Q: What is Zero Data Retention (ZDR) and how does it relate to HIPAA? A: True ZDR means the provider processes your data in real time and immediately discards it—no prompt logging, no retained outputs, and no use for training. ZDR is a contractual promise distinct from HIPAA (you can have one without the other), and this HIPAA-compliant AI coding tools guide explains why ZDR is not automatic and when providers may still retain data for safety monitoring or flagged sessions. Q: How does Anthropic’s Claude Code handle data retention and HIPAA coverage? A: Anthropic provides ZDR via direct commercial contracts on the Anthropic API, but bundled Claude Code seats are not covered by the HIPAA-ready offering and require contacting Anthropic’s account team. Files uploaded via the Files API are retained until explicitly deleted and metrics logging may be exempt from ZDR, and Claude used through AWS Bedrock or Google Vertex follows those cloud BAAs instead of Anthropic ZDR. Q: What should healthcare teams know about OpenAI Codex and ChatGPT for HIPAA use? A: OpenAI offers ZDR to approved enterprise customers (for example, ChatGPT Enterprise) but standard API endpoints can retain inputs and outputs for up to 30 days for abuse monitoring and ZDR must be specifically requested and approved. Workloads that run in OpenAI-hosted containers are not ZDR-eligible, and consumer ChatGPT tiers and personal accounts are not HIPAA-eligible, creating high risk for PHI. Q: How does Google Vertex AI protect customer data and support HIPAA? A: Google Vertex AI builds privacy into the platform with an AI/ML Privacy Commitment that prevents using customer data to train managed models without permission, and Google’s BAA covers Vertex for HIPAA workloads. Published Gemini models may cache customer data in-memory with a default 24-hour TTL to reduce latency, and that caching can be disabled at the project level. Q: Which agent workflows commonly break ZDR and should be avoided for PHI? A: Agent features can silently break ZDR—code execution sandboxes often retain container data for up to 30 days, batch APIs typically hold data for 29 days, and beta features are excluded from ZDR; file uploads, embedding stores, and function call logs can also persist data unless you enforce deletion. The architectural implication is to route PHI-adjacent inference through real-time ZDR-eligible Messages APIs and avoid batch jobs, sandboxes, and beta features for regulated data. Q: What engineering controls should teams implement to stop PHI leaks when using AI coding tools? A: Implement technical controls that don’t depend on memory, such as deny rules, code-review gates, network isolation, blocking consumer AI domains, separate VPCs, and egress controls, and route PHI only through ZDR-eligible real-time inference endpoints. Also mask or tokenize identifiers, strip attachments, disable prompt/event logging and in-memory caches where possible, and enforce pre-commit hooks and code reviews to prevent developers pasting PHI into prompts. Q: How can organizations stand up AI governance quickly? A: Start with an inventory of every AI tool and map each to its contract, tier, and BAA status, then publish a simple one-page policy forbidding consumer ChatGPT for work and disallowing PHI in non-approved tools. Set default-deny for file uploads with auto-delete timers, create two model routes—regulated (ZDR or cloud BAA, no logs, no cache) and unregulated—and monitor vendor terms and settings regularly. Q: What immediate steps prevent developers from accidentally sharing PHI in AI prompts or personal accounts? A: Audit and block consumer AI domains at the proxy while allowing only approved enterprise endpoints, and enforce separate projects or VPCs with strict egress controls so PHI workloads are isolated. Add pre-commit hooks, code-review gates, outbound scans for PHI patterns, and clear policies prohibiting personal ChatGPT accounts for work to make the safe path the easy path.