What is HIPAA?

Why It Matters

HIPAA was written for systems that store and query records. An agent does something those systems do not: it reads health information into a context window, reasons over it, and emits it inside an artifact. The obligation does not change, but every control that satisfied it before now has a gap.

The practical consequence: compliance for AI is decided during architecture, and an audit only discovers whether that happened. Access scope, retention, logging, and the vendor agreement all have to be settled before the first record reaches a model.

The Three Safeguards

Administrative. Policies, workforce training, and the vendor relationships that govern who may touch health data. The business associate agreement lives here.

Physical. Control of the facilities and devices where the data sits. In a cloud deployment this becomes a question about which region and which tenancy.

Technical. Access control, audit controls, integrity, and transmission security. This is where AI systems most often fall short, because the controls assume a database query rather than a model reasoning over retrieved text.

Where It Breaks

Role-based access answers “may this user see this record.” It does not answer “may this agent pull this field into a prompt for this purpose.” Those are different questions, and the second one is where over-retrieval happens.

Encryption protects data at rest and in transit. It says nothing about protected health information flowing through a model’s reasoning in memory, or landing in a trace that ships to an observability vendor.

Audit logs capture queries and writes. They rarely capture the sequence of model inputs and outputs that produced an artifact. An access log that shows nothing unusual is not evidence that a run was appropriate.

How Flytebit Handles It

We treat the safeguards as design inputs. The deployment runs under a business associate agreement, retrieval is bounded by role and purpose, and every run produces a decision record that holds the sources, the model and prompt versions, and the reviewer’s action. Residency rules apply to the retrieval index, the telemetry sink, and the review queue, not only to the database. The industry application is on our Healthcare & Life Sciences page, and the engagement version is our AI governance and risk work.