Suki

Both halves are model driven: ambient capture for documentation, and a voice command layer that interprets natural language instructions to order, navigate, stage, and query. The command layer is what distinguishes the product from passive scribes, and it is entirely an AI capability.

The autonomy ceiling is set sensibly in the product design: voice commands stage orders for clinician review rather than submitting them, which keeps a human decision between the model and any clinical action. Held back from A because escalation behavior when a command is misheard, and the accuracy rate for command interpretation, are not published, and a misinterpreted medication order is a materially higher consequence failure than a flawed note.

No model is named, and there is no model class, architecture description, model card, version or update policy, or published accuracy, error or benchmark figure. Neither a foundation model provider nor a hosting platform is disclosed, which places this behind competitors that name their model provider outright or at least identify the platform layer. The one substantive statement about the technology concerns training data, and it raises more than it settles.

The company states that AI is only as good as the data it is trained on, and that its electronic health record integrations and industry partnerships give it the ability to leverage an extensive dataset to produce the most applicable output. That is a claim that scale of record access improves the model, made without stating whether customer data or patient encounters form part of what is learned from.

Set against it, the developer security documentation states that voice is used only for transcription and summarisation, that recordings are de identified by chunking, and that protected health information is removed from text transcripts, which points toward constrained use of encounter data. Those two statements are not incompatible, but they are not reconciled anywhere public. Whether de identified transcripts are used for model training and improvement is the question to put in writing.

A subprocessor route exists and its contents are not public, which is the middle position and worth distinguishing from having nothing. The trust portal carries a subprocessor page, so the vendor maintains the artifact and a buyer in diligence will presumably receive it, but the portal renders its content behind the platform rather than as a public page and two passes returned no enumerated party.

Nothing else names the chain either: no foundation model provider, no model class or version, and no hosting platform disclosed. What is published is data handling rather than supply chain. Developer security documentation states that voice is used only for transcription and summarisation, that recordings are de identified by chunking, and that protected health information is removed from text transcripts, which describes how content is treated without saying who treats it.

One unresolved tension sits directly on this axis and should be settled in writing. The company markets that its record integrations and industry partnerships let it leverage an extensive dataset to improve output, while the developer documentation says voice is used only for transcription and summarisation.

Both can be true if the training corpus is de identified record derived text rather than audio, but nothing states that, and the question of which parties hold that corpus is exactly what an enumerated subprocessor list would answer. Ask for the list in writing and ask which entries process encounter content.

No peer reviewed study, controlled evaluation or published accuracy or time savings result was located, which separates this record from its two nearest comparison partners, both of which carry substantially stronger evidence positions. What exists is deployment scale and process.

The largest named reference is an expanded partnership with MedStar Health, a system reported at 7.7 billion dollars with more than 300 care locations, putting the product in front of thousands of clinicians across ambulatory specialties, alongside partnerships with Zoom and athenahealth. The company also describes a dedicated clinical operations team reviewing note quality, which is a process control rather than evidence of outcome.

Scale of use does not substitute for evidence of benefit, and this index applies that consistently. One figure is worth pursuing and is recorded here as unverified: third party material reports that roughly 97 percent of ambient notes are linked to supporting documentation from the patient transcript or the electronic health record. If that is vendor published and reproducible it is a groundedness measure rather than a satisfaction measure, and would be genuinely unusual in this category. It could not be confirmed from a Suki source. No dedicated research or outcomes page was located, so this assessment could change materially.

The privacy engineering here is specific in a way most of this category is not. The developer documentation states that neither the clinician nor the patient is identified in voice recordings, that no voice signatures are created or retained, that voice is used only for transcription and summarisation of notes, that recordings are de identified by chunking files, and that protected health information is removed from text transcripts to prevent individual identification.

Third party material additionally reports audio and transcripts deleted after 30 days by default with the generated note retained for the contract term. Held below a higher grade because the model training question is not resolved.

The developer documentation says voice is used only for transcription and summarisation, while the company's homepage says AI is only as good as the data it is trained on and that its electronic health record integrations and partnerships give it the ability to leverage an extensive dataset. Both can be true if the training corpus is de identified record derived text rather than voice, but nothing states that, and a buyer should not have to infer it.

Whether de identified transcripts or record derived data train or improve the models is the question to settle in writing. One structural point is worth noting without criticism: the developer documentation places responsibility for obtaining patient consent on the clinician, which is defensible for a platform sold through integrators but means consent for recording sits on the buyer's side.

The commitment is plain and unhedged, and published in the company's own developer documentation: Suki complies with HIPAA requirements and signs business associate agreements for patient data handling with its customers. The privacy policy goes further, stating that Suki acts as a business associate for healthcare, which is a direct self identification under the statute rather than the vaguer alignment language most vendors in this category use.

A business associate agreement is provided as part of standard enterprise contracting. One structural feature is unusual and worth a buyer's attention, because it shifts an obligation. The developer documentation states that before an integration sends personal data to the platform, the clinician is responsible for obtaining patient consent, including maintaining their own consent policies and securing all necessary authorisations.

That is a defensible allocation for a platform sold through integrators and resellers, but it means consent for ambient recording sits on the buyer's side of the line rather than being handled by the product. Held below a higher grade because the agreement itself is not public: no terms, tier, execution path or self serve route, so the instrument arrives with the enterprise contract after the sales process rather than being establishable beforehand.

A dedicated trust centre exists at trust.suki.ai, and the SOC 2 claim specifies the type, stated plainly on the company's own homepage as SOC 2 Type 2 certified and HIPAA compliant. That combination answers the question this index puts to every SOC 2 claim, and it places this record ahead of competitors whose security pages claim unnamed certifications or omit the type.

Encryption is described in transit and at rest with modern ciphers, alongside run time analysis to detect anomalies or suspicious software behaviour, which is a monitoring control rather than a static claim. Held below a higher grade on the scope of disclosure rather than the strength of the posture.

No ISO 27001, HITRUST, penetration testing statement or subprocessor list was located, and the trust centre's document inventory could not be enumerated in this review, so what sits behind it is unestablished. The strongest position in this category is set by vendors publishing a trust centre that names the scope of the audit alongside additional authorisations, penetration testing and a full subprocessor list.

No FDA clearance, none claimed and none required for ambient documentation. The company states plainly that outputs are decision support and that clinicians must review and validate notes before signing, so the human signature is the control point. But this is the furthest advanced scope expansion in the ambient scribe category, and on this record it is the core positioning rather than an adjacent module.

The company describes itself as a true AI assistant rather than a scribe, and states that it combines documentation, coding, clinical reasoning and question answering in a single solution. Published capabilities include ICD-10 and HCC coding support, chart question answering, voice commands for ordering and charting, and ambient order staging, launching first on athenahealth.

Three separate escalations sit inside that, and a buyer should evaluate them separately rather than treating the platform as one thing. Clinical reasoning and chart question answering are decision support functions rather than documentation, landing in the territory occupied by dedicated clinical reference products. HCC coding shapes risk adjusted reimbursement and carries documented audit exposure.

And ambient order staging places a proposed order into the chart, which is action rather than information. The regulatory rationale for the reasoning and order staging features specifically is worth requesting.

No AI governance artefact of any kind was located: no responsible AI statement, no bias or fairness position, no subgroup or demographic performance analysis, no error taxonomy, no model monitoring description and no published evaluation output.

The company's public material is substantial on security and compliance and silent on AI governance specifically, which is a distinguishable posture rather than a general absence of documentation, and it is the weakest position among its closest competitors: some publish AI training data and bias as a catalogued item a buyer may interrogate, and others at least assert bias mitigation as an objective.

One honest limitation is acknowledged and deserves credit, though it is an operational caveat rather than a fairness disclosure. The company states that transcription and predictive performance depend on audio quality and completeness, and that noisy environments or poor microphones reduce accuracy. That admission points directly at the unaddressed exposure.

Speech recognition accuracy varies measurably across accent, dialect, speech rate and vocal characteristics, and a vendor that already concedes audio conditions degrade its output is well positioned to publish performance by speaker group, care setting and language. None was found. The trust centre was identified but its contents could not be enumerated, so a governance document may sit behind it.

Suki publishes its terms of service, and they are unusually explicit on this axis in both directions. On the negative side the exclusion is total and it is written in capitals rather than buried. The vendor states it does not guarantee the accuracy or completeness of output generated by its artificial intelligence features, places responsibility for final review on the customer, and then excludes any liability whatsoever for the accuracy or completeness of processed data or for any decision made or action taken in reliance on it.

The scope of that exclusion is what makes it notable. Processed data is defined to include clinical notes, orders, diagnoses and transcripts, so a product that stages prescription orders and assists with diagnostic coding disclaims liability for the accuracy of exactly those outputs. It is the clearest published statement in this lane of where the risk is meant to sit, and a buyer should read it as the vendor intends it: the clinician who signs owns the content entirely.

On the positive side, and this is what lifts the grade above the floor, the terms carry an indemnity running from the vendor toward the customer, covering third party claims brought against the customer. Most peers in this category offer no vendor side indemnity at all, and several run indemnity only in the opposite direction.

The located text does not establish that it reaches clinical output rather than intellectual property claims, and the accuracy exclusion elsewhere in the same document suggests it does not, so the diligence question is simply what the indemnity actually covers. The patient position is unchanged: no route, no relationship, no standing.

Deep bidirectional integrations across five major systems including MEDITECH and Oracle Health, not only Epic, plus a partnership program extending to MEDENT, Azalea Health, and WellSky. The voice command layer requires write access to order and staging workflows, which is a materially deeper integration surface than note insertion. Suki Platform additionally lets other healthcare software vendors embed the capability in their own products.

Electronic health record embedding is this record's real strength, and the company stakes its positioning on it, claiming deep real time integrations with the four leading systems, Epic, Oracle Health, athenahealth and MEDITECH, and describing itself as the most embedded ambient AI solution on the market.

Specific integration routes are named rather than implied, including MEDITECH Expanse documentation APIs and athenahealth Ambient Notes at general availability, which lets a practice run Suki inside a standard athenahealth interface. A second and distinct deployment surface is easy to miss. The Suki developer platform and reseller programme let other companies embed the core technology inside their own products, so the software reaches clinicians under other vendors' brands.

That is a real interoperability position and also a governance question, since an end user may not know whose model is producing their note. A telehealth channel exists through a Zoom partnership, MedStar Health is named as a large multi site deployment, and continuous availability is marketed.

Held below a higher grade on two absences: no data residency statement of any kind, no region named and no residency commitment, which matters for a product processing recorded encounters; and no published implementation timeline, uptime commitment or support model.

No published pricing. Suki sells exclusively through enterprise channels, list prices are not posted, and there is no free tier and no free trial, so evaluation runs through a sales demo followed by an enterprise contract.

That last point is a real commercial fact rather than an absence, and it is worth stating because it differs from much of this category. Several ambient scribe competitors publish self serve monthly rates and offer free tiers precisely so a clinician can test accuracy on their own notes before committing. A buyer here cannot try the product or price it without entering a sales process.

One partial disclosure is credited: the product structure is public even though the prices are not, with two named tiers, so a buyer at least knows what is being priced.

Third party and competitor sources report figures in the region of 299 dollars per user per month for the documentation tier and 399 for the assistant tier, with enterprise contracts reported between 350 and 500 or more per user per month depending on integration depth, and setup fees reported between 500 and 2,000 dollars per practice. Those are recorded only so a later reader does not mistake them for published figures. This index does not credit competitor or aggregator estimates as vendor disclosure, and several of the sources reporting them sell directly competing products. No published return on investment figure, payback period or contract structure was retrieved.

Breadth comes from reach rather than from an enumerated specialty list, and it is genuine. Integration with Epic, Oracle Health, athenahealth and MEDITECH covers the electronic health record estate of most United States health systems, from large integrated networks down to ambulatory practices, and the MedStar Health deployment is described as spanning ambulatory specialties across more than 300 care locations.

A Zoom partnership extends coverage to telehealth encounters, a distinct setting many ambient products handle poorly, and the developer platform pushes the technology into third party applications serving settings the company does not sell into directly.

Functional breadth is also wider than a scribe's: the product combines documentation, coding, clinical reasoning and question answering in one assistant, with voice commands for ordering and charting, so it covers more of the clinical workflow than note generation alone. Held below a higher grade for three reasons, the first being the most important.

No performance or accuracy evidence is published for any setting, so breadth of availability is established while breadth of demonstrated usefulness is not. No enumerated specialty list or specialty tuned model claim was retrieved, unlike competitors marketing specialty specific tuning. And no multilingual capability was located, which several competitors publish.