BAA-eligible LLM providers are large language model vendors and cloud platforms that will sign a HIPAA Business Associate Agreement covering specific products, configurations and data retention settings. Coverage is never automatic: it applies only to named services, often requires zero or limited data retention, and excludes consumer and many team plans.
The most expensive AI mistake in healthcare is sending patient data to a model service that is not covered by a Business Associate Agreement, and it happens constantly because coverage rules are confusing and change often. This comparison explains which major providers offer BAAs, what each covers and the traps that catch teams out, based on public vendor documentation as of September 2026. It extends our guide to BAAs with AI providers, drawing on Taction Software’s 200+ healthcare projects since 2013.
Our experts are ready to understand your business goals.
Model quality, speed and price get most of the attention in LLM selection, but in healthcare, contractual coverage comes first. If a model service processes PHI without a Business Associate Agreement, the organization is exposed regardless of how secure the vendor actually is. Coverage also shapes architecture, because covered configurations often restrict features, data retention and which endpoints can be used. Choosing a model before confirming coverage leads to painful rework. The six principles below explain how BAA coverage should shape every healthcare LLM decision, and why it belongs at the start of your AI architecture rather than the end.
HIPAA requires a Business Associate Agreement before a vendor creates, receives, maintains or transmits PHI on your behalf. Our glossary entry on the HIPAA BAA explains the requirement, and it applies to model providers exactly as it applies to hosting or billing vendors.
Providers cover named products, not their whole brand. A vendor may cover its API but not its consumer chat app, or its enterprise workspace but not team plans. Staff using the wrong product with patient data create exposure even when the organization has signed a BAA.
Covered services often require specific settings, such as zero data retention, disabled features or approved regions. Signing the agreement without applying the required configuration can leave data uncovered, so technical setup and contract terms must be verified together before any PHI flows.
Some providers extend coverage to new products through updated agreements. An older BAA may not cover features released after signing, so teams adopting new capabilities should confirm whether their existing agreement covers them or needs amendment before using them with patient data.
Every service in the chain that touches PHI, including logging, vector databases, monitoring and orchestration tools, needs coverage or must be kept away from PHI. Our BAA Network Setup service maps and configures the full chain end to end. Gaps hide there.
A BAA sets obligations for the vendor, but your organization still owns access control, minimum necessary use, audit logging, risk analysis and breach response. Coverage is a prerequisite for compliant AI, not a substitute for the safeguards HIPAA requires you to maintain yourself.
Several major model providers and cloud platforms now offer Business Associate Agreements for specific AI services. The details differ in important ways, from which products are covered to how agreements are executed and what configuration is required. The summaries below reflect public vendor documentation as of September 2026 and will change, so always confirm current terms, eligible products and required settings directly with each vendor before sending PHI. The six options below cover the providers healthcare organizations ask us about most, and our OpenAI vs Anthropic vs Gemini comparison covers model capabilities in more depth.
Anthropic offers a BAA covering its first-party API under a HIPAA-ready configuration and Claude Enterprise once an administrator enables the HIPAA setting. Free, Pro, Max and Team plans cannot be covered, and some newer features and tools sit outside coverage, so confirm each surface before use.
OpenAI offers BAAs for eligible API customers, typically tied to zero data retention on covered endpoints, and for enterprise workspace offerings through its sales team. Consumer and individual plans are not covered, so organizations must keep staff away from personal accounts when handling patient information.
Gemini models accessed through Google Cloud’s Vertex AI can fall under the Google Cloud BAA as covered services. Coverage depends on the specific services and features listed in Google’s current HIPAA documentation, so confirm each Gemini capability appears on the covered list before use.
Microsoft’s Azure OpenAI and related AI services can be covered under Microsoft’s BAA through its product terms, letting organizations use OpenAI models within Azure’s healthcare-ready environment. Our AWS Bedrock vs Azure OpenAI comparison explains how the two platforms differ. Confirm each service’s status.
Amazon Bedrock is listed as HIPAA eligible under the AWS BAA, giving access to models from several providers, including Anthropic, Meta, Mistral and Amazon, within your AWS environment. Confirm model-specific terms and eligible regions, because availability and conditions can differ between models.
Open-weight models hosted in your own cloud account or data center avoid model vendor BAAs entirely, relying instead on your infrastructure provider’s agreement or your own controls. Our on-premise LLM work covers selection, hardware and deployment for these setups. Control comes with responsibility.
Most PHI exposure through AI does not come from sophisticated attacks. It comes from simple coverage gaps: staff using personal accounts, developers testing with real data, new features adopted before agreements are updated, or logs capturing prompts in uncovered tools. These gaps are easy to create and easy to miss because everything appears to work normally. The six traps below are the ones we find most often during AI compliance reviews, and each can be closed with clear policy, technical controls and a documented inventory of which AI services are approved for PHI.
Clinicians and staff often use personal AI accounts to draft letters or summarize notes, unaware those plans cannot be covered. Clear policy, approved alternatives and network controls are needed, because banning tools without offering covered options usually pushes use further out of sight.
Engineers sometimes prototype with real patient data on uncovered accounts or default configurations. Synthetic or de-identified data should be the default for development. Our PHI redaction services and synthetic data work make safe prototyping practical. Real data belongs in validation only.
Providers release new tools and features frequently, and some launch outside existing BAA coverage. Teams should maintain an approved-feature list and review new capabilities before enabling them, rather than assuming everything inside a covered product is automatically included. Review before enabling.
Prompts and responses often flow into observability, analytics or debugging tools that were never reviewed for PHI. Logs can quietly become the largest uncovered PHI store. Every logging destination must be covered or receive only redacted content. Check every destination.
Covered configurations may apply only to specific endpoints, regions or retention settings. A single misconfigured endpoint can send PHI outside coverage. Configuration should be enforced in infrastructure code and verified regularly, not left to individual developer settings. Automate the checks.
Agent frameworks, plugins, retrieval services and third-party tools called by the model may handle PHI too. Each needs review and coverage or must be isolated from patient data, because the model provider’s BAA does not extend to services it calls on your behalf.
Once coverage is confirmed, the right provider depends on model capability for your use case, existing cloud relationships, data residency, cost, latency and how much control you need. Many healthcare organizations choose a cloud platform first and use the models available within it, while others contract directly with model providers for specific capabilities. Some use more than one. The six criteria below help organizations choose a BAA-eligible LLM provider that fits their technical, commercial and compliance needs, and they should be evaluated for the specific use cases you plan to support rather than in general.
If your organization already runs on AWS, Azure or Google Cloud under a BAA, using models within that platform often simplifies contracting, networking, identity and logging. Our AWS vs Azure vs GCP for healthcare comparison explains the broader platform differences.
Different models perform differently on summarization, extraction, reasoning and structured output. Evaluate candidate models on your own test cases, such as clinical notes or payer documents, rather than general benchmarks, because performance on healthcare tasks can differ from published leaderboard results.
Zero data retention or strict retention limits may restrict features such as stored conversations or certain tools. Confirm that your required features work within covered, low-retention configurations before committing, especially for agents and applications that depend on persistent context. Test features early.
Token pricing, throughput limits and hosting costs vary widely. Estimate costs using realistic volumes with our LLM inference cost calculator, including redaction, retrieval and logging overhead that healthcare deployments typically add. Hidden overhead often adds a meaningful share to total monthly cost.
Some organizations require data to stay within their own environment, specific regions or private networks. Self-hosted open models offer maximum control but add operational effort. Our on-prem vs cloud LLM decision framework helps weigh those trade-offs. Requirements drive the choice.
Model capabilities and terms change quickly, so avoid designs that lock you into one provider. Abstraction layers, standard interfaces and portable evaluation sets let you switch or combine providers as coverage, pricing and model performance evolve over time. Portability protects budgets.
Signing a BAA is the starting point, not the finish line. A compliant deployment requires configuration, network isolation, redaction, logging, access control and documentation that prove PHI stays within covered services. Security reviewers at hospitals and health plans will ask for this evidence before approving any AI application. The six steps below describe how we deploy LLMs under BAA coverage for healthcare clients, and our free HIPAA AI compliance checklist lets you check your current setup against the same controls before we speak. Each step produces evidence. Skipping any step leaves gaps.
List every model, platform, tool and logging destination that could touch PHI, including shadow AI use by staff. The inventory becomes the basis for coverage decisions, policies and technical controls, and it should be updated whenever new AI tools are adopted.
Confirm current BAA terms for each service, execute or update agreements and record covered products, configurations and dates. Keep copies in your compliance records, because auditors and customers will ask to see evidence of coverage for every AI service handling PHI.
Enable HIPAA settings, zero or limited retention, approved regions and disabled features exactly as each provider requires. Enforce configuration through infrastructure code and verify it regularly, so a single change cannot quietly move PHI outside covered services. Drift is caught early.
Send only the data each task needs, and redact identifiers where tasks do not require them. Minimization reduces risk even within covered services and aligns AI use with HIPAA’s minimum necessary standard for every prompt and retrieval request. Less data means less risk.
Log prompts, responses, users and model versions in covered, access-controlled systems with tamper protection. Monitor usage for unexpected patterns, such as unusual volumes or access from unapproved applications, and route alerts to named security and compliance owners. Logs need coverage too.
Prepare architecture diagrams, data flow maps, coverage records and control evidence for security reviews and audits. Clear documentation lets your team answer customer and auditor questions about AI quickly, which shortens sales cycles and approval processes significantly. Evidence closes deals.
We help healthcare organizations select BAA-eligible providers, configure compliant deployments and build AI applications on top of them. For new AI products, we work through our productized pathway with fixed prices for each stage. For existing deployments, we assess coverage gaps and fix them. Every engagement produces working infrastructure and documentation you own. The six options below describe how organizations engage us, and for developer capacity on specific platforms you can hire Claude AI developers or hire OpenAI developers through our team. Every stage price is fixed. You keep every deliverable.
Our BAA Network Setup add-on configures the full chain of covered services, from model providers to hosting, logging and vector databases, with configuration verification and compliance documentation your security reviewers can rely on. Gaps are found before any PHI flows through the system.
For new AI applications, the Discovery Sprint selects the provider, confirms BAA coverage, designs the compliant architecture and plans evaluation, ending with a fixed-price quote for building the application. You keep every artifact, including the coverage map and architecture design.
The MVP Sprint at $95,000 and Pilot-Ready Sprint at $145,000 build and harden the application on covered services, with redaction, logging and monitoring included from the first release. Coverage is verified at every stage, and evidence is packaged for your security reviewers before any pilot starts.
For AI already in use, we inventory services, check agreements and configurations, and identify gaps such as uncovered logging or consumer plan use. Assessment and remediation are scoped after an initial review and billed at our $50 blended hourly rate.
For organizations needing private or dedicated environments, our HIPAA-compliant AI hosting work deploys models and supporting services within covered infrastructure, including self-hosted open models where data must stay in your environment. Performance and cost are benchmarked before rollout. Scaling is planned upfront.
After launch, care packages cover monitoring, coverage reviews when providers change terms, configuration checks and model updates, keeping AI deployments compliant as vendors evolve their products and agreements. Configuration drift is checked monthly, and new provider features are reviewed before enablement.
Two questions matter when choosing a partner for BAA-covered AI: do they understand the contractual and configuration details of each provider, and can they build AI applications that pass demanding security reviews. Many AI developers treat compliance as paperwork, while many compliance advisers cannot configure the infrastructure themselves. Our team combines both, drawing on 200+ healthcare projects since 2013 and ISO 27001 certified processes. We sign Business Associate Agreements before accessing PHI and receive no commissions from model providers. The six points below explain what working with us on compliant AI looks like.
We build on Anthropic, OpenAI, Google, Microsoft, AWS and open models, and recommend whichever fits your use case, cloud and budget. We have no reseller arrangements, so recommendations reflect your needs rather than a commercial relationship with any single model provider.
We verify HIPAA settings, retention, endpoints, regions and network isolation with real tests, not screenshots. Verification catches the misconfigurations that silently move PHI outside coverage, which are far more common than most teams expect during their first healthcare AI deployments.
We map every service touching PHI, from models to logging and vector stores, and ensure each is covered or isolated. Full-chain thinking closes the gaps that model-only reviews miss, especially in agent and retrieval systems with many connected components. Nothing is assumed.
We produce the architecture diagrams, coverage records and control evidence that hospital and health plan security teams request. Your team can answer AI security questionnaires confidently in days rather than weeks, which shortens procurement for healthcare AI products. Buyers notice the difference.
For new AI applications, our productized pathway publishes fixed prices, so leaders approve compliant AI investment with a known budget. Gap assessments for existing systems are scoped after review, so you pay only for problems that actually exist. Budgets stay predictable.
Infrastructure code, configurations, documentation and compliance records belong to you. We hand everything over in documented form, so your team can maintain compliant AI deployments internally or continue with our ongoing care packages as providers change terms. No vendor lock-in applies.
These are the questions CIOs, compliance officers, AI teams and founders ask most often about BAA-eligible LLM providers, whether they are choosing a first model, auditing current AI use or preparing for customer security reviews. The answers are short on purpose, reflect public documentation as of September 2026 and are not legal advice, so confirm current terms with each vendor and your counsel. If your question depends on your stack, a short call with our team will help. For a deeper technical walkthrough, read our guide to BAAs with OpenAI, Anthropic and AWS Bedrock.
As of September 2026, Anthropic, OpenAI, Microsoft Azure OpenAI, Google Cloud Vertex AI and AWS Bedrock all offer BAA coverage for specific products and configurations. Coverage excludes consumer and many team plans, so confirm the exact product, settings and current terms with each vendor.
No AI product is HIPAA compliant by itself. Specific business and API offerings can be covered by a BAA when configured correctly, but consumer plans cannot. Compliance depends on the covered product, required configuration and the safeguards your organization implements around it.
Not necessarily. Some providers exclude certain features, tools or newer capabilities from coverage, and older agreements may not cover products released after signing. Maintain an approved-feature list and confirm coverage before enabling new capabilities with patient data. Assumptions are risky.
Yes. Self-hosted open models avoid model vendor BAAs, relying on your infrastructure provider’s BAA or your own environment. They offer strong control but require operational effort for hosting, scaling, security and evaluation, which we help organizations plan and run. Many teams combine both.
Using uncovered consumer AI with PHI can be an impermissible disclosure requiring breach risk assessment and possibly notification. Organizations should provide covered alternatives, train staff and apply technical controls, then investigate and document any incidents promptly with counsel. Prevention costs less.
For new applications, our productized pathway starts with a $45,000 Discovery Sprint, then $95,000 MVP and $145,000 Pilot-Ready Sprints. Coverage gap assessments are scoped at our $50 hourly rate. Model usage and hosting fees are separate. Every stage price is fixed.
Share which models and AI tools your teams use today, how PHI reaches them and any upcoming security reviews. In a 30-minute call we will identify likely coverage gaps and outline how to close them quickly. Book a free consultation.
Your email address will not be published. Required fields are marked *
Our expert reaches out shortly after receiving your request and analyzing your requirements.
If needed, we sign an NDA to protect your privacy.
We request additional information to better understand and analyze your project.
We schedule a call to discuss your project, goals. and priorities, and provide preliminary feedback.
If you're satisfied, we finalize the agreement and start your project.