Reteta

Ambient capture, note generation, pre charting, after visit summaries, referral letters and the coding output in Bcoder are all model produced. No human scribe tier and no services layer. The company is a product company launched out of a venture studio rather than an existing services business adding AI, and both products are sold as the thing itself rather than as a feature of a platform.

The review gate is described plainly, with clinicians reviewing and signing generated notes before they enter the record, and the explainability design is a real oversight mechanism rather than a slogan. If a clinician can trace any sentence in a note back to the moment in the conversation it came from, verification becomes something they can perform in seconds rather than a re-reading of the whole note against memory. That is the difference between a review step that is checkable and one that depends on recall.

Three things weaken it.

The traceability claim is asserted without a demonstration, an error rate or any description of how the trace is presented at review.

Coding suggestions and documentation-improvement prompts surface in real time during the encounter, with no published confidence threshold and no described behavior when the system is uncertain. A prompt arriving during the consultation shapes what the clinician documents while they are documenting it, which is a stronger form of influence than a suggestion offered after the visit.

And the figure that 80 percent of notes are accepted without edits is presented as a benefit when it also describes how often the review step changes nothing. Turned around, it says four notes in five are signed as generated. Whether that reflects a good draft or a light review is exactly what an acceptance rate cannot distinguish on its own.

Nothing is published about what is under the hood. No model or model family is named, no foundation model provider is identified, no architecture is described beyond natural language understanding as a category, no training data is characterized, and no versioning or update practice is published, so there is no way to know when the system changed or what changed with it.

One property is described rather than named, and it is a real one. The vendor states explainability as an architectural commitment: every entry in a generated note is said to trace back to the point in the conversation it came from. That describes a design behavior without identifying what implements it.

An accuracy figure is a performance disclosure rather than a statement about what produces the output, and the figures here compound the problem rather than relieving it. At least four are published that cannot all describe the same thing: a figure across diverse accents, a homepage figure using the unfalsifiable up-to construction, a proportion of notes accepted without edits, and a separate coding accuracy figure. None carries a method, a denominator or an evaluation design, and one of them sits in a blog post comparing the product with a named competitor, which is marketing rather than technical documentation. A vendor publishing four unreconciled numbers is publishing less information than one publishing a single number with a method.

Nothing identifies any party in the chain. No model or model family, no foundation model provider, no hosting arrangement and no subprocessor list is published. End-to-end encryption is stated for protected health information in transit and at rest, which protects content without naming who processes it.

One capability creates a content type that stands apart, because it is derived rather than captured, and the ordinary analysis of a scribe does not reach it.

The after-visit summary generation reads the patient's language preference and infers their health literacy level in order to adapt what it produces. An inferred health literacy level is an attribute the system assigns to a person. It did not come from the patient, the patient is unlikely to know it exists, and it is the kind of derived characteristic that can persist and travel further than the summary it was created for. If it is written to the record, it becomes part of the patient's chart; if it is retained for personalization, it becomes a profile.

The coding product raises the chain question a second time, since a real-time coding suggestion generated during an encounter is produced somewhere, from something, and neither is named.

Nothing published says which patient attributes the system derives, where they are stored, whether they are written to the record, whether the patient can see or contest them, or whether they are excluded from any model improvement process.

No study, controlled evaluation, third party assessment or named customer is published. Adoption is described only as hundreds of satisfied healthcare providers, with no organization named and no deployment count. Every performance figure is vendor stated without method: two hours saved per clinician per day, SOAP notes in under two minutes, burnout reduced by up to 75 percent, an 85 percent increase in claim approval, 70 percent coding productivity gain.

Presence on the Elion and AVIA marketplaces is not validation: the Elion profile is explicitly marked unclaimed, meaning it was generated by the directory rather than submitted and verified by the vendor.

End-to-end encryption is stated for protected health information, and the company describes safeguarding data in transit and at rest.

Nothing further is published. There is no audio retention schedule, no deletion commitment, no de-identification practice and no statement in either direction on whether customer content is used to train or improve models. Encryption without a retention answer is an ordinary position rather than a strong one: it describes how content is protected while it is held, and says nothing about how long it is held or what else is done with it.

The gap is wider here than for a scribe alone, because the platform holds more kinds of content. Encounter audio, generated notes, pre-charting drawn from the record before the visit, after-visit summaries and coding output all sit with the vendor, and they are unlikely to be governed identically.

One derived attribute stands apart. The after-visit summary generation reads the patient's language preference and infers their health literacy level in order to adapt what it produces. That is a characteristic the system assigns to a person rather than one the patient supplied, and nothing published says what attributes of this kind are derived, where they are stored, whether they are written to the record, or whether they are excluded from any model improvement process.

The certification claim is unconfirmed. A certification under a named third party health privacy compliance program has been attributed to the vendor, and it would name an assessing body that could be asked for evidence. The vendor's current product pages and its own launch announcement give only the general formulation: end-to-end encryption and a health privacy compliant framework, with compliance protocols described rather than attested. The named program does not appear.

A claim that cannot be found now is not the same as one the vendor never made, but it cannot be relied on either. Nothing published says whether any third party compliance certification is held, by which assessor, with what scope or from what date.

What is established is thinner. Compliance is claimed, encryption in transit and at rest is stated, and no business associate agreement terms, availability, scope or tier gating are published anywhere.

The integration surface sharpens the agreement question. This product writes notes, codes and summaries back into hospital record systems, so it handles protected health information for organizations whose own obligations flow down to it, and those organizations will require the agreement before anything is connected.

No SOC 2 report of either type, no ISO 27001 certification, no penetration testing statement, no audit report and no trust center appears anywhere. The vendor describes end-to-end data protection, which is a claim about design rather than an examination of it.

One distinction is worth restating because it is the kind of item that inflates a compliance list. A health privacy rule compliance attestation is a different object from a security controls attestation. It speaks to a documented compliance program rather than to an independent examination of technical and organizational controls. Laws, regulators, conformity marks and control attestations are four different kinds of thing, and a list mixing them reads as more than it contains.

The absence matters in proportion to what the platform holds and who it sells to. This vendor integrates with major hospital record systems through interface standards and writes notes, codes and summaries back into them, so it holds encounter audio, generated documentation and coding output for organizations whose procurement functions expect an examined posture as a matter of routine. Selling into that buyer without one is a harder position than selling to a solo practice without one, and it is the first thing a health system security review will ask for.

None of this is a judgment that controls are absent; it describes what a counterparty can verify before contracting. For a company founded in 2024 with fewer than ten people, the absence is expected; the buyer profile is what makes it consequential.

Ambient documentation needs no clearance, and none is claimed. No United States device pathway attaches to a note the clinician reviews and signs.

What governs is payment integrity, and this product sits at the sharp end of it. The platform pairs an ambient scribe with a dedicated coding product delivering real time coding suggestions. The vendor markets real time coding visibility as helping clinicians maximize reimbursement accuracy, describes compressing the clinical documentation improvement cycle from days to minutes by identifying documentation gaps during the encounter, and claims a large increase in claim approval.

The field has now published on precisely this, which makes the concern a documented pattern. A 2025 policy analysis in a peer reviewed digital medicine journal describes the business case for ambient scribes shifting from clinician wellbeing toward revenue capture through more intensive coding, and reports measured effects: an eleven percent rise in physician work relative value units and a fourteen percent increase in documented hierarchical condition category diagnoses per encounter at one health system, more high level evaluation and management visits billed at another, and an oncology study finding documented diagnoses per encounter rising from three to four point one.

That analysis names the question this vendor faces: whether better coding reflects better documentation of care actually delivered, or more intensive coding of the same care. Identifying gaps during the encounter rather than after the visit is the sharper version, because it shapes what the clinician documents while they are documenting it.

One thing counts in its favor. The vendor describes its scribe as explainable and its coding as traceable to the encounter. A suggestion a clinician can trace back to what was said is reviewable in a way an unexplained one is not, and that is the right direction for a product on this gradient.

Mixed, and both halves matter.

The positives are real and unusual for a company of this size. The vendor publishes an accuracy figure for speech across diverse accents and dialects at 86 percent, and it is not a flattering number, which makes it more credible rather than less. Generating the after-visit summary to the patient's own health literacy level and preferred language is an accessibility design decision rather than a marketing line, and it addresses a failure mode most documentation products ignore. Explainability gives a clinician a route to challenge an output.

The figure is not a subgroup breakdown, and that distinction matters. A single aggregate percentage measured across a diverse sample tells a buyer that the system was tested on varied speech; it does not say how performance differs between one accent and another, which is the disclosure that would let a practice serving a particular population know whether the product works for them. No per-group figures, no method, no sample description and no model card accompany it.

Against that, the coding posture puts the product on the reimbursement gradient. Documentation improvement capabilities are described as identifying documentation gaps during the encounter rather than after the visit, compressing a cycle that normally takes days into minutes, and the outcome sold is an 85 percent increase in claim approval alongside claim ready notes. Prompting at the point of care shapes what the clinician documents while they are documenting it. The framing here sits at the defensive end, denial avoidance rather than maximizing intensity, which is the milder version of the pattern.

The strongest point is traceability. Every entry in a generated note is stated to trace back to the point in the conversation it came from, which is a correction route available at the moment of review rather than a policy commitment. A clinician who can check a sentence against the words that produced it can challenge it in seconds.

The vendor also publishes an accuracy figure for speech across diverse accents at 86 percent, and it is not a flattering number. Publishing an unflattering figure at all is a form of accountability.

What is missing is a published error rate with its method and denominator. At least four accuracy claims are published that cannot all describe the same thing: the accent figure, a homepage figure using the unfalsifiable up-to construction, a proportion of notes accepted without edits, and a coding accuracy figure. No methodology, denominator or evaluation design accompanies any of them, and the accent figure sits in a blog post comparing the product with a named competitor rather than in technical documentation.

Nothing stands behind the output either. No warranty, indemnity or remediation commitment is published, and no route is described for a patient who finds an error.

Described as standards-based and bidirectional: secure interfaces over HL7 and FHIR, writing SOAP notes, coding and summaries back into the record, with Epic, Cerner, athenahealth and eClinicalWorks named, plus access to patient information and appointments from inside the platform. Naming a standard and naming systems is the right shape, and it goes beyond simple transfer, where a note is produced and moved by hand.

Two gaps remain.

The phrase bolts onto any EHR is a universality claim, and universal compatibility is usually a consequence of not integrating rather than of integrating well. The named list ends in and more, with no supported-systems page behind it and no per-system documentation, so nothing shows which connections are built and which are asserted.

And nothing is demonstrated. No customer is named on any specific system, no marketplace listing is cited, and no reference deployment is published, so the integrations are described rather than shown running.

The distinction that matters most in practice is what write-back actually produces. A note landing as discrete coded fields is a different product from one landing as a block of text in a free-text box, and the difference determines whether the coding output is usable downstream or has to be re-entered.

Nothing published shows whether write back lands as discrete coded fields or as a text blob, or which of the named systems are live today.

Delivery is a cloud platform. No hosting region, no residency option, no subprocessor list and no on-premise path is published.

What is better established is the integration mechanism, and it matters because the route a scribe takes into the chart varies more than buyers expect. This product connects through secure interfaces and the standard health interoperability protocols, writing notes, codes and summaries back that way. That is a permissioned integration rather than browser manipulation or robotic process automation driving the record system as a user, so its actions carry their own identity and the record system's own access controls apply to them. Where a product operates under a clinician's credentials instead, the audit trail stops being a reliable record of who did what.

What remains unanswered is where the processing happens. Nothing establishes where transcription and note generation run, whether a third-party model service performs either, which provider that would be, or what it retains. The coding product raises the same question separately, since a real-time suggestion produced during an encounter is generated somewhere.

The named integrations are the largest hospital record systems in the United States. A buyer of that size will have a residency position of its own and a subprocessor review process, so this gap is likely to be closed in contracting rather than in published material. That does not help a buyer comparing vendors before a sales conversation.

No price, unit of pricing, tier structure or trial is published, and no third party estimate appears either. The routes in are a demo request and marketplace listings, neither of which exposes a rate.

More than 15 specialties with specialty-specific templates, spanning ambulatory practice and hospital or health system settings.

The output range is wider than the specialty count suggests and is the real distinguishing feature: SOAP notes, pre-charting before the visit, after-visit summaries, and referral letters to other specialties, which is a document type many scribes never produce at all. Pre-charting and referral generation both sit outside the encounter itself, so the product covers more of the documentation day than a capture-and-write tool.

The after-visit summary is the standout. It is generated with education materials, treatment regimens and caregiver notes tailored to the patient's health literacy level and preferred language, which reaches the caregiver as a reader rather than only the clinician and the payer. Most documentation products treat the after-visit summary as a formatting step; treating it as a comprehension problem is a different design decision and a better one.

Two things fall short. The specialty count is modest and, more importantly, no specialty instrument is named anywhere: templates are described by specialty label rather than by the scores, staging systems or measures that specialty actually records, which is the difference between a template library and domain depth. And no deployment is named in any setting, so the ambulatory and hospital claims are both unevidenced.