Speke

The corporate family's core business is human scribing, and the enterprise tier adds trained specialists. That workforce belongs to Speke AI Pro and to separate services the group sells, and it sits outside the self serve product.

Speke AI Express is sold on its own at a published monthly rate, and on that tier the model is the product: ambient capture of the encounter, a generated note, personalization to the provider and write back into the record system, with the clinician reviewing and approving. Remove the model and nothing remains to sell. The vendor's own FAQ describes a mix of proprietary and publicly available models that it tunes further in house, with specialty specific outputs.

The human layer is what a buyer prices at the enterprise tier, where the choice is how much work a person does after the model has done its part.

Speke AI Pro adds quality assurance and editing by trained clinical support specialists before a note reaches the clinician. That review is not part of Speke AI Express, the self serve tier at $150 a month.

The vendor's own material states this both ways. The homepage describes a professional reviewing the note for accuracy before submission as part of how the product works, while the feature lists and the Pro page place that review in the enterprise tier. The feature lists and the Pro page are the more specific of the two.

What Express provides is a clearly stated clinician oversight model. Its own page says providers review, edit and refine the generated note and approve it before it becomes part of the patient's record, including where the note is written back into the record system. One detail goes further: where the system cannot produce a note, it says so in the space where the draft would appear, rather than returning a thin note without warning.

What is missing is any threshold at which the system stops, flags low confidence or routes a case for attention, and anything on what happens after a note is wrong.

No accuracy figure, no model card, no named models and no evaluation methodology. The technical claims offered are credentials rather than disclosures, notably that product and machine learning staff were hired from Amazon and Microsoft, which says nothing about how the system performs. Language coverage is also stated inconsistently across the vendor's own properties, appearing as 25 or more languages in one place and 30 or more in another.

No party in the chain is named, but the architecture is described.

The automated path is stated at the level of its model layer. The vendor advertises the latest generative AI models in both published tiers, and its own FAQ describes a mix of proprietary and publicly available models that it tunes further in house. That places the product in a known class of system even though no model, model family or provider is identified. Alongside it a retention position is published, and it is concrete: once a visit is marked complete, it is deleted from the vendor's system after 72 hours.

On the enterprise tier the chain also has a human half. On Speke AI Pro a workforce of trained clinical support specialists reads recorded clinical conversation and performs quality assurance and editing on draft documentation. Speke AI Express, the self serve tier, carries no such review. A Pro buyer therefore knows a category of party handles the content, without knowing who they are.

What is absent is every identity. No foundation model provider, no hosting arrangement and no subprocessor list. The workforce is not enumerated in any respect: where it sits, whether review is onshore, how specialists are vetted, and whether they are employees or contracted are all unstated. The corporate family operates a separate remote scribe service, which makes the location question concrete rather than hypothetical.

Because the 72 hour deletion commitment is specific, its boundaries matter. Its text does not say whether it covers the audio, the transcript and the draft, whether it applies to copies held by a specialist during their work, or whether it survives in any backup.

Independent third party validation exists, with limits that matter. KLAS published an Emerging Technology Spotlight on Speke, which validates the experiences of organizations actually using it rather than repeating vendor assertions. An Emerging Technology Spotlight is an early stage validation based on a small interview base and is not equivalent to a Best in KLAS ranking, so it evidences real buyer experience rather than comparative performance. Vendor reported figures alongside it, up to 50 percent less documentation time and an average of 2.5 hours saved daily, carry no denominator or design. There is no peer reviewed publication.

The review workforce is the distinctive privacy question here, and it applies at one tier, not both. On Speke AI Pro, trained clinical support specialists perform quality assurance and editing, so recorded clinical conversation and draft documentation are read by people as well as processed by a model, which is a larger and more variable exposure surface. Speke AI Express, the self serve tier, carries no such review. The homepage describes review as part of the general workflow, so the published pages differ on which tier includes it.

One retention position is published and it is specific: a visit marked complete is deleted from the vendor's system after 72 hours. Nothing is published on whether customer data is used to train or tune models, which matters because the vendor says it tunes models further in house, and nothing on deidentification, guardrails or how an inaccurate or unsafe output is handled. Only the retention question is answered.

Compliance is claimed and the agreement is not. The product homepage states that Speke AI is HIPAA compliant, and repeats it inside both published pricing tiers alongside the SOC 2 Type 2 claim. No statement of business associate status, no agreement, template or scope statement is published, and the footer's privacy policy is the only legal document a buyer can open before contact.

The question turns on who touches the data, and the tiers differ on exactly that. The self serve tier is software led. At the enterprise tier a remote specialist employed by the vendor performs quality assurance and editing, produces documentation and routes it into the customer's record system, which means vendor personnel hold working access to the record itself rather than only to an audio file. An agreement scoped to the software does not describe that arrangement, and an agreement written for one tier does not describe another.

Second, and unusually, this product is sold to veterinary practices as well as to human healthcare providers, with veterinary listed among the specialties on the product's own homepage. Veterinary records fall outside the health privacy rule entirely. So a single vendor compliance posture is being applied across customers where the rule governs completely and customers where it does not apply at all. Nothing published shows whether the terms offered to healthcare customers are healthcare terms or a general form written to cover both.

One published fact belongs in the same conversation. The vendor states that a visit marked complete is deleted from its system after 72 hours. That is a concrete retention position, stated on a marketing page rather than in an agreement, and what it covers is not spelled out.

Scale raises the stakes rather than lowering them. This is an established organization operating across a large client base and many care sites with a workforce built over two decades, so the questions concern a mature operation rather than a startup's gaps: who the workforce are, whether they are employees or subcontracted, where they work, and what their record access reaches beyond the encounter they are documenting.

The product homepage states that Speke AI is HIPAA compliant and SOC 2 Type 2 certified, and repeats the certification inside both published pricing tiers.

A trust portal also exists. The footer links a trust page, which resolves to a hosted trust center for the wider HealthChannels group, and the portal may hold the report.

Two questions remain. No audit period, auditor or scope statement accompanies the certification, and a report has both a scope and a clock. And the portal covers the corporate family rather than this product, and nothing says whether the report's scope reaches the Speke platform specifically or only the group.

Scope matters more here than for a pure software vendor, because of what this organization's model places inside the boundary. At the higher tier the vendor's own trained personnel review encounters, author notes and route them into the customer's record system. That is a standing internal access surface operated across a large client base. The open questions sit there: how personnel are vetted, how access is provisioned and revoked, what a reviewer can reach beyond the encounter they are documenting, whether their activity is logged in a form the customer can audit, and what happens when someone leaves.

Recognition in a healthcare research firm's technology spotlight appears in the vendor's material, with customers giving high performance grades. That is a customer satisfaction assessment and a real market signal; it is not a security examination.

No clearance claimed and none required for ambient documentation.

Two things stand out.

The vendor states that the platform processes audio in more than twenty five languages and generates a note in English, and then says explicitly that it cannot act as an interpreter and that a native speaker of the recorded language will need to corroborate the note's accuracy. That is a vendor publishing the limitation of its own capability alongside the capability, and stating what mitigation is required. The point bears directly on what a clinician is signing: where an encounter took place in a language the reviewing clinician does not speak, the signature attests to a translation the signatory cannot check, and this vendor says so in its own material.

The second is the service structure. At the enterprise tier a remote specialist employed by the vendor performs quality assurance and editing on the documentation and routes it into the customer's record system. The clinician then reviews and approves. That is a different object from a clinician editing a machine draft. The documentation is worked by a third party from audio, and the clinician's signature attests in part to another person's account of an encounter they themselves conducted.

This applies only at the enterprise tier. Quality assurance and editing by trained specialists is listed as an enterprise tier feature and not in the self serve tier, while the homepage describes review as part of the general workflow. Which object is being signed therefore depends on the contract.

The product is also marketed to veterinary practices alongside human healthcare. That does not change the device analysis, since neither is regulated as a device here, but it does mean one product spans two entirely separate regulatory worlds.

Contains a precise language scope disclosure, standing alone rather than inside a formal governance document. Speke states that it processes audio in 25 or more languages and generates the note in English, then adds explicitly that it cannot act as an interpreter and that a native speaker of the recorded language will need to corroborate note accuracy. That does two things: it names the limitation, and it assigns the verification duty to a specific person rather than leaving it unowned. No formal transparency document covering accent and dialect performance, unsupported use cases or subgroup considerations is published, and no fairness or subgroup analysis.

Specialist quality assurance before the clinician sees the note belongs to Speke AI Pro only. The homepage describes that review as part of the general workflow; the feature lists and the Pro page confine it to the enterprise tier.

On Speke AI Express, the self serve tier, the correction mechanism is the clinician's own review. Providers review, edit and approve the note before it enters the record, and the system states when it could not generate a note rather than producing one silently. Review before commit is a real mechanism for challenging an output.

Nothing stands behind the output. No accuracy figure with a method or denominator, no error rate, no warranty, indemnity or remediation commitment, and no route for a patient affected by an inaccurate note. The performance claims offered are credentials rather than measurements, notably that the engineers previously worked at large technology companies. Language coverage is stated inconsistently on the homepage, as more than 25 languages in one section and more than 30 in both pricing tiers.

Bidirectional record system integration was announced in September 2024 with the claim that integration is available for all industry leading and midmarket systems, positioning the product as record system agnostic.

The breadth is partly named and partly asserted, and the difference matters. Five systems are identified on the product's own pages, with logos: Epic, athenahealth, ModMed, Veradigm and eClinicalWorks. That is a real answer for a practice running one of them. Beyond those five the claim of universal coverage is not enumerated, and no certification program membership or partner directory listing is published to support it.

Bidirectional write back is the meaningful part of the claim. Naming a system establishes that a connection exists; it does not establish what that connection does, whether the note lands as structured content or as text, or whether the same mechanism is available at every plan. Integration depth also tends to vary by customer instance rather than by vendor, particularly on the larger platforms.

Named integrations with the major platforms are established, while the wider agnostic claim rests on assertion.

Residency here is a two part question covering both storage and workforce access, and more is known about one half than the other.

On storage, no hosting region, residency option or subprocessor list is published, and nothing establishes whether a third party model service processes the audio or what it retains.

On workforce, more is now known. The service runs in tiers and the tiers differ in who handles the encounter rather than only in turnaround. At the entry tier the software produces the note. At the higher tiers a remote scribe reviews the visit data, performs the documentation and routes it into the customer's record system, with bidirectional record integration available at those tiers specifically. So the answer to where the data goes depends on which tier is purchased, and a buyer selecting on price or turnaround is also selecting an access model without necessarily realizing it.

Where that workforce sits is not published, and for a service of this kind it is a residency fact rather than an operational detail. A note reviewed and authored by a person in another jurisdiction has crossed a border whatever the servers do.

Language handling adds a further question. The platform processes audio in more than twenty five languages and the vendor states that a native speaker of the recorded language must corroborate accuracy, which implies a review capability spanning those languages. Who performs that corroboration, and where, is not published.

The vendor publishes a rate and currently sells two tiers.

Speke AI Express, the single user tier, is priced from $150 a month, billed monthly, with a direct subscribe route that needs no sales conversation. Speke AI Pro is quoted. A buyer can therefore size the entry cost before contact; the enterprise tier carries no figure.

The tiers are separated by how much human work is funded rather than by seats or volume. Pro adds quality assurance and editing by trained clinical support specialists. A third tier, Plus, appears in a 2024 announcement and is absent from the current site.

Coverage is wide and unusually well spread across care settings rather than concentrated in outpatient, which is the more useful property for a multi site buyer. Roughly 80 specialties spanning ambulatory, acute and urgent care, with emergency department depth inherited from the parent group's core scribe business, plus veterinary medicine, which the product's own homepage lists among its specialties.

Audio processing across 25 or more languages, with the honest caveat that output is English and that a native speaker must corroborate accuracy. The published figure is inconsistent, appearing as 25 or more in one section and 30 or more in the pricing tiers on the same page. Site specific templates, macros and dot phrases carry over per provider, which is the practical form specialty coverage takes for a clinician switching from a human scribe.