How Amazon Connect Health brings agentic AI to the point of care

Amazon Connect Health brings agentic AI to the point of care, helping clinicians reclaim the nearly two hours they spend on administrative work for every hour of direct patient care. Agentic AI is artificial intelligence that autonomously performs multi-step tasks on behalf of users. On March 5, 2026, AWS announced Amazon Connect Health, a purpose-built agentic AI service that handles clinical documentation, patient insights, and medical coding directly within existing electronic health record (EHR) workflows. Amazon Connect Health delivers two sets of capabilities: patient engagement (patient verification and appointment management, pre-integrated with Amazon Connect) and point of care (patient insights, ambient documentation, and medical coding, available via a unified SDK for health technology developers). This post focuses on the point-of-care capabilities and shows how EHR companies, healthcare independent software vendors (ISVs), and tech-enabled provider organizations can use Amazon Connect Health to embed these capabilities into their existing workflows using a single, unified Software Development Kit (SDK). We walk through the clinical workflow end to end, explore the architecture, and share a fully working sample application on GitHub.

Documentation, coding, and pre-visit preparation consume time that could be spent with patients, and the burden is a leading driver of clinician burnout. Healthcare organizations have explored a range of AI tools to address this, but most require standalone applications, separate workflows, or significant integration effort. Rather than asking clinicians to adopt another tool, Amazon Connect Health meets them where they already work: inside the EHR.

The integration challenge in healthcare AI

Healthcare builders face a familiar challenge. Clinicians need AI that reduces work, not AI that creates a new tab to check, a new login to manage, or a new interface to learn. At the same time, engineering teams need solutions that are modular, standards-based, and straightforward to integrate into complex EHR architectures.

Amazon Connect Health addresses both sides of this equation. It offers three point-of-care capabilities (patient insights, ambient documentation, and medical coding) through a single SDK with a consistent interface. Each capability is fully managed and independently adoptable. You can start with pre-visit patient insights this quarter, add ambient documentation next quarter, and layer in standalone medical coding after that, all through the same client, same authentication model, and same domain structure. No new integration projects are required.

A unified domain model

At the center of Amazon Connect Health is the concept of a domain, a logical container that groups all resources for an organization or environment. Every API call references a domain identifier, which means all capabilities share the same organizational boundary, the same access controls, and the same audit trail.

This design is intentional. In healthcare, traceability matters. When a pre-visit summary references a lab result, and that same lab result appears in the ambient documentation transcript, and the resulting medical code links back to the clinical note, all of that needs to be auditable under a single organizational context. The domain model makes this possible without requiring builders to stitch together separate systems.

The SDK is available for multiple languages including Java, Python (Boto), JavaScript, .NET, Go, C++, Ruby, Rust, and Swift, with the same operations accessible through the AWS CLI.

Before the visit: patient insights from longitudinal records

The first touchpoint in a clinical workflow is preparation. Before a patient walks in, the care team needs context. What happened at the last visit? Are there outstanding lab results? Which chronic conditions need attention? Are there risk adjustment opportunities?

Amazon Connect Health’s patient insights capability automates this preparation. It reads structured and unstructured Fast Healthcare Interoperability Resources (FHIR) data from AWS HealthLake, including conditions, medications, encounters, observations, diagnostic reports, and clinical notes. It then produces a visit-specific summary that includes patient background, health events since the last visit, and Hierarchical Condition Category (HCC) recapture opportunities for risk adjustment.

Amazon Connect Health writes the output as a structured summary to Amazon Simple Storage Service (Amazon S3), organized into clinical narrative sections. Each statement in the narrative includes evidence references, which are pointers back to the specific FHIR resources in HealthLake that support it. This means a clinician reading “Glucose 120 mg/dL” in the summary can trace that value directly to the source Observation resource, giving full transparency into how the AI arrived at its conclusions.

For EHR builders, the integration pattern is straightforward: submit a job with the patient’s FHIR identifier and a pointer to the HealthLake datastore, poll for completion, and read the structured output from Amazon S3. You can then render the summary in whatever format fits your existing pre-visit workflow, whether that is a sidebar panel, a preparation screen, or a printable brief.

What this looks like in practice

Consider a primary care clinic preparing for a morning of follow-up visits. As the schedule loads, the EHR triggers patient insights jobs for each upcoming patient. By the time the clinician opens a patient’s chart, the pre-visit summary is already populated:

• A narrative overview of the patient’s history, woven from conditions, medications, and recent encounters
• A timeline of health events since the last visit, including lab results, specialist referrals, and medication changes, each linked to its source FHIR resource
• Visit priorities highlighting conditions that need attention, including HCC recapture opportunities that might otherwise be missed

The clinician walks into the exam room already informed, without having spent 10 minutes clicking through chart tabs.

During the visit: ambient documentation that listens and structures

Once the visit begins, Amazon Connect Health’s ambient documentation capability takes over. It captures patient-clinician conversation in real time through a bidirectional audio stream and produces structured clinical documentation automatically.

The technical flow works as follows: the EHR application captures microphone audio in standard Pulse-Code Modulation (PCM) format (16kHz, 16-bit, mono) and streams it to the service through a persistent session. As the conversation progresses, transcript events stream back in real time, giving the clinician (or the EHR) visibility into what the service captures. When the session ends, Amazon Connect Health generates and writes four structured outputs to Amazon S3:

• Clinical documentation: Subjective, Objective, Assessment, Plan (SOAP) notes formatted according to a configurable note template
• Medical codes: International Classification of Diseases, 10th Revision (ICD-10) diagnosis codes and Current Procedural Terminology (CPT) procedure codes with confidence scores, generated directly from the conversation content
• After-visit summary: A patient-friendly summary of the visit, written in accessible language suitable for patient portals or printed handouts
• Transcript: The complete conversation transcript with speaker attribution

The medical codes generated from the streaming session typically appear approximately 30 seconds after the clinical document. This reflects the additional processing required to analyze the full conversation context and map it to standardized code sets.

What this looks like in practice

A clinician starts a follow-up visit with a patient who has hypertension and recurring headaches. The EHR’s consultation mode activates ambient documentation with no separate app, no recording device, just a microphone icon in the existing interface.

The clinician and patient have a natural conversation. The clinician examines the patient, discusses medication options, and recommends a follow-up in two weeks. When the visit concludes, the EHR transitions to a review screen where the SOAP notes are already populated:

• Subjective: Chief complaints of headache and elevated blood pressure, as described by the patient
• Objective: Blood pressure reading of 140/90, physical exam findings
• Assessment: Essential hypertension (I10), tension headache (R51.9)
• Plan: Start lisinopril 10mg daily, follow up in two weeks

The medical codes appear alongside the notes: ICD-10 codes I10 and R51.9, CPT code 99213, each with a confidence score. Clinician reviews, makes any edits, and approves. Total documentation time: under two minutes, compared to the typical 10–15 minutes of manual charting.

After the visit: medical coding from final documentation

Medical coding is available both as an automatic output of the ambient documentation pipeline and as a standalone capability. The standalone path is particularly valuable for two scenarios: when clinical notes are authored manually (not through ambient documentation), and when a clinician edits AI-generated notes and needs codes regenerated to reflect the final documentation.

The standalone medical coding API accepts clinical text, whether the full encounter note, a SOAP summary, or any clinical narrative, and returns ICD-10 and CPT codes with confidence scores. It does not require access to the patient’s longitudinal record or any external data source. It analyzes the text you provide and identifies the appropriate codes.

This design supports a clean edit-and-regenerate workflow. A clinician reviews the AI-generated SOAP notes, adds a detail about a secondary condition, removes a finding they disagree with, and saves. The EHR sends the updated text to the medical coding API, and the codes refresh to match the clinician’s final documentation. The codes always reflect what the clinician approved, not what the AI initially suggested.

Note: Medical coding is currently in gated preview. Contact your AWS account team to request access. Patient insights and ambient documentation work independently without it.

What this looks like in practice

After reviewing the ambient-generated SOAP notes, a clinician adds a note about the patient’s reported dizziness and removes a mention of a condition they don’t believe is clinically relevant. The EHR detects the edit and regenerates codes from the updated text. The new code set includes an additional ICD-10 code for dizziness (R42) and drops the removed condition, ensuring the billing codes accurately reflect the clinician’s final clinical judgment.

Composing capabilities in a clinical workflow

The three capabilities work together across the arc of a patient encounter, though you can adopt each one independently:

1. Before the visit: Patient insights reads the longitudinal record from HealthLake and produces a pre-visit summary with evidence-linked clinical narrative and HCC recapture opportunities. The clinician arrives informed.
2. During the visit: Ambient documentation captures the conversation and generates structured SOAP notes, medical codes, an after-visit summary, and a full transcript. The clinician focuses on the patient, not the keyboard.
3. After the visit: The clinician reviews and edits the documentation. If edits are made, standalone medical coding regenerates codes from the final text. The after-visit summary is shared with the patient through the portal.

All three steps use the same domain identifier, the same credentials, and the same SDK. The patient insights output can inform the ambient session by providing clinical context, and the ambient session output feeds directly into medical coding. For EHR builders, this means a single integration surface for the entire point-of-care AI workflow.

Architecture considerations for EHR builders

When you evaluate Amazon Connect Health for your environment, consider the following architectural factors: Data residency and FHIR compatibility: Patient insights reads from AWS HealthLake, which stores data in FHIR R4 format. If you already use HealthLake, you can point directly to your existing datastore. If your data lives in other systems, you will need to establish a FHIR-based data pipeline into HealthLake.

Audio capture and streaming: Ambient documentation requires real-time audio streaming from the client application. The audio format is standard PCM (16kHz, 16-bit, mono), which is compatible with most browser-based and native audio capture implementations. The streaming connection uses a bidirectional protocol, meaning transcript events flow back to the client in real time.

Output storage and retrieval: Amazon Connect Health writes all outputs (summaries, clinical documents, codes, transcripts) to Amazon S3. You control the output location and can integrate Amazon S3 event notifications to trigger downstream workflows such as populating chart fields, initiating billing processes, or sending patient communications.

Security and compliance: Amazon Connect Health runs on HIPAA-eligible AWS infrastructure. The AWS Identity and Access Management (IAM) permission model uses the health-agent signing service, and you can scope permissions to specific actions and resources. All data in transit and at rest is encrypted.

Regional availability: Amazon Connect Health is available in US East (N. Virginia) and US West (Oregon), with both runtime and streaming endpoints in each region.

Try it yourself: sample EHR integration on GitHub

To help builders get started quickly, we have published a fully working sample application that demonstrates all three Amazon Connect Health capabilities integrated into an EHR-style interface. The repository includes:

• A Python Flask backend that orchestrates patient insights jobs, Amazon Bedrock-powered pre-visit synthesis, and standalone medical coding calls
• A Java WebSocket service that bridges browser audio to the ambient documentation streaming API
• A frontend that renders pre-visit summaries with evidence-linked FHIR references, real-time transcription, SOAP notes with editable medical codes, and patient-friendly after-visit summaries
• AWS CloudFormation templates for deploying the full stack to AWS (Amazon ECS on AWS Fargate, Amazon CloudFront, Application Load Balancer, Amazon Cognito authentication)
• A demo mode that runs entirely from cached responses, requiring no AWS credentials or live API calls

The sample application walks through the complete clinical workflow described in this post: pre-visit preparation powered by HealthLake data, ambient documentation during the encounter, and code regeneration from edited notes. We built it as a reference implementation that builders can explore, adapt, and use as a starting point for their own integrations.

You can find the repository at aws-samples/sample-amazon-connect-health-point-of-care on GitHub.

Looking ahead

Amazon Connect Health represents a shift in how AI is delivered to clinical workflows, not as a separate application, but as a set of capabilities that embed directly into the systems clinicians already use. The modular, SDK-based approach means healthcare builders can start small, prove value with a single capability, and expand over time without rearchitecting their integration.

For EHR companies and healthcare ISVs, this opens a path to differentiate their solutions with AI-powered documentation, coding, and clinical preparation, without building and maintaining those capabilities in-house. For provider organizations with internal development teams, it offers a way to reduce clinician burden measurably and quickly.

The administrative burden on clinicians is not a new problem, but the tools to address it are finally catching up. Amazon Connect Health brings agentic AI to the point of care, where it can make the most difference, giving clinicians back the time they need to focus on what matters most: their patients.

To get started, try the sample application on GitHub in your development environment. Begin by implementing patient insights against a test HealthLake datastore, then layer in ambient documentation and medical coding as you expand your integration. To learn more, visit the Amazon Connect Health product page, review the AWS HealthLake documentation, or contact your AWS account team to discuss how these capabilities can integrate into your clinical systems.

Amazon Connect Health is built on HIPAA-eligible AWS infrastructure. AWS has been recognized as the Best in KLAS winner for Public Cloud for two consecutive years.