AI for PHI Analysis: What Zero Data Retention Means

Meta description: "Zero data retention" is not standard across AI vendors. Here's what it means in practice, why it matters for PHI, and how to verify a vendor's claim.

TL;DR

• Zero data retention (ZDR) means an AI provider processes your PHI without writing it to persistent storage after returning a response, but this claim varies widely across vendors and often includes carve-outs for abuse monitoring, safety logs, and subprocessor chains.
• ZDR is one required control under HIPAA's Security Rule, not a substitute for encryption, access controls, audit logging, and minimum necessary standards.
• Verify vendor ZDR claims by requesting explicit BAA language, subprocessor lists, and log retention schedules, not just marketing copy.

What does zero data retention actually mean for PHI?

Zero data retention for AI means a provider processes your protected health information without writing it to persistent storage after generating a response. The prompt and output are discarded once the inference call completes. That is the marketing definition. The operational reality is messier.

If you are building a broader compliance program around AI tools, the HIPAA-Compliant AI for Healthcare Providers: The Complete 2026 Guide covers the full landscape of safeguards, vendor evaluation, and implementation requirements. This post focuses specifically on what ZDR means, what it excludes, and how to verify a vendor's claim before processing PHI.

OpenAI generates abuse monitoring logs for all API feature usage and retains them for up to 30 days, even under a ZDR arrangement. Anthropic's flagged sessions can be held for up to two years regardless of your arrangement, including under ZDR and including under a HIPAA agreement. These are carve-outs most healthcare teams do not read carefully.

The distinction matters because HIPAA does not recognize "zero retention with exceptions." Either PHI is retained or it is not. When a vendor claims ZDR but retains metadata, abuse-monitoring logs, or safety flags, your organization remains liable for that retained data under 45 CFR §164.312(b), which requires audit controls over every record.

How PHI moves through an AI system

When you send a clinical note to an AI tool, that data moves through four stages, each with its own retention risk.

Stage 1: Inference Input. Your prompt arrives at the API endpoint. The vendor can log the request, cache it for performance, or flag it for abuse monitoring. ZDR means the content is not stored after the response is delivered, but that applies only to the inference layer.

Stage 2: Model Processing. The model generates a response in memory. No persistent storage occurs here. This is where ZDR claims are technically accurate.

Stage 3: Operational Logging. After the response is returned, the vendor logs metadata: timestamps, query length, user ID, error codes. This metadata is not the clinical content, but it can re-identify patients. HIPAA's 18 identifiers include dates and any unique identifying number or code. A vendor that claims "we don't store the content, only metadata" is not HIPAA-compliant.

Stage 4: Subprocessor Chains. Your vendor's zero-retention policy does not automatically bind their cloud infrastructure provider, their LLM provider, or their logging service. HHS sample BAA provisions require that subcontractors be bound by the same PHI protections. Gaps in subprocessor coverage are a common audit finding.

ZDR controls risk at stages one and two. Stages three and four require explicit contractual language and verification.

What zero data retention does NOT cover

Abuse Monitoring Carve-Outs. If you process 1,000 patient summaries per day and the vendor retains abuse-monitoring logs for 30 days, that is 30,000 PHI-adjacent records stored, even under a "zero retention" contract.

Model Training vs. Operational Retention. Training opt-out means the provider will not use your inputs to improve its models. ZDR goes further: the provider does not persist your data at all after returning a response. You can have one without the other. Many healthcare teams confuse these two distinct commitments.

Safety and Compliance Holds. If a vendor's safety team flags a session for potential abuse or legal compliance, the data is retained. This carve-out is rarely addressed explicitly in BAAs.

Metadata and Audit Logs. Timestamps, user IDs, and query patterns can re-identify patients under the 18-identifier standard, even when payload content is not stored.

How to verify a vendor's zero data retention claims

Marketing claims are not contracts. Use this checklist before signing a BAA.

Step 1: Request the BAA and subprocessor list. The BAA should explicitly state that ZDR applies to all PHI, name the retention period for any exceptions, and list all subprocessors who handle PHI. If the vendor cannot name their subprocessors, they cannot guarantee ZDR.

Step 2: Confirm subprocessor BAAs. For each subprocessor, confirm a BAA is in place that covers the specific services used. If your vendor uses AWS Bedrock, note that Amazon Bedrock supports model invocation logging that can write full request and response data to CloudWatch Logs or S3. Disabled by default at the Bedrock level is not the same as disabled by contract.

Step 3: Request log retention schedules. Ask: "What logs are retained after processing? For how long? Where are they stored? Who has access?" If the vendor cannot answer, they have a marketing claim, not a ZDR policy.

Step 4: Verify encryption in transit and at rest. ZDR does not substitute for encryption. Confirm TLS 1.2 or higher for all API calls containing PHI.

Step 5: Request audit evidence. Ask for a third-party audit report documenting data deletion procedures. If the vendor cannot provide it, the claim is contractual only.

Red flag language to reject

• "We do not retain data for training purposes" (does not address operational logs)
• "Data is deleted after 24 hours" (not zero retention)
• "Subprocessors are listed upon request" (should be provided proactively)

Does zero data retention make an AI tool HIPAA-compliant?

No. This is the most dangerous assumption in healthcare AI procurement.

NIST SP 800-53 and HIPAA's Security Rule require covered entities and their business associates to implement technical safeguards under 45 CFR §164.312, including encryption, access controls, and audit logging. ZDR is one safeguard among many. It does not address who can access the tool, whether PHI is encrypted in transit, whether audit trails exist, or whether your team is applying minimum necessary standards before submitting data.

A tool is HIPAA-eligible if the vendor will sign a BAA and has implemented basic safeguards. A tool is HIPAA-compliant only if your organization has configured it correctly and integrated it into a broader compliance program. When HHS OCR investigates a breach, they examine whether all required safeguards were implemented, not just one.

Comparing zero retention policies across major AI vendors

The table below reflects publicly available terms as of mid-2026. Confirm current terms directly with vendor legal teams before processing PHI.

OpenAI requires enterprise agreement activation and retains abuse monitoring logs for up to 30 days even under ZDR. Microsoft Azure AI requires an approved application to disable abuse monitoring, not just a contract. Anthropic's flagged sessions can be held for up to two years regardless of your arrangement, a significant carve-out for healthcare use cases. AWS requires explicit opt-out of CloudWatch logging at the configuration level.

Hathr.AI provides zero data retention by default with no carve-outs, built on FedRAMP High Infrastructure and following Government-approved Claude, NIST 800-53 controls, FedRAMP High Infrastructure. BAA is provided on signup, subprocessor disclosure is upfront, and no abuse-monitoring or safety-review carve-outs apply to customer PHI. For teams processing high volumes of clinical documents through Hathr.AI for healthcare, this architecture eliminates the vendor verification overhead that other providers require significant configuration or enterprise negotiation to approximate.

Frequently asked questions

Does zero data retention mean the vendor cannot use my PHI to train their models?

No. Training opt-out and ZDR are separate controls. A vendor can offer training opt-out while retaining operational logs indefinitely. Request both commitments explicitly in your BAA.

How is ZDR different from training opt-out?

Training opt-out prevents your data from improving the vendor's models. ZDR prevents your data from being stored at all after processing. If you need both, request both in writing.

What specific language should a BAA include to enforce ZDR?

The BAA should state: the vendor will not retain prompts or outputs after processing; the retention schedule for any exceptions; the vendor will delete all PHI within a specified period upon request; and all subprocessors are bound by the same terms. Per HHS sample BAA provisions, subcontractor binding is a required element.

Is ZDR sufficient as a sole HIPAA safeguard?

No. ZDR addresses data persistence but not access control, encryption, audit logging, or minimum necessary standards. It is one layer of a multi-layered compliance program.

Next steps

Request a BAA from any vendor before processing PHI. Verify subprocessor terms, request log retention schedules, and ask for audit evidence that ZDR is enforced at the infrastructure level, not just contractually. Document your compliance program, including BAAs, subprocessor agreements, and data minimization policies.

Team of 10 or more? Book a 20-minute Enterprise Interview to walk through security review, BAA, and rollout.

Try Hathr.AI free. HIPAA-Compliant, built on FedRAMP High Infrastructure. BAA provided on signup. Zero data retention by default. No usage limits, no seat minimums. Start a free trial ->