Picture a skilled nursing admissions coordinator who is tired of copying referral updates between spreadsheets. They describe a dashboard to an AI coding assistant: show open referrals, flag missing documents, and draft a follow-up. By lunch, the dashboard looks real. By the end of the day, somebody suggests uploading yesterday’s referrals to test it.
That last step is where a useful experiment can become a privacy problem. The issue is not that AI wrote the code. It is that nobody has yet established where the patient information goes, who can see it, or which services are permitted to process it.
AI-assisted software can be part of a HIPAA-compliant operation. “Vibe coded” is neither a certification nor a disqualification. Compliance depends on the organization, the use case, the complete data flow, the required agreements, the safeguards, and how the system is operated.
This is an educational design and development guide, not legal advice, a compliance assessment, or a certification. Your privacy, security, and legal teams should determine what applies to your organization.
1. A working prototype is not a compliance assessment
Vibe coding means describing a feature in ordinary language and using AI to generate or revise the code. It is useful for exploring layouts, trying workflows, building test fixtures, and turning an idea into something a team can discuss. It can also generate code that appears convincing without enforcing the security decisions the application needs.
A login screen does not prove that one facility cannot read another facility’s records. A green “secure” badge does not prove that uploads are private. An encrypted database does not stop a debugging tool from recording an unredacted referral.
Treat generated code as a draft that needs review. Use AI to accelerate implementation, not to sign off on its own work. For patient-facing or clinical AI, also assess accuracy, human oversight, and patient safety; HIPAA compliance alone does not establish clinical suitability.
2. Draw the data boundary before writing another prompt
HIPAA applies to covered entities and business associates in the circumstances the rules describe. A healthcare-themed website is not automatically a HIPAA-regulated system, and a contact form is not automatically outside the rules either. Assess the information collected and the organization’s role, rather than guessing from the page’s appearance.
Start by listing what the feature creates, receives, maintains, or transmits. Include names, contact information, appointment requests, diagnoses, referral documents, notes, identifiers, and any combinations that identify a person in connection with their care.
Then trace every destination:
- The browser, application server, database, file storage, and backups.
- The AI builder, coding assistant, model endpoint, and connected tools.
- Email, text messages, analytics, session recordings, error monitoring, and support tickets.
- Preview deployments, developer laptops, downloaded spreadsheets, and exported logs.
Do not limit the inventory to the main database. A patient’s name in a URL can end up in a log. A screen recording can capture a referral. An error report can include a form submission. In healthcare AI, the seemingly secondary systems are often where the boundary breaks.
Prototype with entirely synthetic patients. Create fictional records from scratch. Removing names from real records is not, by itself, HIPAA de-identification. HHS describes two methods, Safe Harbor and Expert Determination, with specific requirements. Read the HHS de-identification guidance.
3. A BAA is necessary where required, but it is not the whole answer
A business associate agreement, or BAA, addresses a business associate’s responsibilities for protected health information. HHS explains that a cloud provider that maintains electronic PHI can be a business associate even when the information is encrypted and the provider cannot decrypt it. Read the HHS cloud computing guidance.
“The vendor offers a BAA” is the beginning of the conversation. Confirm the actual agreement is executed and covers the exact service, account, product tier, model endpoint, and feature you intend to use. Check exclusions, retention, training uses, support access, subprocessors, and connected services.
For example, a covered database service does not automatically cover a separate AI chat product from the same company. A covered production environment does not automatically cover a public preview. A signed BAA does not repair a public storage bucket or missing authorization check.
The tool that generates the code, the code it generates, and the services that run that code are different parts of the system. Approval of one does not establish approval of the others.
4. A safer way to vibe code a healthcare application
Step one: make the scope deliberately small
Write down the users, the task, the minimum information needed, and what the application must not do. A public service-page assistant that answers approved questions is a different project from an assistant that reads medical records. Do not quietly expand the first into the second.
Step two: build against synthetic data
Keep the prototype isolated from real patient systems. Do not paste referrals, screenshots, production logs, credentials, or medical records into an unapproved coding assistant to fix an error. Reproduce the problem using invented records and redacted technical details.
A useful starting prompt is explicit about the boundary:
Build a referral workflow prototype using fictional records only.
Do not connect to production systems or external AI services.
Keep credentials out of browser code.
Define facility-scoped access rules before implementing queries.
List every proposed external dependency for human approval.
Generate tests proving users cannot access another facility's records.
This prompt is a development instruction, not a security control. A qualified reviewer must verify that the generated application actually follows it.
Step three: review the architecture and the code
As practical engineering safeguards, check server-enforced authorization on every read and write, private file access, secrets management, session handling, input validation, and dependency risks. Test what users are denied, not just what administrators are allowed. Use appropriate encryption and govern keys and backups; do not assume a vendor’s default settings are suitable for your workflow.
Step four: inspect the AI boundary separately
If the production feature sends information to a model, approve that path explicitly. Minimize the data sent, verify relevant contractual terms and retention settings, and prevent patient content from leaking into diagnostic logs. Patient documents can contain malicious instructions, so treat their text as data, not as permission to change access rules or call tools.
For actions such as sending a message or updating a record, define narrow permissions and meaningful human approval. A fluent answer is not evidence that an action is appropriate.
Step five: document operational ownership
HHS’s cloud guidance emphasizes risk analysis and risk management, alongside the other applicable HIPAA obligations. Identify who reviews access, investigates incidents, manages vendor changes, restores backups, trains staff, and approves releases. Document decisions and unresolved risks. A system is not finished just because the generated code compiles.
5. What this looks like for a skilled nursing admissions tool
Return to our admissions dashboard. The safer first version uses fictional names, mock referral PDFs, and invented facility IDs. It can demonstrate the workflow without touching actual patient information.
Before a production pilot, the team maps the referral journey, approves the hosting and storage services, resolves required BAAs, and restricts each user to the facilities they support. File links require authorization. Notifications say an update is available rather than copying a diagnosis into ordinary email. Monitoring avoids capturing patient documents.
The team then tests an adversarial case: a user at Facility A changes a request to the identifier of a Facility B referral. The server must reject access even if the interface would normally hide that referral. The team also tests exported files, old links, removed users, backups, and support access.
If an AI summary is added later, that is a new data flow to evaluate, not a cosmetic upgrade. The coordinator should be able to inspect the original document, verify the summary, and correct it. The application must not turn an unverified AI summary into an automatic admissions or clinical decision.
6. The questions to answer before real data enters
- Scope: Have the organization’s HIPAA roles and the feature’s data types been assessed?
- Vendors: Are all services that touch PHI identified, with required agreements and service boundaries confirmed?
- Access: Have denial tests verified separation between users, roles, and facilities?
- Visibility: Can PHI appear in prompts, logs, analytics, previews, email, URLs, or support tools?
- Resilience: Have backup restoration, access removal, incident response, and relevant retention needs been addressed?
- Ownership: Who accepts remaining risks and approves production use?
If any answer is unknown, keep the application in a synthetic-data environment while the team resolves it. This list organizes the discussion; it is not an exhaustive compliance checklist or a substitute for a documented assessment.
7. Move quickly on the prototype. Be deliberate about the patient data.
The strongest use of healthcare AI is not skipping engineering, privacy, or security. It is reducing the time spent on repetitive implementation so the team can spend more attention on the decisions that matter.
For a healthcare website, separate public marketing content from patient-data workflows. Keep sensitive submissions out of general-purpose marketing forms, and direct them to the organization’s approved secure channels. For an application, design the data boundary before connecting live systems.
SNF Web Design works on medical website design, healthcare web development, and accessible websites. Talk with us about the website experience and the boundaries your privacy and security teams need respected. No website design, platform, or AI tool can independently guarantee your organization’s compliance.
Discuss Your Healthcare WebsitePrimary sources
- HHS: Guidance on HIPAA & Cloud Computing
- HHS: Methods for De-identification of Protected Health Information
Published November 20, 2026. Vendor terms and regulatory requirements can change. Verify current requirements for your exact use case before processing patient information.
