HIPAA · Building AI Automations

Building HIPAA-Compliant AI Automations: AWS vs. Google Cloud

A vendor-neutral guide for healthcare and med-tech teams choosing a platform to build AI automations — calling AI models by API, wiring up agents, and moving patient data through them compliantly. Written to be readable by a clinical or executive decision-maker, with enough builder detail to actually choose. No winner is declared.

Built from official AWS and Google Cloud documentation, rechecked November 6, 2026. Every claim is cited inline to its source. Service lists and rules change fast; confirm against the live pages before you build.

Review the pre-build decisions

Start here — the decision you're actually making

You're choosing where to build AI automations that touch patient data. Read this framing first.

An "AI automation" here means software your team builds that calls an AI model through an API (a programmatic request — e.g., "summarize this note," "extract the diagnoses," "answer this patient question") and chains those calls into a workflow or agent (software that reasons over several steps and uses tools on its own). If that workflow ever touches PHI (Protected Health Information — patient data tied to a person), it falls under HIPAA, and the platform underneath has to support that. Both AWS and Google Cloud do.

Three ground rules shape everything below:

  • No cloud — and no AI model — is "HIPAA certified." There's no such stamp. AWS says it plainly: "There is no HIPAA certification for a cloud service provider."1 Google says the same — "there is no certification recognized by the US HHS for HIPAA compliance."5 Be skeptical of any AI vendor claiming to simply be "HIPAA compliant" out of the box.
  • It's a shared responsibility. The cloud secures the infrastructure and the model service; your team is responsible for configuring it correctly and only sending PHI to approved services in an approved way.1 5 This is where most AI-automation compliance mistakes happen — see the #1 risk.
  • The BAA is the gate. Before any PHI flows through your automation, your organization signs a BAA (Business Associate Agreement), the contract in which the vendor commits to protect that data. Both clouds make it self-service.1 5
Not legal advice This guide explains how each platform supports building HIPAA AI automations so your team can ask the right questions. Whether a specific build is compliant depends on configuration and a signed BAA, confirmed with your legal / privacy / compliance owner.

The 30-second answer

If you read one section, read this one.

Both vendors publish HIPAA-supported service scopes and BAA processes. This guide helps engineering and compliance teams compare a proposed data flow; neither a model catalog nor a signed contract establishes that every model, feature or deployment is suitable for PHI. AWS supported scope Google supported scope

The differences that actually matter for AI automations are narrow, and each platform has a distinct strength:

  • AWS exposes explicit inference retention modes. Evaluate processing Regions, project overrides, model requirements and separate storage surfaces. AWS retention controls
  • Google publishes supported-service scope and feature-specific retention controls. Its training restriction is not a promise that prompts are never retained. Google retention controls

If you're already building on one of them, the strong default is to stay — you should not switch clouds for compliance reasons alone.

Side-by-side comparison

Note: Confirm the exact services, models and settings for your deployment before sending patient data.

Six dimensions that matter when building AI automations on patient data. Each cell gives the finding plus what it means for you, cited to its source. Genuine ties are labeled honestly.

DimensionAmazon Web Services (AWS)Google CloudBottom line
1 · Signing the BAA (the gate)Self-service in a console tool (AWS Artifact); one acceptance can cover all your organization's accounts. PHI must run on a service from the HIPAA-eligible list.1 2Self-service in the console; covers Google's infrastructure plus a named list of covered products. PHI must run on a covered product.5TieBoth self-serve and free; both gate PHI to an approved list of services.
2 · Calling AI models by APIAmazon Bedrock — API access to many foundation models incl. Claude, Llama, and Amazon's own, all HIPAA-eligible, with SDKs.2Gemini Enterprise Agent Platform — API access to Gemini plus a model garden that also includes Claude.5TieBoth give HIPAA-eligible API access to top models — including Claude on either.
3 · Building & running agentsBedrock AgentCore (HIPAA-eligible) for multi-step agents, plus Guardrails to filter sensitive data in/out.2Assess the exact managed agent runtime against the BAA-supported service list. Encryption and network controls are feature-specific, not blanket assurances.TieBoth are production-grade agent stacks. Match to your team's skills.
4 · Keeping PHI safe in model calls (the risk)Per-Region retention modes and mantle project overrides. Models requiring AWS review are blocked under none. Legacy provider_data_share does not currently send content to providers. AWS retention modesTraining restrictions do not guarantee zero retention. Review abuse monitoring, grounding, Interactions API storage, session resumption and caching for the exact model. Google retention controlsTie — different mechanismsAWS gives a switch you must set; Google gives a standing commitment plus controls. Both need per-model diligence.
5 · Prebuilt healthcare building blocksHealthLake (clinical data store), HealthScribe (ambient note drafting), Comprehend Medical (clinical-text extraction) — all HIPAA-eligible.2Cloud Healthcare API and Healthcare Data Engine for ingesting/managing clinical data in standard health-data formats.5Slight edge: AWSMore ready-made clinical AI today; both have standards-based data stores.
6 · Cost & supportModel/service usage and support are priced separately. AWS Support lists Business Support+, Enterprise and Unified Operations. Bedrock pricingGoogle offers its supported HIPAA services at the same product pricing available to other customers; model, location and support tier still determine cost. Google model pricingCompare a concrete workload and support contract; HIPAA eligibility does not make total deployment costs equal.

Editorial comparison, not a score of compliance or total cost. Review the exact services and data flow before choosing.

1 · The BAA gate

The first thing your organization does on either cloud, before any PHI moves through an automation.

The BAA is the vendor's legal commitment to protect the patient data you send it. On both clouds it's self-service — no procurement cycle, no lawyer-to-lawyer negotiation for the standard agreement. On AWS you accept it in a console tool called AWS Artifact, and a single acceptance can apply to every account your organization manages.1 On Google Cloud you accept it in the console.5

On both clouds, signing the BAA does not bless the entire platform for PHI. AWS gates PHI to a published HIPAA-eligible services list;2 Google's BAA covers its infrastructure and a published list of covered products. In practice both work the same way — check the list, keep PHI on it.

Agreement coverage must be confirmed for the relevant account or organization before PHI. Do not assume same-day approval or universal zero cost from a self-service workflow.

2 · Calling AI models by API — the core of any AI automation

Compare the exact model version, endpoint, processing location and contractual scope before sending patient data. Catalog availability and HIPAA eligibility are separate checks.

  • AWS — Amazon Bedrock. A single API/SDK to call many foundation models, including Anthropic's Claude, Meta's Llama, and Amazon's own models. Bedrock is HIPAA-eligible.2
  • Google Cloud — Gemini Enterprise Agent Platform. API access to Google's Gemini models plus a model garden that also includes Claude.5 Current Google Claude API catalog

Editorial decision: choose a model only after its account access, supported-service coverage and data-handling requirements are reviewed. Neither cloud is automatically suitable for every model workload.

3 · Building and running agents

Beyond one API call: frameworks to chain steps, use tools, and run agents in production.

  • AWS lists Bedrock AgentCore as HIPAA-eligible. Guardrail text filters can block or mask detected PII probabilistically; tool-call arguments, logs and traces need separate protection. AWS filter limitations
  • For a Google managed agent runtime, confirm the exact service and enabled features in the BAA scope. Review feature-specific encryption and network controls; an SDK or framework name alone does not establish coverage. Google supported services and limitations
Product names do not establish BAA coverage. The current Google list separately names Gemini Enterprise Agent Platform and Vertex AI Workbench; check the exact service and enabled features. Google supported products

What it means for you: a tie. Both are credible, production-grade agent platforms.

4 · The #1 PHI risk — what happens to your prompts

Note: Confirm the exact services, models and settings for your deployment before sending patient data.

This is where AI-automation compliance is won or lost. "The platform is HIPAA-eligible" does not mean every model is safe for PHI by default.

When your automation sends a prompt to a model, you must know: does anyone store or see that prompt afterward? A prompt can contain PHI, so retention and provider-sharing are the core HIPAA questions.

AWS — an explicit retention switch 3

  • none blocks inference models requiring retention; review the separate flagged-image abuse-detection exception.
  • aws_review permits AWS retention and review where the model requires it; content remains within AWS.
  • provider_data_share is legacy; AWS says it does not currently share content with model providers. New configurations should use aws_review where review is required.

Configure every processing Region and check mantle project overrides. A model whose required mode exceeds your effective mode is unavailable; setting none is an editorial choice when your risk assessment requires ZDR, not a universal HIPAA rule. AWS mode and scope documentation Model-specific retention and image exceptions

Google Cloud — a standing commitment plus controls 6

Google restricts training without permission or instruction, but retention depends on features. Advanced AI may prevent ZDR. Search grounding keeps derived queries/context for up to three days; Maps grounding keeps data for 30 days. Interactions API store defaults to true. Live session resumption and implicit caching can keep data for up to 24 hours; supported models may use durable caching without CMEK. Assess or disable each applicable feature. Current Google retention and caching controls

Before PHI, confirm the BAA covers each service and assess model, endpoint, Region, storage, logs and tool destinations. Encryption controls vary by feature; do not infer that enabling a customer-managed key covers every retention surface. AWS controls Google controls

Editorial conclusion: neither platform choice alone establishes compliance; review the complete data flow and contractual obligations.

5 · Prebuilt healthcare building blocks

How much clinical capability you get without building it yourself.

  • AWS has the deeper set of ready-made clinical AI, all HIPAA-eligible: HealthScribe (turns a visit conversation into a draft note), Comprehend Medical (extracts diagnoses, medications, and procedures from free-text notes), and HealthLake (a clinical data store using FHIR).2
  • Google Cloud offers the Cloud Healthcare API and Healthcare Data Engine — strong, standards-based services for ingesting and managing healthcare data (FHIR, DICOM, HL7v2) — but fewer turnkey "clinical AI" features.5

What it means for you: a slight edge to AWS if you want clinical AI capabilities working out of the box.

6 · Cost & support

Note: Confirm the exact services, models and settings for your deployment before sending patient data.

Is there a "HIPAA tax," and who picks up when your automation breaks?

Cost: Google states supported HIPAA services use the same products and pricing available to other customers. This does not promise equal costs across vendors or all configurations. Obtain a workload-specific estimate rather than treating a BAA as a price guarantee. Google HIPAA features and pricing

Support: AWS lists Business Support+, Enterprise and Unified Operations; Google lists Standard, Enhanced and Premium. Compare severity-specific response targets, coverage hours, exclusions and quoted fees for the plan you buy; response targets are not resolution guarantees. AWS Support plans Google Customer Care

Editorial decision: estimate model usage, storage, logging, network and support together. The BAA is a contractual requirement, not a total-cost comparison.

Price & confirm for your specific build

Note: Confirm the exact services, models and settings for your deployment before sending patient data.

Compliance is broadly a tie. These build-time variables often decide the real cost and feasibility.

  • Model pricing varies by provider/model, location, tier and usage. Compare input/output, caching, grounding and other feature charges for the intended workload. AWS model pricing Google model pricing
  • Model & region availability under the BAA. Confirm the model versions you want are available and BAA-eligible in your required region.
  • Where inference runs (data residency). When a PHI-containing prompt is sent, which region processes it? Verify you can pin it to your required geography.
  • Rate limits & latency. Real-time clinical workflows can hit throttling. Check each platform's rate limits and typical latency for your call volume.
  • Audit logging. Both clouds log activity, and both can log model calls — but prompt/response logs can contain PHI, so those logs must be encrypted and access-controlled.
  • Fine-tuning / bring-your-own-model. If you plan to fine-tune on clinical data, confirm that path is HIPAA-eligible and understand where the training data and resulting model live.

Which is the better fit if…

Note: Confirm the exact services, models and settings for your deployment before sending patient data.

No single winner — but clearer answers once you know your situation.

…you already build on one of them

Editorial judgment: start with existing infrastructure when it meets the workload requirements. Service eligibility does not establish that either cloud supports every proposed PHI feature or model.

…you specifically want Claude under a BAA

Either works. Claude is available HIPAA-eligible on AWS Bedrock2 and in Google's model garden.5

…your compliance team wants a provable "retain nothing" control

AWS exposes explicit retention modes and policy controls. Evaluate model availability under your effective mode and review separate abuse-detection exceptions. AWS retention controls

…you want the simplest "what's in scope" story

Lean Google Cloud. Its BAA spans the whole infrastructure and its privacy commitment is platform-wide.5

…you want clinical AI out of the box (ambient scribe, clinical NLP)

Lean AWS. HealthScribe and Comprehend Medical are hard to match today without building it yourself.2

...you need to compare regulated-workload costs

Use the actual model, location, features and support quote. Google states supported HIPAA products use its standard product pricing; this alone is not a reason to declare a cheaper platform.

Before you build — five things to confirm

This guide is a starting map, not a compliance sign-off. Verify these with your team or vendor.

  • Assess every model and endpoint before PHI. If the risk assessment requires ZDR, configure none in each AWS processing Region and check project overrides; on Google assess every enabled retention feature. AWS retention Google retention
  • "Eligible/covered" is not "compliant." A service being on the approved list only means you're allowed to use it with PHI once your BAA is signed and it's configured correctly.
  • The lists and brands change. Both vendors update their approved-service lists regularly — re-check the live pages when you build.
  • Price out the build-time variables. See Price & confirm for what actually decides cost and feasibility.
  • This is not legal advice. Final HIPAA sign-off belongs with your legal, privacy, or compliance owner.
The one rule to carry away On both AWS and Google Cloud, HIPAA compliance is a shared responsibility: the vendor secures the platform, but signing the BAA and configuring your automation correctly — especially each model's data handling — are on you. Choosing the platform is the easy part; building on it compliantly is the ongoing job.

Sources

Every numbered citation above links here. Original source collection: June 12, 2026.

Amazon Web Services

  1. 1AWS HIPAA Compliance — self-service BAA via AWS Artifact, shared-responsibility framing, no-HIPAA-certification statement.
  2. 2AWS HIPAA Eligible Services Reference — current HIPAA-eligible service scope, checked November 6, 2026; customers still configure their workloads.
  3. 3Amazon Bedrock — Data retention — retention modes, zero-data-retention behavior, model availability under the effective retention mode, IAM/SCP enforcement.
  4. 4AWS Support Plans — Business Support+, Enterprise, and Unified Operations; 24/7 coverage; response targets; named Technical Account Manager.

Google Cloud

  1. 5HIPAA Compliance on Google Cloud — self-service BAA, covered-products list (Gemini Enterprise Agent Platform, Vertex AI Workbench, Cloud Healthcare API, Healthcare Data Engine), no-HIPAA-premium statement.
  2. 6Google platform retention and caching controls — training restriction, abuse monitoring, grounding, storage and caching exceptions; checked November 6, 2026.
  3. 7Google Cloud Customer Care overview — Standard, Enhanced, and Premium tiers; Technical Account Manager at Premium.
Original research: June 2026. Built from live official AWS and Google Cloud documentation on 2026-06-12. Both vendors change their approved-service lists, AI/agent offerings, model data-handling, support terms, and pricing over time — re-verify against the linked sources before building a real PHI workload. Not legal advice.