Oracle Health Clinical AI Agent

Graded on the established platform not algorithm precedent applied to ModMed and Elation Health in this same category. The agent is genuinely generative, but the moat is the Oracle Health electronic health record and its installed base rather than the model, and no organisation buys Oracle Health because of this agent. The grade describes the mechanism, not the quality of the tool.

Worth re checking on refresh: the next generation Oracle Health EHR is positioned as voice first with AI embedded directly in workflow rather than bolted alongside it, which would move this grade if the model becomes the reason the platform is chosen.

The sharpest oversight question in the ambient category, and the disclosure does not match the reach. Since February 2026 the agent drafts prescriptions, laboratory tests, imaging studies and follow up appointments from ambient listening, inferring dose, frequency and preferred pharmacy from patient history and clinician favourites.

Drafting a medication order from overheard conversation is a materially different risk surface from drafting a note, because an error is a clinical action rather than a text correction. Oracle describes output as ready to review, so a human remains in the loop, but no confidence threshold, no order level accuracy rate, no abstention behaviour and no description of what happens when the conversation is ambiguous have been published. Most autonomous, least disclosed, the same pattern this index graded C for Hello Patient.

No model card, no named models, no published limitations and no accuracy figures for either note or order generation. Capability is described in unfalsifiable terms such as advanced reasoning and semantic understanding. Oracle's own product page carries a forward looking disclaimer stating that the content outlines general product direction, is for information purposes only, may not be incorporated into any contract, and that the nature and extent of AI assistance may evolve and is subject to change. A buyer cannot determine from published material where this system degrades.

Nothing published identifies what is in the chain. No model or model family is named, no foundation model provider is disclosed, no hosting arrangement is described for this product specifically, and no sub processor list was located in two passes over the vendor's material.

The vendor operates its own cloud and its own record platform, so the chain is probably short in practice, but a chain that is inferred from corporate structure is not a disclosed one, and a buyer cannot confirm from anything public whether encounter content stays inside the group. One announced direction makes the question more pressing rather than less.

Leadership has publicly described software arriving in 2027 that would let record system customers use ambient listening to populate clinical trial data for pharmaceutical companies and researchers. Whatever the merits of that product, it describes encounter derived content moving to parties outside the treating institution, and this is the axis on which a buyer would expect to see who those parties are and under what terms. Nothing addresses it. Establish in writing which entities process encounter audio and derived text today, and which would under any research pathway.

The only published figure is customer reported time saved, more than 200,000 hours of documentation time for United States doctors as of February 2026. That is an adoption aggregate relayed by the vendor, not a study: no denominator of clinicians, no comparison group, no peer reviewed publication and no accuracy evaluation located. Graded C on the standing precedent in this index that scale of use does not substitute for evidence of benefit, the same call applied to Infervision, QuantHealth and Anterior.

No product specific statement on training data use, audio and transcript retention, or de identification was located. The earlier assessment identified this as the highest value axis to close and the second pass confirms it, because a secondary use of encounter derived data has now been announced publicly.

Company leadership has described software arriving in 2027 that will let record system customers participate in clinical trials using ambient listening tools to populate trial data, presented as enabling pharmaceutical companies and researchers to start trials quickly and as a valuable new revenue source for community and rural hospitals.

That is a clear statement of intent to derive value from encounter content beyond the care of the patient in front of the clinician, and it is the vendor's own. It may well be implemented with consent, de identification and institutional review, and none of that is in question here. What is unstated is the architecture: whether the ambient capture feeding trial data is the same capture that produces the note, what a patient is told, what consent is obtained and by whom, and whether a health system can adopt documentation without adopting the research pathway.

Those questions should be asked before the capability ships rather than after, because the answer determines what a patient consented to when they agreed to be recorded for their own visit.

The ordinary disclosures remain absent alongside it. No retention period for audio or transcripts, no de identification practice, and no statement on whether customer content trains models. The output now includes prescription orders, which are derived clinical data with their own retention and reuse questions.

Ask for the retention schedule, the training position in contract language, and the consent architecture for any research use.

No product specific business associate agreement posture for this agent was retrieved, and this index does not assert terms it has not seen.

Agreements plainly exist. This vendor is a record system supplier to United States health systems at national scale, and every one of those relationships is papered. What is absent is any published statement scoped to this agent, so a prospective buyer cannot establish before contracting what the terms cover.

The scope question has grown considerably since the earlier assessment, and that is the substance here. When the agent drafted notes, its footprint was encounter audio and a clinical note. It now drafts laboratory, imaging, prescription and follow up orders, and the surrounding platform handles prior authorisation requests and billing. So protected health information generated from a recorded conversation flows onward to payers, pharmacies and laboratories through the same vendor's automation.

That means the agreement a health system holds needs to describe more than documentation. Establish whether it covers order transmission, prior authorisation submission and billing derived from ambient capture, whether each is separately scoped, and where responsibility sits if an order generated from a misheard conversation reaches a pharmacy.

A second question follows from the announced research pathway. Where encounter derived data is proposed for clinical trial recruitment and population, the business associate framing may not be the right one at all, since research use typically runs through a different consent and oversight route. Establish which framework would govern it.

Ask for the agreement scoped to this agent, the list of downstream recipients, and the position on research use.

No product level attestation for this agent was located, as distinct from the parent's general cloud compliance programme.

That distinction is the whole of this record and it is the most tempting version of it in the index. This vendor operates one of the largest enterprise cloud compliance programmes in existence, spanning a long list of certifications, government authorisations and regional frameworks. A buyer looking for assurance will find an enormous amount of it, and almost none of it is about this agent.

The pattern is familiar from elsewhere in this lane, where vendors publish their infrastructure provider's certifications in place of their own. Here it is a first party version, which makes it harder to spot: the compliance programme genuinely belongs to the same company, so the inheritance feels legitimate. It still answers a different question. Platform level authorisation describes how the cloud is run. It does not describe how this specific agent handles encounter audio, what its own controls are, whether it was separately assessed, or whether it falls inside the scope of any particular report.

Scope is therefore the thing to ask about rather than existence. Attestations certainly exist. Establish which of them names this agent, whether ambient capture and the order generation path are inside the assessed boundary, and what period the relevant report covers.

The expanding function set raises the stakes. The agent now drafts prescriptions and other orders and touches prior authorisation and billing, so the surface any assessment would need to cover has grown considerably since the product was documentation alone.

No clearance claimed and none required for documentation. The distinction from the earlier assessment stands and remains important: this vendor released a next generation ambulatory record system pending final regulatory approvals, which refers to health information technology certification rather than any device review of the AI. Certification of the record system says nothing about the agent writing into it.

The scope has expanded substantially and the boundary questions now run well past documentation. Order creation shipped in early 2026, drafting laboratory tests, imaging and diagnostic studies, new and refill prescriptions, and follow up appointments from ambient listening. Prior authorisation requests, billing and clinical trial eligibility identification are described alongside it.

One feature deserves separate and careful attention, because it is the sharpest instance of commercial influence on a clinical recommendation anywhere in this index. Company leadership has described the system proposing the right billing code, the right lab and the right medication based on what is covered by the payer, and identifying care gaps and next best actions based on payer contracts. Medication selection weighted by coverage is a real and defensible clinical consideration, since an unaffordable prescription is not a good prescription. But a system that surfaces the covered option is optimising against a contract as well as against the patient, and where the same platform also handles the billing, a buyer should establish what the recommendation is ranked on and whether the clinician is shown the alternatives that were not surfaced.

A second tension is worth resolving directly. Product material describes orders as drafted; other statements describe agents sending a prescription order. Establish which it is, per order type.

A patient facing component also exists, giving patients plain language explanations of diagnoses and test results and answering questions about their records.

No fairness, subgroup or accent and dialect statement located, which is a notable absence for a voice first product. What Oracle does publish on the subject is a disclaimer rather than a disclosure: that AI assistance may evolve over time and is subject to change, and that product descriptions may not be incorporated into any contract. That allocates risk away from the vendor without telling a buyer anything about how the system behaves. Contrast Microsoft Dragon Copilot in this same category, which names accent and dialect underperformance explicitly.

This is the clearest published statement of non commitment located in the index so far, and it is worth reading carefully because it is affirmative rather than a mere absence. The product page carries a forward looking disclaimer stating that the content outlines general product direction, is for information purposes only, may not be incorporated into any contract, and that the nature and extent of artificial intelligence assistance may evolve and is subject to change.

A buyer should sit with what that means on this axis. The vendor has pre emptively removed its own description of the product from the set of things it can be held to, so a claim that the system failed to do what was represented has no representation to attach to. Where other vendors publish limitations that create something to argue with, this publishes a clause that forecloses the argument. Nothing else offsets it.

No accuracy or error figure is published for note or order generation, capability is described in unfalsifiable terms, no published limitations exist, and no performance guarantee, remediation commitment or indemnity was located. The output now includes prescription orders, which raises the stakes of a wrong one considerably above a documentation error. Set against that, the operating model does place a licensed clinician between the system and the patient, and that clinician signs. Ask for representations about the artificial intelligence functionality inside the contract itself, since the published material states plainly that it does not travel there.

Depth is high and breadth is nil, which is the whole shape of this product. The agent is native to Oracle Health, embedded in clinician workflow rather than launched alongside it, and now writes structured orders back into the record rather than pasting text. It is also unavailable to anyone not on Oracle Health, so it competes only inside that installed base. For an Oracle site this is the strongest possible integration; for anyone else it is not a candidate.

No residency detail specific to this agent was located. Delivery is a cloud service within the vendor's own health stack, and mobile capture applications are published for both major phone platforms.

One fact from the second pass is worth recording because it partially answers the question this axis asks first, though not for the component under assessment. The vendor has stated that the generative feature in its patient portal, which produces plain language explanations of diagnoses and results, is built on a named third party foundation model provider. So at least one clinical AI component in this platform runs on an external model service.

That does not establish what processes the ambient encounter, and it should not be read across. It does establish that external model services are in use somewhere in the stack, which makes the unanswered version of the question for the scribe worth pressing rather than assuming a wholly first party pipeline. Ask specifically which provider processes encounter audio and transcripts, what that provider retains, and under what enterprise terms.

Residency itself is unstated for this agent. That is more consequential than for a small vendor because this platform is deployed across health systems with their own residency commitments, government adjacent customers, and in some cases sovereignty requirements, and because the vendor operates regional and government cloud regions that a buyer may reasonably assume apply here.

Assume nothing from the parent's regional footprint. Establish which region hosts the agent for a given deployment, whether ambient audio is processed in that region or centrally, and whether the answer differs for the order generation path.

No published rate card and no standalone price. The agent is sold within the Oracle Health relationship, so its cost is embedded in an enterprise EHR contract and is not separable or comparable from outside. That is the structural reason bundled incumbent pricing cannot be benchmarked against the per provider per month figures the pure plays are quoted at.

Bounded and partly forward looking. Confirmed available to United States ambulatory providers, with order creation also United States only. Acute care functionality has been announced for 2026 rather than confirmed shipped in this pass. No specialty count or language coverage published, which is a gap relative to the pure play vendors in this category that enumerate both.