature set.

The company now describes itself as an Ambient Operating System rather than a scribe, and its published capabilities include automated E/M coding, HCC and ICD-10 coding, AI pre charting, and real time clinical insights. Its own marketing states that the platform maximises reimbursements and increases revenue. That changes the object under review in two ways a buyer should separate.

A system proposing the codes a note will be billed under shapes reimbursement, which creates an incentive gradient note drafting does not have, and HCC coding in particular drives risk adjusted payment and carries documented regulatory and audit exposure. Separately, real time clinical insights delivered during an encounter is a different function from retrospective documentation and moves toward the decision support boundary. Worth establishing which modules are in scope at the version being bought, and asking for the regulatory rationale for the insights feature specifically.

A governance artefact exists and is more than nothing. A Trust and Safety Suite was announced in February 2024 with a dedicated page, framed around system safety, transparency, data privacy and human centric values, and claiming continuous AI training and monitoring for accuracy and bias mitigation built on AWS SageMaker and Bedrock, alongside a stated collaboration with Covenant HealthCare to align the technology with real world clinical needs.

Naming bias mitigation as an explicit objective puts this above vendors that never raise it. But it is assertion without method or result. There is no bias evaluation output, no demographic or subgroup performance analysis, no error taxonomy, no monitoring thresholds and no description of how bias is measured or what would constitute a failure.

The exposure is specific rather than theoretical for this product class: speech recognition accuracy varies measurably across accent, dialect, speech rate and vocal characteristics, and the affected populations overlap with those already underserved. A vendor claiming bias mitigation on a system transcribing patient speech should be able to publish performance by speaker group. None was found. The announcement is also now more than two years old with no published follow through, which is worth noting when weighing the commitment.

The published terms of use contain two clauses that pull in opposite directions, and reading them together is the whole assessment. The first is a broad disclaimer: the company does not warrant or guarantee that the service will provide accurate or informative recordings, nor that information obtained through it will be accurate, relevant, complete or current, and the customer is made solely responsible for verifying accuracy, for making diagnostic and clinical decisions, and for compliance with the laws around them.

The second is a genuine warranty with a remedy attached, which is more than several peers in this lane offer. The company represents that the service will be provided in a professional and workmanlike manner consistent with general industry standards and will conform in all material respects to the specifications and functionality in the documentation, and where it fails to do so it commits, at its own expense, to use commercially reasonable efforts to correct the non conformity.

A stated remedy is what turns a warranty from a sentence into an obligation, and a buyer should note that many vendors publish the warranty and omit the remedy. The reason this stays mid band rather than rising is what the warranty covers. It guarantees that the product behaves as documented, not that the note is correct, and the same document expressly disclaims accuracy.

So the vendor stands behind the product doing what the manual says and does not stand behind the clinical content it produces. Those are different promises and the second is the one a health system is exposed to. One practical caution for anyone verifying this: a separate and unrelated company operates under a similar name at a different domain and publishes its own terms. Read the terms at the vendor's own domain.

Genuinely bidirectional rather than one way push: reads problems, medications, labs, and prior notes before the encounter and writes structured data back into discrete EHR fields after, syncing against the EHR schedule, across Epic, athenahealth, NextGen, and eClinicalWorks. Reading the chart before the visit is what enables the longitudinal note, and it is a materially deeper integration than transcript only scribes.

Delivery is clear and the channels are credentialed. The product is listed on the AWS Marketplace, a real third party distribution channel with its own baseline requirements, and runs as an iOS application requiring no additional hardware or software, which is a materially lower deployment burden than cart or ambient hardware models elsewhere in this category.

Electronic health record integration is bidirectional, with Epic, athenahealth and eClinicalWorks named, and documentation written directly into the record rather than copied across. Infrastructure is disclosed at the platform level through the named AWS dependency. Two things hold it below a higher grade, and both concern what a buyer cannot establish.

No data residency statement of any kind was located, no region named and no residency commitment offered, which matters for a product retaining recorded encounters under an open ended retention clause. And the installed base is asserted rather than evidenced: thousands of clinicians and some of the largest health systems in the United States are claimed, but Covenant HealthCare is effectively the only customer named in public material, where Abridge names more than twenty. No published implementation timeline, uptime commitment or support model was located either.

No published pricing of any kind. Access requires booking a demo, and the vendor publishes no rate, tier structure, per clinician figure or self serve path. Third party estimates exist and disagree widely, which is itself evidence of a fully gated model: roughly 350 to 500 dollars per provider per month from one review site, approximately 750 dollars per month from another, and a 2,000 to 12,000 dollar annual compliance and support band from a third.

A spread that wide across aggregators is what full gating produces, and none of it is vendor disclosure. The asymmetry is the notable part, and it is well developed here. The company publishes specific and quantified return claims: chart closure times as low as 1.6 minutes, up to a 75 percent reduction in documentation time, the potential to see two additional patients per day, and explicit statements that the platform increases revenue and reimbursement.

Every number on the benefit side is published and every number on the cost side is withheld, so a buyer is given the return on an investment whose size is not disclosed. The category contrast shows this is a choice rather than a constraint, since several ambient scribe competitors publish self serve monthly rates openly.

The clearest specialty positioning in the lane, stated as a deliberate focus rather than breadth: oncology first with a reported 40 percent of US cancer care visits, plus tuned models for cardiology, urology, gastroenterology, orthopedics, neurology, and nephrology, and more than 110 languages. The company is explicit that the fit is complex longitudinal specialty care rather than short low acuity visits.