Is custom AI safe for HIPAA, finance, and other regulated workflows?
Yes, with care. The major AI providers (OpenAI, Anthropic, Google, AWS Bedrock) offer Business Associate Agreements (BAAs) on their enterprise/API tiers, making HIPAA-compliant custom AI possible since 2023–2024. The risk is not the model, it’s the integration. Most HIPAA leaks in custom AI projects come from logging, prompt-history features, or third-party tools that weren’t covered by the BAA.
For a regulated small business, a dental office, a small clinic, a financial advisor, a law firm with sensitive cases, the question of whether custom AI is "safe" usually conflates two things: is the model itself compliant, and is the build done in a way that respects the regulation? The first is a yes-or-no answered by the BAA. The second is engineering discipline.
On the model side, OpenAI signs BAAs on their Enterprise and dedicated API tiers. Anthropic signs BAAs on Claude for enterprise and the API at certain commitment levels. Google offers HIPAA coverage on Vertex AI. AWS Bedrock covers most regulated industries via the AWS BAA. None of these cover the consumer-tier products (ChatGPT free/Plus, Claude.ai free); a regulated business should not use those for protected data.
On the build side, the failures are mundane: developer logs that capture full prompts (including PHI), a third-party analytics tool (Sentry, Logflare) that wasn’t covered by a BAA, a prompt-cache that stored a customer’s record longer than retention policy allows. The mitigation is the same as any HIPAA-conscious software build: data minimization, encrypted-at-rest storage, audit logs, vendor BAA inventory. Custom AI doesn’t change the discipline, it just adds two or three more vendors to the BAA list.
Key facts
- OpenAI: BAA available on Enterprise plans and zero-data-retention API tiers.
- Anthropic: BAA available for Claude API at certain enterprise commitment levels.
- Google: HIPAA coverage on Vertex AI under standard GCP BAA.
- AWS Bedrock: HIPAA covered under standard AWS BAA, many enterprise builds default here for this reason.
Common follow-ups
Can a small dental office actually get a BAA from OpenAI?
Through the standard API: yes, on the zero-data-retention tier. Through ChatGPT Team: yes, with a BAA. Through ChatGPT Plus or free: no, never use those for PHI.
What about state laws: California CCPA, New York SHIELD?
These add disclosure and consent requirements but don’t generally prohibit AI per se. The same engineering discipline (data minimization, encryption, audit logs) covers both.
Sources
Related answers
- What data do I need to give an AI agent for it to be useful? →
- What is AI realistically useful for in a small law firm in 2026? →
- What are the best AI agents for small businesses in 2026 (and when to build your own)? →
- What does "custom AI" mean for a small business? →
Want a website built to be cited by Google and AI answer engines? Drop your URL, if it’s a fit, we’ll rebuild it for free.
See if you’re a fit →