Oracle Health Clinical AI Agent

The agent is generative throughout, drafting notes, orders and charge codes, transcribing dictation and summarizing charts, but it is a layer of the Oracle Health record rather than a product bought on its own. Organizations choose Oracle Health for the record system and its installed base, and the agent comes with it. Oracle now presents the agents as part of a voice first EHR with AI embedded in the workflow, which would change that picture if the AI becomes the reason the platform is chosen.

The reach is wide and the oversight is described in a phrase. Since February 2026 the agent drafts prescriptions, laboratory tests, imaging studies and follow up appointments from ambient listening, completing them from prior orders, the patient's history, the physician's favorites and organizational preferences, and since August 2026 it suggests professional fee charge codes. Oracle describes physicians reviewing and approving the orders and confirming the codes before submission, and clinicians reviewing, editing and signing all documentation.

That review step is the whole of the published oversight structure. There is no confidence signal on a drafted order or code, no order level accuracy rate, nothing on what the agent does when the conversation is ambiguous, and nothing on what happens after a drafted order turns out to be wrong. A wrong prescription is a clinical event, so the review gate carries the entire safety case.

Oracle describes the agents as working together as a system with semantic reasoning, sharing context in near real time on a unified architecture, and its product page says the outputs are interpretable and explainable so users can see the basis for a recommendation. No model card, named model, version or accuracy figure for notes, orders or codes is published.

The product page closes with a disclaimer that the AI features describe general product direction, may evolve and are subject to change, and may not be incorporated into any contract, so the published description of how the system works is not something a customer can hold Oracle to.

Nothing published identifies the parties behind this agent. No model or model provider, no hosting arrangement specific to the product and no subprocessor list is named. Oracle runs its own cloud and record platform, so the chain may be short, but that is inferred from corporate structure rather than disclosed.

External models are in use elsewhere on the platform. Oracle announced in September 2025 that conversational features in the Oracle Health Patient Portal are built on OpenAI frontier models, so a wholly first party pipeline for the ambient agent cannot be assumed. Leadership has also described software arriving in 2027 that would let customers use ambient listening to populate clinical trial data for pharmaceutical companies and researchers, which would move encounter derived content to parties outside the treating institution. Who would receive it, and under what terms, is not stated.

The published figure is time saved. Oracle reports that note generation has saved physicians more than 400,000 hours across US health organizations in nearly two years since launch, up from more than 200,000 hours in February 2026, and it publishes a case study on Beacon Health System's use of the agents and a quote from Billings Clinic.

These are customer reported aggregates without a clinician denominator, comparison group or method, and no peer reviewed study or accuracy evaluation of notes, orders or codes is published.

No product specific statement covers retention of encounter audio or transcripts, de identification, or whether customer content trains models. That gap carries more weight because a secondary use of encounter data has been announced: company leadership has described software arriving in 2027 that would let health systems take part in clinical trials using ambient listening to populate trial data, presented as a revenue source for community and rural hospitals.

What is not stated is the architecture behind it: whether the capture that feeds trial data is the same capture that produces the note, what a patient is told and who obtains consent, and whether a health system can adopt documentation without the research pathway. The output now includes orders and charge codes, which are derived clinical data with their own retention and reuse questions.

Nothing published describes safety engineering on the model: no guardrails, no hallucination controls and no process for handling a safety event.

No business associate agreement position scoped to this agent is published. Oracle supplies record systems to US health systems at national scale and those relationships run under contract, but nothing public lets a prospective customer see what the terms cover for the agent before contracting.

The scope question has grown with the product. The agent now drafts laboratory, imaging, prescription and follow up orders and suggests charge codes, so protected health information captured from a recorded conversation flows on toward pharmacies, laboratories, payers and billing through the same vendor's software. Nothing published says whether the agreement a health system holds covers order transmission, coding and billing derived from ambient capture, or which framework would govern encounter data used for the clinical trial pathway leadership has described.

No attestation scoped to this agent is published. Oracle runs a large cloud compliance program spanning certifications, government authorizations and regional frameworks, and that program describes how the cloud is operated rather than how this agent handles encounter audio, drafts orders or suggests codes.

The assurance is first party, since the compliance program belongs to the same company, which makes it easy to read as covering the agent. Nothing published says which report names the agent, whether ambient capture and the order and coding paths sit inside an assessed boundary, or what period a relevant report covers.

No clearance is claimed for the agent. The next generation ambulatory EHR was released pending health IT certification, which concerns the record system rather than any device review of the AI.

The agent's functions now run well past documentation: drafting prescriptions, laboratory tests, imaging and follow up orders since February 2026, suggesting professional fee charge codes since August 2026, and summarizing charts. Oracle describes each as drafted for the clinician to review and approve, and nothing published sets out why these functions sit outside device regulation.

One described feature puts commercial weight on a clinical recommendation. Company leadership has described the system proposing the right billing code, laboratory test and medication based on what the payer covers, and surfacing care gaps and next best actions from payer contracts. Coverage is a legitimate clinical consideration, but nothing published says what a recommendation is ranked on or whether the clinician sees the alternatives that were not surfaced.

Oracle's product page says the agents give interpretable, explainable outputs so users understand the basis for a recommendation. No fairness statement, subgroup analysis or accuracy by accent, dialect or speaker is published for what is now a voice first product.

The page's closing disclaimer, that AI assistance may evolve, is subject to change and may not be incorporated into any contract, allocates risk rather than describing how the system behaves.

Review before signing is the control on offer: physicians review and approve drafted orders and suggested charge codes before submission, and review, edit and sign all documentation.

Nothing stands behind the output, and the product page goes further than silence. Its disclaimer says the content outlines general product direction, is for information only, may not be incorporated into any contract, should not be relied on in purchasing decisions, and that the nature and extent of AI assistance may change. The vendor's own description of the agent is therefore not something a customer can hold it to. No accuracy figure, published limitation, guarantee, remediation commitment or indemnity covers notes, orders or codes, and a wrong prescription is a clinical event rather than a text correction.

The agent is native to Oracle Health and works inside the clinician's workflow there. It writes drafted notes and structured orders back into the record, places charge code suggestions within the orders workflow, takes dictation into any text field, and summarizes history, labs and medications from across the chart.

It is not available on any other record system, so it competes only inside the Oracle Health installed base. For an Oracle site the integration is as close as it gets; for anyone else it is not an option.

The agent is delivered as a cloud service within Oracle's own health stack, with mobile capture apps for both major phone platforms, and its released functions are available to US customers. No residency detail specific to the agent is published: nothing says which region hosts it for a given deployment, whether ambient audio is processed in that region or centrally, or whether the order and coding paths differ.

That matters because Oracle Health serves health systems with their own residency commitments and government adjacent customers, and Oracle runs regional and government cloud regions a customer could reasonably assume apply. The parent's regional footprint says nothing about this agent. Elsewhere on the platform, the patient portal's conversational features run on OpenAI models, so external model services are part of the stack.

The agent has no published price and no standalone list price. Oracle sells it within the Oracle Health record system relationship, so its cost sits inside an enterprise EHR contract and cannot be separated or compared with per clinician scribe pricing from outside. Oracle's product page adds that the pricing of the features it describes may change at Oracle's sole discretion.

Each released function is available in the United States only: note generation, order creation since February 2026, and coding, dictation and chart review since August 2026, with professional fee coding aimed at ambulatory workflows. Oracle's page describes clinical support across inpatient, outpatient and emergency settings and a nursing agent for capturing charting details by voice, while several administrative agents are marked as planned rather than shipped.

No specialty list and no language coverage is published, so fit for a given specialty mix cannot be checked from public material.