HIPAA compliant AI agents
HIPAA compliant AI agents, by mechanism, not by promise
HIPAA compliant AI agents need proof, not promises. Ambient scribes, chart summaries and billing agents are new PHI boundaries. The HIPAA Security Rule NPRM pulls each one into your risk analysis. DataShield hands agents tokens, not identifiers. We tokenize datasets at ingest. Agents query over MCP. Turning a token back is a walled-off step, and we log it.
A health breach costs $7.42M and 279 days to close, on average (IBM 2025). The control is not a promise in a BAA. It is proof you can check.
PHI de-identification for LLM agents: three failure points, closed
Agents see PHI they don't need
The minimum-necessary standard applies per agent, per handoff. DataShield sorts fields at ingest: NAME, DOB, MRN and the rest. It swaps each identifier for a deterministic, join-preserving token. Agents get answers, not identifiers. We roll up quasi-identifiers too, so mixing fields can't pin a row back to a person. See Ontology.
No proof of who accessed what
HIPAA §164.312(b) calls for audit controls over ePHI use. We scope, meter, and tie each governed tool call to an agent. We seal each one into a SHA-256 hash chain with Ed25519-signed checkpoints. A check tells you if someone changed, added to, cut, or trimmed a log's tail. Try it on a sample log.
Re-identification is all-or-nothing
One-way redaction destroys clinical value. The vault can turn our tokens back, under policy. A billing step that truly needs the identifier gets it, walled off and logged. Erasure means crypto-shred, and it leaves the audit chain whole. Details in Security.
The HIPAA Security Rule NPRM puts AI tools in your risk analysis
The proposed Security Rule update asks you to list AI tools that touch ePHI. Put them in your risk analysis. What kind of ePHI, how much of it, shown to whom, and where the output goes. OCR's risk-analysis enforcement drive runs into 2026. Fines reach $73,011 per violation, with a $2.19M yearly cap. The first health GenAI case has already landed: Texas AG v. Pieces Technologies, over hospitals feeding live patient data to a GenAI summarizer.
No law bans PHI near an LLM outright. What the rules ask for is risk analysis, minimum-necessary access, BAAs and an audit trail. A data plane that de-identifies at ingest gives you that. So does a hash-chained audit trail.
"Our LLM vendor has a BAA" is not a control. A BAA parcels out blame for one vendor node. It skips pipeline entry, app logging, and agents chaining tools. Those are the places the 2025 MCP incidents actually happened. DataShield runs self-hosted, in your own cloud or data center. Or on a dedicated single-tenant server we run. Either way, you hold the keys. A BAA conversation is welcome. Read the architecture.
HIPAA compliant AI agents: buyer questions
Does DataShield make our ambient scribe or chatbot HIPAA compliant?
No product makes you compliant on its own. Compliance is your program. DataShield brings the technical controls that program needs. We tokenize PHI at ingest and roll up quasi-identifiers (Ontology). We authorize each tool call on its own, and you can revoke mid-session (Auth). Audit logs stay tamper-evident. They map to §164.312(b) by design.
Does DataShield intercept prompts to strip PHI in real time?
No. We make tokens at ingest, per dataset. There is no inline traffic interception. Agents query already-tokenized data over MCP, so identifiers never enter the query path. Turning a token back is a separate, walled-off vault step. We log it.
Can we still join patient records after de-identification?
Yes. Tokens are deterministic and join-preserving. The same MRN always maps to the same token. Joins across datasets still work. The identifier itself stays out of the query path. We roll up quasi-identifiers to a set k, so mixing fields can't match rows back.
How do we handle a patient erasure request if data is tokenized?
Erasure means crypto-shred. Destroy the key material and the tokens go dark for good. The audit chain stays whole, and you can still check it. You keep the proof trail without keeping the identifiers.
Is a self-hosted AI agent platform HIPAA compliant?
HIPAA has no product certification, so no platform is "HIPAA certified". What you can get is alignment. Self-hosted means PHI stays in your own estate, on your keys. It never leaves for a vendor tenant. These controls line up with the Security Rule. Tokens at ingest. Per-call authorization. Audit logs that map to HIPAA §164.312(b) by design. A BAA is available on request.
PHI tokenized at ingest. Queried over MCP. Turned back only under audit, in your own estate, with your keys.
Get an instant quoteYou've seen the proof
Ready for a number? Scope your deployment and we'll price it against your own economics.
Get your quote →