The Direct Answer
Claude — Anthropic's model family — is not HIPAA compliant out of the box. Claude usage can be made HIPAA-aligned in two ways:
- Through Anthropic's API under a Business Associate Agreement (BAA) signed with Anthropic, on approved tiers.
- Through Amazon Bedrock or Google Vertex AI, where the hyperscaler's BAA covers the inference path and Claude runs inside the customer's cloud account.
The Claude.ai consumer product and the Claude for Teams tier do not support HIPAA workloads. Neither does Claude for Enterprise unless you have an explicit BAA in place.
That is the contractual layer. The harder question — the one your compliance team will ask — is where PHI lives during inference, what the audit chain looks like, and who you fall back to if Anthropic changes its terms or its model lineup.
What HIPAA Actually Asks of Your AI Stack
HIPAA's Security Rule applies to every system that handles Protected Health Information. An AI inference endpoint that receives a prompt containing PHI is a covered system, full stop. The Security Rule requires:
- Access control — only authorized workforce members can submit PHI prompts.
- Audit controls — every prompt that touched PHI is logged and reviewable.
- Integrity controls — PHI isn't tampered with in flight or at rest.
- Transmission security — PHI is encrypted between every hop.
- Authentication — identities are verifiable across the entire chain.
A BAA is the legal commitment that the business associate (Anthropic, AWS, or Google) will meet these safeguards. The technical implementation determines whether the safeguards actually hold up under an audit.
The Three Routes to "Claude with a BAA"
Anthropic Direct API
Anthropic offers a BAA on its API under specific commercial tiers. The covered scope is the API itself — prompts, responses, and the inference processing performed by Anthropic's infrastructure. The customer is responsible for everything upstream (workforce access, prompt construction, identity) and downstream (storage of responses, downstream system integration).
This is the cleanest route if you want Claude as your primary model and you can accept that PHI is processed on Anthropic's infrastructure.
Amazon Bedrock
Claude is available as a managed model on Amazon Bedrock. Bedrock runs inside your AWS account, in your VPC, with AWS's BAA covering the entire infrastructure layer. The inference still calls into a managed Claude endpoint, but the data path stays inside your AWS perimeter.
For organizations already on AWS with HIPAA-compliant accounts, this is often the path of least friction. The trade-off is that you're now coupled to AWS for that workload.
Google Vertex AI
Claude is also available through Vertex AI on Google Cloud, under Google's BAA, with the same shape as the Bedrock path: managed model, customer cloud, hyperscaler BAA. The trade-off is again the hyperscaler coupling.
What All Three Routes Have in Common
In all three routes, the inference is happening on someone else's infrastructure — Anthropic's, AWS's, or Google's. The BAA covers the legal layer. The technical control plane is partially yours and partially the vendor's.
For many healthcare organizations, that is fine for general-purpose workloads. For PHI-heavy workloads — clinical documentation, medical coding, claims adjudication, member services — the leadership team often wants tighter control:
- Audit logs that flow into the hospital's SIEM, not a vendor's dashboard.
- Identity that maps to the hospital's IdP, not a vendor's user store.
- Network paths that never leave the hospital's data center for the most sensitive prompts.
- A fallback plan if Anthropic deprecates a model the clinical workflow depends on.
That is where the SaaS-with-a-BAA posture starts to feel thin, and where most organizations begin asking what a self-hosted AI posture looks like.
The Self-Hosted Alternative
A self-hosted AI platform changes the architecture in three ways:
- The AI platform itself runs inside the covered entity. Not in an Anthropic data center, not in a vendor's managed cloud — in the hospital's own VPC, data center, or air-gapped environment.
- The model is your choice, per workload. Frontier models (Claude, GPT, Gemini) through their APIs for non-PHI workloads when frontier quality matters; open-weights models (Llama, Mistral, Qwen, Phi) running locally for PHI-touching workloads.
- The governance layer is yours. Every prompt, every response, every model invocation logs into your SIEM. Every BAA you sign covers the actual flow. The audit story is the story you control.
This is the model the ibl.ai platform implements. The customer gets the complete source code on day one. The platform deploys inside the customer's infrastructure. Claude, when used, is reached through the routes above — through Anthropic's BAA-covered API or through Bedrock — but the platform sitting in front of it keeps the audit trail and the governance layer under the customer's control.
For PHI workloads that should never leave the perimeter at all, the platform routes to a local open-weights model. No external API call, no BAA needed, no vendor dependency on that path.
This is the build-and-buy posture: a production platform on day one (you don't engineer from scratch), with full source-code ownership (you aren't locked in if priorities change). It is the architecture most HIPAA-aware health systems we work with end up at after their first generation of SaaS AI deployments.
A Decision Framework for Healthcare Buyers
For a hospital or health system evaluating Claude for clinical use, the practical decision tree is:
- Is this workload PHI-heavy? If yes, prefer a self-hosted platform with a local model for inference, and reserve Claude for non-PHI augmentation.
- Is this workload sometimes-PHI? Route through a self-hosted platform that can call Claude via Bedrock or Anthropic's BAA-covered API, with the audit trail captured in your SIEM.
- Is this workload definitely non-PHI? Any of the BAA-covered Claude routes is fine; pick on price, latency, and developer experience.
The key is that you do not have to pick one model and one architecture. The whole point of a sovereign AI platform is that you can route per workload, change models as the field evolves, and keep the governance layer constant.
What to Take Away
- Claude is not HIPAA compliant by default.
- Claude can be used in HIPAA-regulated workflows under Anthropic's BAA, through Amazon Bedrock, or through Google Vertex AI — each with a BAA from the respective vendor.
- A BAA is the contractual layer. The technical layer — where PHI lives, who owns the audit trail, what your fallback is — is what compliance auditors test.
- The strongest HIPAA posture is a self-hosted AI platform with model choice per workload, audit logs in your SIEM, and full source-code ownership.
See how ibl.ai's self-hosted and private LLM platform handles HIPAA-aligned Claude routing, and how the Healthcare solution is structured. For procurement, the Claude for Education alternative page is the side-by-side on ownership, model choice, and cost.