DeepCura

The agent network is the product and there is no non AI business underneath. The architecture decomposes the work into eight separately scoped agents rather than one model producing a note, a design choice with real consequences for how failures surface and where a buyer can intervene.

Wide autonomous reach with no published limits. The orders agent drafts laboratory tests, imaging, durable medical equipment and prescriptions; the billing agent generates CPT codes with linked diagnoses; the referrals agent drafts letters and handles prior authorization; the receptionist answers calls unattended; and the scribe writes back per field into structured EHR data rather than dropping text for a clinician to place.

Per field write back is the deepest form of automatic entry, and it removes the natural pause that transfer based products preserve. No agent has a published confidence threshold, accuracy rate, abstention behavior or review gate.

Architecture is described clearly, with each of the eight agents given a defined scope and per field write back specified rather than implied. What is entirely absent is measurement: no accuracy figure, no model card, no named models, no error rate and no evaluation methodology for any of the eight agents, including the ones generating prescriptions and billing codes.

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. The architecture is described in detail, with eight agents carrying defined scopes and per field write back, so the shape of the system is visible while nothing says what runs it or who else touches the content moving through it.

An eight agent design also implies more model calls per encounter rather than fewer, so whatever unnamed parties are involved are involved repeatedly, and the agents generating prescriptions and billing codes are handling the most consequential content in the product. End to end encryption is stated and a business associate agreement is offered to all covered entities, which bounds obligations without naming anybody.

A 4.6 out of 5 rating on a software review site is the only third party signal. No study, controlled evaluation, accuracy benchmark, clinical outcome data or named health system deployment. DeepCura itself publishes a large volume of comparative content about its own market, reviewing and ranking its own competitors, which makes it a self interested source; those pages do carry a dated disclaimer telling readers to verify with the vendor.

End to end encryption is stated and a business associate agreement is offered to all covered entities rather than only to enterprise accounts, which matters for a self serve product an individual clinician can adopt. No retention schedule, deidentification practice or statement on whether customer data trains models is published, and those are the disclosures that matter most for patient data.

States HIPAA compliance and provides a business associate agreement to all covered entities, available through a self serve path rather than requiring enterprise negotiation, which is a lower barrier for a small practice. The agreement terms are not published for inspection before signup.

One phrase here is doing work it has not earned. DeepCura describes its security practices as SOC 2 aligned.

Aligned is not certified, attested or audited. It describes an intention to resemble a control framework, and any organization can claim it without an auditor ever being involved. There is no report behind it, no scope, no period and no independent opinion, which are the four things a SOC 2 report actually consists of. A buyer reading the phrase quickly will take away that an examination happened. None did.

Against that, CASA certification is claimed and is a real external assessment. It examines application security for cloud applications and is independent. It is also narrower than SOC 2 in scope, covering the application rather than the organization's control environment, so it does not substitute for one.

The net position is one real but narrow assessment, presented in language that implies a broader one. What is missing is a full organizational examination, named with its type and period, of the kind that produces a report a buyer can request.

The phrase belongs to a family designed to be skimmed past, alongside SOC 2 ready, SOC 2 in progress, HIPAA compliant infrastructure and enterprise grade security. None of the five is an attestation, and each is compatible with no auditor having been engaged.

No clearance claimed and none required for the ambient note itself. What is shipping goes beyond the note, because these agents do not draft for review, they write.

The platform deploys eight agents around the scribe. Diagnoses are extracted, mapped to SNOMED and ICD-10 and written to the problem list. Allergies are captured with severity, coded and written to the allergy section. Medications are coded and written to the medication list. Orders covering labs, imaging, durable medical equipment, vaccines and prescriptions are placed through the record system's order interface. Billing codes are pushed to the charge slip. Referrals including prior authorization are automated, and follow up appointments are booked during the visit.

Writing structured clinical data into the chart is a different object from producing a narrative note a clinician signs. The signature is a clear control point for a note. It is far less clear what the control point is for a coded allergy already written to the allergy list, and nothing published says where human confirmation sits for each agent, or whether the review step governing the note governs the rest of them.

One consequence is the sharpest safety question here. The vendor states that coding allergens to standard vocabularies and validating them against the record system's own database ensures every allergy entry works with that system's built in clinical decision support and drug interaction checking. Machine generated allergy data therefore becomes the input to a downstream safety system the clinician treats as authoritative. An error in capture does not stay in the note. It propagates into interaction checking, and nothing in the workflow marks that entry as machine derived.

Markets appear to be United States only.

Two governance relevant designs, no governance disclosure. The billing agent produces CPT codes with linked diagnoses and the platform markets billing integrity, which sits at the milder end of AI coding claims. Patient intake is offered multilingually, with patients completing forms in their preferred language, yet no fairness statement, subgroup analysis or performance disclosure by language is published. Marketing a multilingual patient facing intake without any evidence of cross language performance is the specific gap.

There is no accuracy or error figure, no published limitations and no warranty, indemnity or remediation commitment, and the scope of that absence matters more than the architecture might suggest. The product is described as eight agents with defined scopes and per field write back specified rather than implied, which is useful structural detail. No measurement exists for any of the eight, including the agents that generate prescriptions and billing codes.

Those two outputs carry the highest consequences: a wrong medication or dose reaches a patient, and a wrong code shapes what is billed and creates federal false claims exposure for the practice submitting it. A vendor precise enough to specify which fields each agent writes back to could be precise about how often each one is right. One thing stands in its favor.

A business associate agreement is offered to all covered entities rather than reserved for enterprise accounts, which matters for a product an individual clinician can adopt, because it is the instrument that makes their use of it lawful and many self serve vendors gate it. It governs the handling of protected information rather than the correctness of output, but it is the right default.

Deeper than the price point suggests, with its limits stated openly. Bidirectional per field write back is supported for seven named systems, Epic, eClinicalWorks, athenahealth, OptiMantra, AdvancedMD, Veradigm and CharmHealth, and per field structured write back is the distinction that separates real integration from dropping a block of text into a note. The vendor also states plainly that providers on any other EHR can still use the product but must copy notes manually. Seven integrations is a narrow list, and no certification program membership or integration architecture is published.

The first question for any ambient scribe is answered by name: which company actually processes the encounter.

Encounter processing runs on a third party model service, and the vendor not only says so but lets the provider choose which one from inside the platform, naming the major commercial model families as the options. Speech recognition uses a named commercial medical speech engine. The platform is described as running on a major public cloud, with a completed Google CASA Tier 2 security assessment and business associate agreements available to covered entities.

Naming the model providers is real transparency, though other ambient scribes also name their model provider or build their own models.

Two weaknesses remain. No hosting region is named, no residency option is offered, and no subprocessor list carrying retention terms is published, so a buyer knows which companies process the encounter but not what those companies retain, for how long, or under what enterprise terms this vendor holds with them.

And the model choice, useful as it is, moves that question onto the customer without giving them the answer. Selecting a different model changes which company processes the consultation, and the retention posture that follows is not published per option. A practice choosing between model families from a dropdown is making a data processing decision with no data processing information attached to it, which is a heavier choice than the interface suggests.

The commercial disclosure here is unusually complete, and the detail is the evidence rather than the adjective.

The full per seat volume ladder is published rather than implied: $129 a month for seats one and two, 116 for seats three to five, 110 for six to ten and 101 for eleven and above, with each seat billed at its own tier so the blended average is explained rather than left to the buyer to work out. An annual alternative is published at $999 per seat. A free trial requires no credit card, cancellation is self service with no fees, and checkout is direct rather than routed through a sales conversation.

The part that stands out most is the savings comparison. When the vendor compares its bundle against buying the components separately, it names the comparator products and dates the estimate. That converts a savings claim from an assertion into something a buyer can reconstruct and disagree with, and it is the difference between a disclosed methodology and marketing arithmetic.

The published ladder, the annual rate, the trial terms and the dated comparator list are all checkable facts.

What a buyer still cannot see from the pricing page is what the agents beyond documentation cost to run, since order generation, billing and referral handling are bundled into the seat price rather than priced separately, and usage limits are not stated.

Ambulatory across a claimed 40 to 50 specialties, sold to physicians, therapists and practice managers, with multilingual patient intake extending coverage to non English speaking patients at the front of the encounter rather than only during it. The specialty count is asserted without enumeration, and no inpatient, emergency or acute capability is documented.