Amazon Web Services has published a reference architecture for an agentic telehealth visit that begins before the clinician joins and continues through documentation and coding. Patient-engagement agents can verify identity, manage appointments and collect the patient’s stated reason for the visit. Point-of-care agents can assemble a source-linked record summary, draft a note from the conversation and suggest billing codes.[1]
AWS brings verification, summaries, documentation and coding into a shared workflow using FHIR-compatible records. Parts of the published example require additional implementation: pre-visit intake and structured write-back use custom agents, verification and scheduling need EHR integration, and medical coding requires account enablement. Those dependencies matter when a health system estimates the work needed to deploy the complete visit.[1]
The six stages in the published telehealth workflow
| Measure | Value | Context |
|---|---|---|
| Identity and scheduling | Live verification and confirmed booking | Requires integration with the health system’s record and schedule.Source 1 |
| Pre-visit intake | Records the patient’s stated reason | The agent is not meant to diagnose or provide clinical advice.Source 1 |
| Patient insights | Source-linked summary of conditions, medication, labs and care gaps | The clinician reviews each statement against the underlying FHIR resource.Source 1 |
| Ambient documentation | Transcript and draft structured note | The physician reviews and customises the note.Source 1 |
| Structured write-back | Staged FHIR resources | Approval occurs before data is written to the clinical store.Source 1 |
| Medical coding | Suggested ICD-10 and CPT codes with confidence and supporting text | A coder reviews the suggestions.Source 1 |
The boundary between administration and medicine matters
Confirming a date of birth, finding an open appointment and writing a booking are administrative actions with clear system-of-record checks. Summarizing a longitudinal record is more complex but can remain evidentiary when every statement links to a source. Assessing symptoms, choosing treatment or deciding whether a diagnosis is correct crosses into clinical judgement. The AWS example explicitly keeps the intake agent from making those decisions.
That separation should be visible in the interface. A patient must not mistake a pre-visit collection script for medical advice. A clinician should be able to distinguish reported patient language from model-generated interpretation. A coder should see which phrase supports a suggested code. The more fluent the agent becomes, the more important it is to preserve these labels.
Where an agent can act—and where professional authority remains
| Step | Agent contribution | Required authority |
|---|---|---|
| Appointment | Verify information and write a confirmed booking | Health-system rules and patient consent.Source 1 |
| Record summary | Retrieve and organize source-linked facts | Clinician checks relevance and accuracy.Source 1 |
| Visit note | Draft structured documentation from the encounter | Clinician edits and signs the record.Source 1 |
| Diagnosis and treatment | May surface existing data or possible care gaps | Licensed clinician makes the medical decision.Source 1 |
| Billing code | Suggest a code with evidence and confidence | Coder or authorized reviewer approves submission.Source 1 |
Automation can move work rather than remove it
A draft note saves time only if correcting it takes less effort than writing the note directly. A record summary helps only if the clinician can quickly verify it and important omissions are rare. Coding suggestions can increase revenue or reduce denials, but aggressive false positives can create compliance risk. The complete measurement should include generation time, review time, corrections, rejected suggestions and downstream errors.
This is especially important when systems share one data foundation. A mistaken field can propagate from an intake summary into a note and then into a code. Reuse removes duplicate integration, but it can also make one early error more influential. Source links, versioning and explicit approval should travel with the data, not disappear when it moves between agents.
A single audit trail is a genuine systems advantage
AWS argues that one platform can give verification, summaries, documentation and coding a common domain model, access control and audit trail. If implemented well, that makes it easier to reconstruct who or what created each entry and which reviewer approved it. The value is not simply fewer vendors. It is a more coherent chain of evidence from patient statement to final record.[1]
Agentic healthcare will be judged by the reliability of the complete encounter. The strongest system will make administrative work faster while keeping clinical authority unambiguous, evidence close to every claim and corrections easy to propagate. An agent can coordinate the visit; responsibility for the patient cannot be delegated to a fluent interface.
Sources & further reading
- Agentic AI on Connect Health: An end-to-end telehealth visit Amazon Web Services · 6 Oct 2026 · Accessed 8 Oct 2026
- AWS HealthLake Amazon Web Services · 8 Oct 2026 · Accessed 8 Oct 2026
- Amazon Connect Health Amazon Web Services · 8 Oct 2026 · Accessed 8 Oct 2026