HIPAA compliant AI, three ways to deploy it
HIPAA compliant AI means the whole system around the model meets the HIPAA Privacy and Security Rules: a Business Associate Agreement, encryption, access controls, audit logs, and no training on your patients' data. Mother Nature AI provides that three ways: access to our health models under a BAA, the same models inside your private cloud, or air-gapped models fine-tuned for your organisation and delivered on hardware you control.
Mother Nature AI maintains HIPAA-compliant safeguards for applicable protected health information. Business Associate Agreements available for covered entities.
The definition
What “HIPAA compliant” means for an AI system
There is no HIPAA certificate. The Department of Health and Human Services does not certify software, so any vendor claiming to be “HIPAA certified” is describing a self-assessment at best. What exists is a set of obligations on covered entities and their business associates, and an AI product is compliant when the organisation deploying it can meet those obligations with the product in place.
In practice that comes down to a handful of controls. PHI must be encrypted in transit and at rest. Every person and service that can read it must be identified, authorised and logged. The vendor must sign a BAA accepting its share of liability, and must not reuse your data for anything outside the contract, including training its models. The system should be designed on the minimum-necessary principle, so the model sees only the fields a task requires. And the organisation must document a risk analysis showing it thought about all of this before go-live.
Large language models add two questions that older software did not raise: where do prompts and outputs go, and what is the model learning from them. If you cannot answer both in writing for a given product, it is not ready for PHI.
Deployment
Hosted, private cloud, or air-gapped
The same health models, deployed where your policy allows. Most organisations start hosted and move inward as their use grows.
Hosted API with a BAA
Clinics, digital health products and care teams that want to move fast.
- PHI processed under a BAA
- No training on your data
- Regional data pinning
- Days to deploy
- Data leaves your network, under contract
Private cloud
Health systems and payers with their own cloud tenancy and security review process.
- Models run inside your VPC
- Your keys, your logs, your IAM
- Fine-tuned on your documents
- Weeks to deploy
- Requires your cloud and security team's time
Air-gapped appliance
Research institutions, government and anyone whose policy says PHI never leaves the building.
- No outbound network path
- Models fine-tuned for your specialty
- Managed updates delivered offline
- Fixed, predictable cost
- Capacity is what the hardware can serve
Fine-tuned for your organisation
Private-cloud and air-gapped deployments can be fine-tuned on your own protocols, formularies, note templates and specialty literature, so outputs match how your clinicians already write and decide. The training data, the resulting model weights and every inference log stay inside your boundary. Shared hosted models are never trained on customer PHI.
Due diligence
Eight things to get in writing from any AI vendor
We answer all eight. Use the list on everyone else too.
- A signed Business Associate Agreement before any PHI is sent
- Customer data contractually excluded from model training
- Encryption in transit (TLS 1.3) and at rest (AES-256)
- Role-based access, MFA, and tamper-evident audit logs
- Data residency you can specify, and subprocessors you can see
- Minimum-necessary design: the model sees only what the task needs
- A private-cloud or air-gapped option when policy requires it
- Documented incident response and breach notification timelines
Use cases
What teams run on it
Clinical documentation
Structured notes and visit summaries from the encounter, in your templates, to cut charting time.
Lab and result explanations
Plain-English, patient-ready explanations of markers and trends, reviewed by the clinician before release.
Lab result explainer →Patient summaries and intake
Provider-ready summaries from wearables, medications, symptoms and records, aligned to how your team makes decisions.
Patient summary →Interaction and safety checks
Drug, supplement and herb interactions with clear explanations, drawing on a library of >250 evidence-graded profiles.
Interaction checker →Care facilities also run VitalIQ, our resident-monitoring dashboard, on the same infrastructure.
Frequently asked questions
What makes an AI system HIPAA compliant?
HIPAA does not certify software. A system is compliant when the organisation running it meets the Privacy, Security and Breach Notification Rules for any protected health information (PHI) it touches: a signed Business Associate Agreement with the vendor, encryption in transit and at rest, access controls and audit logs, minimum-necessary data handling, workforce training, and a documented risk analysis. An AI model is one component inside that program, not a substitute for it.
Can I use a public chatbot with patient data?
Not with identifiable PHI unless the vendor will sign a BAA and contractually excludes your data from model training. Most consumer chat products do neither. Pasting a patient's note into one is a reportable disclosure.
Is Mother Nature AI HIPAA compliant?
Mother Nature AI maintains HIPAA-compliant safeguards for applicable protected health information. Business Associate Agreements available for covered entities. Health data is encrypted at rest using AES-256 and in transit using TLS 1.3. Details are published on our Security and Privacy page.
What is an air-gapped LLM and when do you need one?
An air-gapped model runs on hardware you control with no outbound network path, so PHI never leaves your environment. It is the right choice for organisations with data-residency rules, research data that cannot be shared, or security teams that will not approve a cloud vendor. The trade is operational: you own patching, capacity and model updates, which is why we deliver it as a managed appliance with fine-tuning on your own documents and workflows.
Does fine-tuning on our records create a privacy problem?
Only if it happens on infrastructure you do not control. Fine-tuning inside your private cloud or on an air-gapped appliance keeps the data, the resulting weights and the inference logs within your boundary. We do not train our shared models on customer PHI.
How is PHI de-identified for AI use?
Under the Safe Harbor method, 18 identifier types are removed before data is used for analytics or model evaluation. Where identifiers are needed for care delivery, the data stays inside the covered workflow under the BAA rather than being de-identified.
What should we ask any HIPAA AI vendor before signing?
Will you sign a BAA. Is customer data ever used for training. Where is data processed and stored, and can it be pinned to a region. What is logged, for how long, and who can read it. How are subprocessors disclosed. Can we get a private or on-premises deployment. What is the breach notification timeline. A vendor who cannot answer these in writing is not ready.
Explore more
Request a demo
Tell us where your PHI is allowed to live and we will show you the deployment that fits: hosted under a BAA, your private cloud, or air-gapped.