Avo

The foundation is curated clinical content, not a model.

Avo Pathway, described by the company as its foundational product, is a library of digitized medical society guidelines and calculators, and the stated business model invites content publishers and evidence based organizations to embed their intellectual property into the platform's copilots. A major clinical reference publisher is named as a knowledge partner doing exactly that. The company's own architecture diagram puts it plainly: record system data, knowledge bases, reasoning, and workflows, with the knowledge base as a distinct layer beside the model rather than inside it.

The asset is the dataset: strip out the generative layer and a valuable guideline library, a no code copilot builder and four deep record system integrations remain, and a health system would still pay for them. The platform sells itself on being more than a point solution, which is the same fact stated commercially.

The model does substantial work on top of it, and that work is not cosmetic. It reads the chart, produces differential diagnoses, suggests dosing, drafts notes from ambient capture across multiple participants, and pends orders. Those are model outputs with clinical consequence, not formatting.

So the content is the moat and the model is the leverage, and choosing this product means choosing both. That split says nothing about whether the product works.

The platform suggests far more than a note, and nothing published limits any of it.

Beyond note generation the platform drafts orders, identifies care gaps, produces differential diagnoses, generates treatment recommendations and suggests dosing. Dosing in particular is a recommendation with direct patient consequence, and the company publishes a figure claiming a 35 percent improvement in dosing quality compliance, which tells a buyer the function is used at scale without telling them how often it is right.

Two real controls exist. Orders are pended rather than placed and notes are drafted rather than filed, so a clinician signs before anything takes effect. And recommendations carry their source, so whoever reviews the suggestion can see which guideline produced it rather than being asked to trust it.

What is missing is everything that would tell a clinician when to look harder. No confidence threshold, no abstention behavior, no statement of what the system does when the patient data is incomplete or conflicting, no escalation path, and no acceptance or override rate. A sign off gate is only as good as the attention brought to it, and nothing in the product distinguishes the suggestions that deserve the most.

The company positions itself explicitly against alert driven decision support, arguing for standardizing care with love rather than alerts, and that is a considered stance on alert fatigue rather than a marketing line. But a philosophy is not an oversight mechanism, and a product that declines to interrupt has to be right more often rather than less.

No accuracy figure, model card, named model, hallucination rate or evaluation methodology is published.

What is described is the architecture rather than the artifact, and at that level the account is clear. The company sets out four layers, record system data, knowledge bases, reasoning and workflows, and the knowledge layer is documented in detail: named medical society guidelines, recovery milestones, consensus pathways, and a major clinical reference publisher embedding its content into the platform. The integration layer is named too, with the platform described as SMART-on-FHIR and a partner supplying a FHIR enabled layer for one record system. That settles what the system reads and what it reasons over.

What produces the reasoning is unidentified. No model, model family or foundation model provider, no version, no update or release cadence, and nothing describing how a customer learns the behavior has changed. For a platform deployed inside four record systems and maintained by the health system rather than the vendor, silent model change is a live operational question rather than a theoretical one.

The nearest thing to a performance statement in the vendor's material is a customer testimonial saying the user has not seen hallucination issues, which is impression rather than measurement.

For a product generating differential diagnoses and dosing suggestions, published error characterization matters more than it does for a note writer, and none exists.

The vendor draws its data line between patient and clinician, and the two halves go opposite ways.

On patient content the commitment is strong and architectural: clinical data collected to assist decision making, including data drawn through record system integrations, is processed only in real time and is not stored or saved after processing, and patient clinical data is stated not to be stored, saved or sold. That matters here because the integrations read deeply, across care summaries, radiology reports, laboratory results, vitals, medications, notes, problem list items, diagnoses and demographics. Not retaining the data at all answers more than an encryption claim does.

On clinician content the same policy goes the other way. Usage information is defined to include the clinician's name, national provider identifier, email, specialty, geography and institutional affiliations. The site participates in a healthcare communications network under which registration includes registration with that network, and professionally relevant contact information is disclosed to the network and its members by default. The policy then subordinates itself, resolving conflicts in favor of the network's policy. So identified clinician data flows to a third party network by default while patient data is not retained at all.

One supplier is named in the knowledge layer, a major clinical reference publisher embedding its content into the platform, and a partner is named supplying a FHIR integration layer for one record system. Those are real disclosures and they cover content and connectivity rather than inference.

On the model layer there is nothing: no model, model family or foundation model provider, and no subprocessor register covering the wider set.

Named customers and vendor reported percentages, with no method behind any of them.

The customers are real and specific: Children's of Minnesota, Driscoll Children's Hospital, PM Pediatric Care and Harbor Health, alongside an implementation of a RAND consensus pathway for vertebral fragility fractures and a partner deployment written up as a case study. Named institutional deployment at that level counts for something, but it is not evidence of benefit.

The figures are the company's own and none carries a method. The home page publishes a 35 percent decrease in note time per patient, a 94 percent reduction in implementation cycle, $11.6 million in annual cost savings across a multi national ambulatory organization, $7.5 million in increased annual reimbursement through documentation improvement, and a 35 percent improvement in dosing quality compliance for high risk medications. No denominator, no period, no baseline, no site named for most of them and no reference standard. The last figure is the interesting one, because dosing is the output with the most direct patient consequence, and a compliance improvement is not an accuracy measurement.

A marketplace listing attributes to the company a claim that randomized controlled trials showed roughly 50 percent clinician time savings and a 30 percent improvement in emergency department admission delays. Those words appear nowhere on the company's own site, and no citation, publication, author list or trial registration is published for them. As it stands, the trial claim is unsupported.

The patient data commitment is architectural rather than procedural. The privacy notice states that when patient related clinical data are collected to assist clinical decision making, including data collected via electronic health record integrations, that data is processed only in real time and is not stored or saved after processing, and that patient clinical data is not stored, saved or sold.

That matters here because the integrations read deeply, covering care summary data, radiology reports, lab results, vitals, medications, notes, problem list items, diagnoses and demographics. A no retention commitment is a different and stronger answer than an encryption claim. The weak side concerns the clinician rather than the patient.

The same policy defines usage information as including the clinician's name, national provider identifier, email, specialty, geography and institutional affiliations, and separately discloses that the site participates in a healthcare communications network, under which registration at the site includes registration with that network and the clinician's contact and professionally relevant information is disclosed to the network and its members by default.

The policy then subordinates itself, stating that where its own notice conflicts with the network's policy, the conflict resolves in favor of the network's policy. A vendor that will not retain patient data while routing identified clinician data to a marketing network by default has drawn its privacy line between the patient and the clinician.

Neither the full privacy notice nor the terms of use mentions HIPAA, a business associate agreement, a covered entity or a business associate anywhere. The notice is structured as a GDPR shaped instrument, with rights of access, rectification, erasure, portability, restriction and objection, complaint routes to national data protection authorities, and specific California and Nevada notices, but the United States health privacy framework is simply absent from it.

That is a real gap for this particular product rather than a formality, because the platform reads labs, vitals, medications, clinical notes, problem lists and diagnoses out of Epic, MEDITECH, athenahealth and Juno, and writes clinical notes and pended orders back. A vendor operating that deep inside the chart handles protected health information on its customers' behalf, the role a business associate agreement governs.

Absence from the public legal documents is not proof that no agreement exists, and an enterprise subscriber agreement very probably contains one, since health systems of the size named as customers would not sign otherwise. SOC 2 Type II is separately attested and linked from the site footer, which is the adjacent assurance. Before a sales conversation, nothing about the agreement or its execution path can be seen.

SOC 2 Type II is stated with the type explicitly specified, surfaced in the site footer on every page and linked to a dedicated announcement.

Specifying Type II matters: it is an audit of controls operating over a period rather than a point in time design review, so stating it says the controls have been examined across months of real operation. It sits in the persistent footer rather than buried on a compliance page.

What is absent is everything that would let the claim be checked. There is no trust center, so the attestation rests on the vendor's own announcement rather than a retrievable report. No audit period, assessor or scope statement appears. No ISO 27001, HITRUST, penetration testing statement, vulnerability disclosure policy or named subprocessor register is published.

The privacy notice describes safeguards only in general terms, an internal framework of policies and minimum standards with need to know access limits, without naming an encryption standard, a key management practice or a control framework. Attestation and description are different goods, and only the first is here.

Scope is open, because the platform spans standalone applications, embedded deployment inside four record systems and an order pending path into Epic. Whether the attestation covers the write path as well as the hosted service is not stated.

No FDA clearance, none claimed and none apparently sought. The clinical decision support exclusion position is argued harder in this vendor's own legal text than at any comparable product, and the terms of use do the work. The services are stated to offer health related information in an easily accessed format for educational purposes only, the information is stated as not intended to replace clinical decision making by a qualified professional, clinicians are stated to be ultimately responsible for decisions on diagnostics, symptoms and therapeutic treatments, the company states plainly that it does not give medical advice and does not provide medical or diagnostic services, and it states that its content is not intended for use by patients.

The substantive basis for the exclusion is present rather than merely asserted, since recommendations are anchored to named published guidelines including medical society pathways, recovery milestones and consensus pathways, so a clinician can go and read the source. The open question is that this product does not only inform, it acts. It pends orders into Epic and drafts clinical notes back into Epic, athenahealth and Juno.

An educational purposes disclaimer fits comfortably around a reference that answers a question and less comfortably around a system that stages an order in the chart, however firmly a human remains the one who signs it. The company publishes no regulatory rationale for the order pending function itself, as distinct from the platform in general.

The strength is a structural property rather than a disclosure document.

Recommendations are anchored to named published sources, including medical society guidelines, recovery milestones and consensus pathways, rather than to model inference or to an institution's own historical patterns. A major clinical reference publisher is named as a knowledge partner embedding its content into the platform, and its chief executive is quoted describing delivery of that content under the governance health systems expect. A clinician can trace a recommendation to its source and disagree with it on the evidence, which is a different and more durable property than a fairness statement.

The no code builder puts content governance with the organization rather than the vendor: the health system decides which guidelines are deployed and maintains them with minimal technical support.

Two things are absent and the second is specific to this product. No fairness statement, subgroup analysis or performance breakdown by any patient or clinician group is published. And no accent or dialect performance disclosure exists for the ambient capture, which the product page says generates notes in multiple languages with three or more participants. A system that transcribes an interpreted, multi party encounter is operating in exactly the conditions where recognition error falls unevenly, and the people most affected are the ones least able to check the note.

Industry funding also raises a governance question. The privacy notice states that life sciences companies fund and distribute clinical guidelines through the platform and present clinicians information about trials and therapies. Industry funded content inside a decision support surface is a governance matter, and nothing published describes how funded content is labeled to the clinician reading it or kept separate from society guidance.

Two real contestability mechanisms, and nothing standing behind the output.

The mechanisms are architectural rather than procedural. Recommendations are anchored to named published sources, including medical society guidelines, recovery milestones and consensus pathways, with a major clinical reference publisher named as a knowledge partner embedding its content into the platform. A clinician who doubts a recommendation can go and read the source it came from and disagree with it on the evidence, which is source traceability. Separately, orders are pended rather than placed and notes are drafted rather than filed, so a human signs before anything takes effect, which is review before commit.

What does not exist is any characterization of error. No accuracy figure, no hallucination rate, no published limitations, no evaluation methodology, and no warranty, indemnity or remediation commitment. The closest thing to a performance statement in the company's material is a customer testimonial reporting that the user has not seen hallucination issues, which is an impression from one person about the absence of a phenomenon they would only sometimes detect, and it is the weakest form of evidence.

The output is what makes that absence consequential. This product generates differential diagnoses and suggests dosing. A dosing suggestion is an input to a prescription rather than documentation of a decision already made, and an error in it can reach a patient. A differential shapes what is investigated and what is ruled out, so an omission is as harmful as a wrong inclusion and considerably harder to notice. For output of that kind, published error characterization is the basis on which a clinician decides how much scrutiny to apply.

The company publishes a 35 percent improvement in dosing quality compliance, which is an outcome claim with no method rather than an error rate, and it does not stand behind the output.

Named bidirectional integrations, verifiable listings and production evidence.

Named bidirectional integrations with major systems: Epic, MEDITECH Expanse, athenahealth and Juno Health, each documented separately with what it reads and what it writes. Epic reads care summary data including cross organization records, radiology reports, laboratory results, vitals, medications, notes, problem list items, diagnoses and demographics, and sends back clinical notes and pends orders directly into the chart. athenahealth reads a comparable set and receives clinical notes back. Juno drafts notes into the record using ambient transcription with formatting and templates. MEDITECH is read only today, reached from the external applications dock in the patient chart.

Verifiable in marketplace listings: the company links its Epic Showroom listing and its MEDITECH Alliance profile directly, and states availability on the athenahealth Marketplace. Those are third party listings that can be opened rather than claims to take on trust.

Evidence the connection runs in production: named customers on named systems, including a value based care organization whose deployment is written up as its own case study.

Two further things are published. The integration effort is given as a number, an Epic build stated at under twenty hours, which directly answers the question a chief information officer asks first. And the standards position is named rather than implied: the platform is described as SMART on FHIR, with a partnership providing a FHIR enabled integration layer for real time access to MEDITECH clinical data at the point of care.

Staging an order in the chart is a different order of integration from reading it.

Delivery spans standalone web and native mobile applications plus embedded deployment inside four named record systems, and the implementation shape is published rather than gated.

The Epic integration build is stated at under twenty hours, after which clinicians reach the product from an activity tab. Access routes are named per system: the external applications dock in MEDITECH Expanse, the left hand sidebar within a patient chart in athenahealth, with activation handled jointly by the vendor and an athenahealth representative. The company describes the install as one time, with no technical involvement needed for later maintenance and updates, which is a meaningful operational claim for a small hospital and one that also means the customer does not control when behavior changes.

A data residency statement exists: the privacy notice states that the services and their servers are located in the United States, and warns non United States users that their information may be transferred to and processed in a jurisdiction whose privacy laws may be less protective than their own.

Named deployments span a wide size range, from children's hospitals and large academic systems to smaller value based care organizations, with dedicated propositions for rural and small hospitals, large and mid size hospitals and ambulatory care.

Two points remain open. The residency statement is a country and nothing finer, with no region, tenant isolation model, failover location or subprocessor register, and it sits in tension with the company's self description in the same document as a global organization. And no customer count, uptime commitment or support model is published.

No pricing page, no published enterprise rate and no rate card.

What the vendor does publish is real but peripheral and buried. The terms of use disclose actual consumer subscription prices for two specialty applications, at $19.99 a year for one after a thirty day free trial and $9.99 a year for the other after a ninety day trial, with auto renewal terms and cancellation windows stated. Those are published prices, but they price two narrow applications rather than the platform a health system actually buys, and they sit inside legal text rather than on a pricing page.

The enterprise tier, which is the whole commercial relationship for the named hospital customers, is gated behind a demo request. A comparative pricing claim without a number is made on the products page, describing competitive pricing against other ambient solutions, which asserts a position on cost while disclosing nothing that would test it.

What is discoverable is the shape rather than the figure: the platform is sold to provider organizations with separate propositions by size, the implementation effort is published as a number, and the product is licensed rather than transacted per encounter. That is enough to establish how the product is sold without establishing what it costs.

One structural fact is a revenue fact. The privacy notice states that commercial clients including life sciences companies use the vendor's marketing solutions to fund and distribute clinical guidelines and to present clinicians information about clinical trials and innovative therapies. Part of the content a clinician sees at the point of care is industry funded, and that is disclosed in a privacy policy rather than in a published advertising or sponsorship policy. For a product whose value rests on the trustworthiness of its guideline content, where the funding sits is itself a commercial term.

Coverage is named unusually well and only part of it has anything behind it.

What is named is broad and specific. Ambient capture handles three or more participants, which is the encounter shape most scribes quietly fail at and the norm in pediatrics and in any consultation using an interpreter. Clinician coverage extends beyond physicians to advanced practice providers, medical assistants, dieticians and behavioral health providers. Workflow coverage spans admission, rounding, discharge, charting and care planning rather than the outpatient visit alone, and the company publishes separate propositions for rural and small hospitals, large and mid size hospitals, and ambulatory care. Many note templates are offered.

Part of that is supported. The multi participant and pediatric claims sit against named children's hospital customers, and the value based care claim against a named partner deployment written up as a case study.

The language claim is not. The current product page says notes can be generated in multiple languages, with no number, no list of which languages, and nothing evidencing performance in any of them. Earlier material put the figure at fifty or more; that number no longer appears on the company's site. For a product whose stated strength is the interpreted, multi party encounter, the language list and its evidence are the substance of the claim rather than a detail, and an unquantified multiple languages is a weaker statement than fifty or more.

Surfaces tested for the language claim: home page, products page, EHR integrations page and the resources library.