HIPAA-compliant AI development
Five safeguards make an AI build HIPAA-compliant: a BAA chain that covers the model provider (Claude via AWS Bedrock, GPT via Azure OpenAI), a no-training API tier, PHI redaction before every model call, an audit log on every PHI read and write, and role-scoped access controls.
Expect $25K–$50K and 4–8 weeks — the compliance substrate is what moves a build out of the $10K–$25K tier. Wolrix shipped a HIPAA-aware telemedicine platform in 8 weeks and a UK digital pharmacy in 10.
True HIPAA compliance is a covered-entity certification — the application vendor (Wolrix) ships the substrate, the client signs the BAA and runs certification. The substrate is five safeguards: BAA covering the model provider, a no-training API tier, PHI redaction before every model call, an audit log on every PHI read and write, and row-level access controls. AI features always sit behind human-in-loop on irreversible PHI actions. We shipped a HIPAA-aware telemedicine platform in 8 weeks — doctor portal, video consults, e-prescriptions, Stripe, S3 — and a UK digital pharmacy in 10.
The five safeguards that make an AI build HIPAA-compliant
Every one of these shipped on the telemedicine and pharmacy builds. Miss any one and the build fails audit.
| Safeguard | What it means | How Wolrix ships it |
|---|---|---|
| BAA covering the model provider | The covered entity signs a Business Associate Agreement with every entity that stores or processes PHI — hosting, database, email, and the AI provider itself. | Claude via AWS Bedrock under the AWS BAA, or GPT via Azure OpenAI under the Azure BAA. Never a consumer endpoint for PHI. |
| No-training API tier | Model inputs and outputs are contractually excluded from provider training data, with zero data retention where the tier supports it. | Enterprise API tiers only. This is a contract term you verify before the first line of code, not a toggle you find later. |
| PHI minimization and redaction | Identifiers are stripped or tokenized before the prompt leaves your infrastructure. The model sees the minimum necessary to do its job. | An application-layer redaction step in front of every model call — the default on every Wolrix healthcare build, regardless of provider BAA status. |
| Audit logging | Every PHI read and write is recorded: who accessed what, when, and through which route. Auditors ask for this first. | A Postgres audit table written from day one. Retrofitting audit logging into a live schema is the most expensive mistake we see. |
| Access controls | A user sees only the records their role permits. A clinician sees their patients, not the whole database. | Row-level security in Postgres plus role checks on every API route. Enforced in the database, not just the UI. |
The phrase “HIPAA-compliant software” is a category error.
HIPAA is a regulatory framework that applies to covered entities — healthcare providers, health plans, and clearinghouses that handle PHI. Software vendors are business associates, governed by a Business Associate Agreement (BAA) signed with the covered entity.
A vendor cannot be “HIPAA-compliant” in isolation. The vendor builds technical safeguards that hold up under audit. The covered entity runs the compliance program, signs BAAs with every business associate (hosting, AI provider, email, analytics), and goes through the formal certification.
What you should ask a vendor is not “are you HIPAA-compliant”. It's “does your application layer pass a real HIPAA audit when our covered entity runs it”. The answer for Wolrix builds: yes — we've shipped builds that have.
Vendor vs covered entity
What Wolrix ships (the substrate)
- •Encrypted-at-rest schemas in Postgres
- •Audit log table on every PHI read and write, shipped day one
- •Row-level security so a clinician sees only their patients
- •Secrets in Vercel + AWS-managed env, never in code
- •Role-based access on every API route
- •Human-in-loop gates on irreversible PHI actions
- •AI provider routing under BAA-eligible deployments (Bedrock or Azure OpenAI)
What the covered entity owns (the certification)
- •Signing the BAA with hosting and AI vendors
- •Designating a HIPAA security officer
- •Running the certification process and audit
- •Workforce training and access reviews
- •Incident response process and breach notification
- •Annual risk assessment
- •Patient-facing notice of privacy practices
The three-layer pattern for healthcare AI
AI as a tool with a human approval gate on irreversible actions, never as the final layer. This is the architecture we ship on every healthcare build.
Retrieval against your data
AI features pull context from your authoritative data store (Postgres + pgvector or a managed RAG layer), not from the model's training set. The model never invents a drug name, a dosage, or a patient detail — it cites a retrieved record.
Function calling for structured output
Where AI proposes a write action, output is constrained to a typed schema (Zod or JSON Schema). The application validates the proposal against business rules before it ever touches the database.
Human-in-loop on irreversible actions
Anything that touches PHI in a way you can't undo — sending a prescription, signing a note, billing a claim, messaging a patient — routes through a human approval queue. AI drafts; clinician signs.
The 8-week telemedicine reference build
Doctor-patient portal with live video consults, e-prescription flow, Stripe billing, S3 document handling, audit log on every PHI read and write. Doctor and patient roles, scoped sessions, consent capture. Video consults via WebRTC with recordings encrypted at rest. E-prescription workflow with provider sign-off. Stripe payments with insurance-aware itemization. Anonymized; references on call after NDA.
What HIPAA compliance adds to an AI build
The delta between a regulated and a non-regulated build is the compliance substrate: the audit table, row-level security, the redaction layer, consent capture, and human-approval queues. That work is why healthcare builds land in the Scale tier. Full tier breakdown on the pricing page.
| Build | Cost | Timeline | What's in it |
|---|---|---|---|
| Non-regulated AI build (Build tier) | $10K–$25K | 2–4 weeks | No PHI. Standard auth, no audit table, no redaction layer. |
| HIPAA-aware AI build (Scale tier) | $25K–$50K | 4–8 weeks | Audit logging, row-level security, PHI redaction, consent capture, human-approval queues. |
| Telemedicine platform (shipped) | Scale tier | 8 weeks | Video consults, e-prescriptions, Stripe, HIPAA-aware audit log. |
| UK digital pharmacy (shipped) | Scale tier | 10 weeks | Prescription verification flow, Stripe, regulatory compliance. |
Scoping starts with a 15-minute intro call or the booking page. Both shipped healthcare builds are documented in the case studies; model-choice tradeoffs for regulated workloads are in Claude vs GPT for healthcare software.
HIPAA + AI questions
What makes AI development HIPAA-compliant?
Five safeguards: a BAA chain that covers the model provider (Claude via AWS Bedrock, GPT via Azure OpenAI), a no-training API tier so PHI never enters provider training data, PHI minimization and redaction before every model call, an audit log on every PHI read and write, and role-scoped access controls enforced in the database. Plus a human approval step on any irreversible PHI action — AI drafts, a clinician signs.
How much does HIPAA-compliant AI development cost?
Expect $25K–$50K and 4–8 weeks for a HIPAA-aware AI build — the Wolrix Scale tier. A comparable non-regulated build runs $10K–$25K in 2–4 weeks. The difference buys the compliance substrate: audit logging from day one, row-level security, PHI redaction, consent capture, and human-approval queues. Wolrix shipped a telemedicine platform in 8 weeks and a UK digital pharmacy in 10.
Do you sign a BAA?
Wolrix is the development vendor, not the hosting or AI vendor. The BAA is between the covered entity and the entities that store or process PHI (Vercel, AWS, Anthropic, OpenAI). We help structure the BAA chain on the spec call and recommend BAA-eligible deployment paths (AWS Bedrock for Claude, Azure OpenAI for GPT).
Can AI write to a patient record without a human approving?
Not on a Wolrix build. AI features always sit behind a human approval step on irreversible PHI actions. AI drafts the note, the clinician signs. AI proposes a prescription, the prescriber approves. AI suggests a message, the nurse sends. This is the third layer of the anti-hallucination architecture.
What healthcare apps have you actually shipped?
A HIPAA-aware telemedicine platform with video consults, e-prescriptions, Stripe, S3, audit logging — 8 weeks. A UK digital pharmacy with regulatory compliance, prescription verification flow, Stripe — 10 weeks. Both anonymized under NDA; references on a call.
How do you handle the model API itself under HIPAA?
Two paths: Anthropic Claude via AWS Bedrock under the AWS BAA, or OpenAI via Azure OpenAI under the Azure BAA — both on no-training enterprise tiers. Direct API calls without an enterprise BAA are never used for PHI workloads. Application-layer PHI redaction runs on every request regardless of provider, so the model sees the minimum necessary even inside a BAA-covered deployment.
More on shipping AI software end-to-end
Building HIPAA-aware AI?
Free architecture audit in 24 hours. We map your PHI flows and tell you what ships in 8 weeks.
Top Rated Plus Upwork · 100% JSS · 42 projects · $200K+ earned · 100% satisfaction guarantee