PHI does not leave. The assistant knows what it may not say.
Two questions decide whether AI goes live in a health setting: where the patient data goes, and what happens when the model is confidently wrong.
Building AI for hospitals and payers rather than running one? Yours is sell into the enterprise.
What your compliance office asks, and what answers it
| The question | The answer here |
|---|---|
| Where does PHI go? | It stays in your environment; Swiftward never receives it. Where you route a prompt to a hosted model, redaction runs before it leaves. |
| Will you sign a BAA? | Yes. |
| Minimum necessary: does the model see more than it needs? | A rule strips fields before the prompt leaves and restores them in the reply. The model sees placeholders. |
| Which de-identification method? | Safe Harbor is a list of identifiers, so it is a rule you write and read rather than a model you trust. Most of those identifiers already have a detector; the rest are yours to add. Expert Determination is a judgment your expert makes, and we hold the evidence they base it on. |
| If something did get out, would you know? | The record shows what was redacted and what was not, per decision. That is what a breach assessment needs. |
The re-identification nobody writes a rule for
Safe Harbor strips eighteen identifiers, and a rare diagnosis is not one of them. The year of birth may stay, and so may the first three digits of a ZIP code. Put those three in one prompt and the patient can be identified, and a scanner reading one field at a time passes it.
A rule can act on the combination, because it sees the whole request rather than one field at a time.
The redaction layer also runs a named-entity check that recognizes medical terms, names and places a regular expression cannot describe. What the redaction layer recognizes.
When the model is confidently wrong
A clinical assistant that answers with certainty and no source is more dangerous than one that says nothing.
A rule can require a citation — and, where you supply a lookup function, check that it is found in the source you point it at. It can check the answer against the role the assistant was given, and send anything that fails to a clinician rather than to the patient.
How that check works: role and judge checks. Where the handover goes: human review.
Fitness, coaching and mental health count too
They hold health data without being clinical, and their users assume the same care. The controls are identical; the regulator may not be.
Where your regulator goes further than HIPAA
42 CFR Part 2. Substance use disorder records carry stricter handling than the rest of PHI. Where those records are fields you store, classify them once and it governs every screen, export and forwarded event. Where they arrive inside a prompt or a tool response, they are a separate detection class with its own redaction rules, kept apart from the rest of PHI. Both are declarations you write; neither is inferred.
The clinical boundary. Wherever you and your regulatory counsel have drawn the line between support and clinical decision support, a rule holds it. The assistant does not produce a diagnosis, a dose, or a recommendation you have not authorized it to produce. Drawing the line is your call; holding it on every request is what this layer is for.