UrgisoftAI-built apps

AI-built apps ยท November 2026

Is your Lovable app HIPAA compliant? What an AI-built health app needs before real patient data

Short answer: not on Lovable itself. Lovable's terms (read November 21, 2026) say you "agree not to upload, input, or otherwise provide through the Services any protected health information subject to HIPAA". So if your app will hold patient data for a clinic, a health plan or someone working for one, build and test it in Lovable with made-up data, and run the live app and its data on services that sign a Business Associate Agreement (BAA) with you. Below: how to tell whether HIPAA applies at all, which pieces need a BAA, and seven ways AI-built health apps leak data even on the right services.

This page is a technical checklist, not legal advice. Whether HIPAA applies to your app, and what your contracts must say, is a question for a healthcare lawyer. The same goes for apps built with Bolt, Base44, Replit or Cursor: check each builder's own terms before any real patient data goes in.

First: does HIPAA apply to your app?

Your app is...HIPAA?What to read
Used by a clinic, practice, hospital or health plan for its own patients or members, or built and run for one of themYes, usually. The clinic is a "covered entity"; you, and every service that stores or sees the data for you, are "business associates" and need a BAA.HHS's health apps page and its 2016 "Health App Use Scenarios".
A consumer app people choose and use on their own (a symptom diary, a fitness or cycle tracker), not offered by a clinic or planUsually not HIPAA, per HHS's scenarios. But the FTC's Health Breach Notification Rule can apply instead.The FTC's guide to that rule.

The FTC rule is easy to miss: it counts "sharing of covered information without an individual's authorization" as a breach, and its own example is health data sent to an ad network. A breach means telling users within 60 calendar days, and the FTC lists civil penalties of up to $53,088 per violation. So the ad pixels and analytics an AI builder adds without asking matter even when HIPAA does not apply.

Which pieces need a BAA (if HIPAA applies)

PieceWhat its own documentation saysWhat to do
The builder (Lovable, including Lovable Cloud)The terms forbid putting HIPAA protected health information through the service.Build and test with made-up data only. Never paste a real patient's details into the builder's chat or its database.
Database, login and files (Supabase)"You will need to be at least on the Team Plan to sign a BAA"; turn on the HIPAA add-on, mark the project High Compliance, enable Point in Time Recovery, SSL enforcement and network restrictions, keep Postgres connection logging on, enforce MFA for everyone in the organization, and do not store protected health information in public Storage buckets.Connect the app to your own Supabase project in an organization with a signed BAA, and clear every warning the Security Advisor shows.
Hosting (for example Vercel)Signs BAAs "with eligible Pro and Enterprise customers"; Pro teams buy the HIPAA BAA add-on in Billing settings.Move the live app off the builder's hosting to a host that has signed a BAA with you (moving a Lovable or Bolt app to Vercel, step by step).
AI features (for example OpenAI's API)API organizations can enable a BAA themselves in Settings once they have a history of API use; OpenAI says it does not offer a BAA for ChatGPT Business. No patient data in a prompt to any AI service until that service has signed a BAA with you.
Email, text messages, error tracking, analyticsEach vendor decides; many sign no BAA.Either a BAA, or that tool never receives health details (send "you have a new message, log in to read it", not the message).

A BAA alone does not make an app compliant. HIPAA also asks for risk assessments, policies, access controls, audit logs and breach procedures; the services above only cover their part.

Seven ways AI-built health apps leak data, and the fix for each

  1. Tables anyone can read. Row Level Security is off, or a policy says true, so anyone who copies the public key from the browser can download every patient row. Turn RLS on for every table and limit each policy to auth.uid() (or the clinic the user belongs to). Then test it: log in as patient A and ask the database for patient B's rows. How to run that test: six security checks you can run yourself.
  2. Uploaded files in a public bucket. Lab results, insurance cards and photos uploaded to a public Storage bucket can be opened by anyone with the link, and Supabase's HIPAA guidance names this one directly. Make the bucket private, give it policies like the tables, and hand out short-lived signed links.
  3. Health details in logs. AI-written code often logs whole requests and records while debugging: console.log(patient), a function that prints the request body, an error tracker that captures form fields. Search the code for console.log and print(, remove the ones that print records, and turn off form-field capture in the error tracker.
  4. Ad pixels and analytics on health pages. A tracking script on a page that shows a condition, an appointment type or a medication sends that page's address and contents to a third party. Remove trackers from signed-in pages, and check the browser's Network tab on those pages for requests to ad and analytics domains.
  5. Patient data sent to an AI model. A "summarize this visit" or "chat with your records" feature puts health details into a prompt. Without a BAA with that AI provider, turn the feature off for real data or strip names and identifiers before the call.
  6. Health details in emails and texts. Reminder and notification code written by the builder usually puts the details in the message. Send a notice with a link to log in instead, or use a provider that signs a BAA.
  7. Real records in the builder. Test data copied from real patients, a spreadsheet of members imported to "try it out", screenshots of a real chart in the chat. That is protected health information going through a service that forbids it. Use made-up data in the builder; real data only ever lives in the BAA-covered project.

Check it before a single real patient signs up

(1) Log in as patient A and try to read patient B's records, both through the app and with the public key straight against the database. (2) Open an uploaded file's link in a private window while logged out. (3) Search the code and the live logs for a test patient's name. (4) Open the signed-in pages with the Network tab open and list every outside domain the page talks to. (5) Write down each service the app uses (builder, database, host, AI, email, texts, error tracking, analytics) and next to it either "BAA signed on (date)" or "never sees health data". If a line has neither, the app is not ready for real patient data.

This is a checklist, not a study, and not legal advice. Read on November 21, 2026: Lovable's terms (the line on protected health information), Supabase's HIPAA page, HIPAA projects guide and shared responsibility model (Team plan, project settings, MFA, public buckets), Vercel's compliance page (BAAs for Pro and Enterprise), OpenAI's BAA help article, the FTC's Health Breach Notification Rule guide and HHS's health apps guidance. The seven leaks are general engineering practice. Vendors change their plans and terms often, so read their pages again before you sign anything. Urgisoft is not affiliated with Lovable, Supabase, Vercel, OpenAI or any other company named here.

Other problems

Six security checks you can run yourself. Login and Supabase errors: seven causes. Deploy to Vercel or Netlify. A blank page after publishing: six causes. Stripe payments not working: seven causes. Base44 bug fixing. Replit app bug fixing. Fix AI-generated code: seven common problems. Convert a Lovable app to iOS. Lovable to WordPress. Lovable app not showing in Google? Seven fixes. Slow on phones? Our demo shop, measured before and after: 14.5 s to 1.4 s. Not sure which problem you have? A ten-minute check.

Want the leaks found and fixed?

We fix apps made with Lovable, Bolt, Base44, Replit, Cursor and other AI builders: tables anyone can read, public file buckets, records in the logs, trackers on signed-in pages, and the move to your own database and host. We tell you in plain words what was wrong and how we tested it. Fixed prices on Fiverr: one bug from $40, up to five bugs with the app working live from $150. No fix, no charge. We never ask for your password. We do not certify HIPAA compliance or sign BAAs for your vendors; that stays with your lawyer and each service.

See the gig on Fiverr