Patient privacy does not stop at the practice door

Practices tend to think of privacy as a matter of what is said, and to whom. AI moves the question. The risk is no longer only in what someone tells a caller. It is in what gets copied, sent and stored somewhere else.

What makes an AI tool HIPAA compliant

HIPAA compliant AI is not a label a vendor prints on a website. It means the practice knows what patient information leaves the office, which vendor receives or stores it, under what agreement, who can read it and how long it is kept. A tool that cannot answer those questions in writing should not see a chart.

The vendor is part of the chain

HIPAA does not ban AI, and it does not ban the cloud. It asks who handles protected health information on the practice’s behalf, and on what terms. The Office for Civil Rights at HHS put it plainly in its guidance on cloud services:

“When a covered entity engages the services of a CSP to create, receive, maintain, or transmit ePHI (such as to process and/or store ePHI), on its behalf, the CSP is a business associate under HIPAA.” — HHS Office for Civil Rights, Guidance on HIPAA and Cloud Computing

The same guidance adds that this holds even when the provider only stores encrypted data and has no key, and that using such a service requires a business associate agreement.

An AI service that receives a visit note, a referral letter or a patient message is in the same position. If it processes or keeps that text for the practice, the question is not whether the tool is clever. It is whether there is an agreement, and what it covers.

Physicians already use it

This is not a future problem. According to an American Medical Association survey, 66% of physicians reported using health care AI in 2024, up from 38% the year before. The same survey found that physicians list data privacy assurances among the things they need before adopting it further.

Physicians are cautious for good reason. But the use that worries a privacy officer most is often not the physician’s. It is a browser tab opened between two calls by a staff member who wants to help, with a consumer tool whose terms can change without notice. Nobody signed anything, and nobody knows what was sent.

A ban alone does not fix this. A practice that forbids AI and offers nothing in its place gets hidden use, which is the hardest kind to audit.

What a practice should be able to answer

None of these questions is technical. A practice is entitled to a written answer to each, and to wait until it has one.

What an installed agent changes

An agent built to measure and installed at the practice answers these questions differently. Records stay on the practice’s own machines. For each task, only the text that task needs is sent to the AI model, and the practice knows which text.

We do not claim that nothing leaves. What leaves is known, limited to the task, and can be described to a patient or a compliance officer. The difference with the browser tab is not one of degree. In one case the practice knows. In the other it does not.

We have not delivered an engagement to a US medical practice. What we describe here is what the guidance requires and what we have seen of the work in medical offices.

What does not move

The agent gives no medical advice. It does not assess a symptom, does not reassure, and never poses as a clinician. Anything that goes out in the practice’s name is approved before it leaves.

It does not decide what counts as protected health information either. That judgment stays with the practice and its privacy officer, along with the responsibility. The agent works inside a perimeter the practice has set, and it stops at the edge.

---

Read more: what an agent takes over at a medical practice and what happens to the practice’s data. Pricing starts from $1,700 per agent, then $170 per month: see what a custom agent costs.