THE SHORT ANSWER
If your clinic does not have the required BAA covering the AI service, do not put PHI into it.
Paying for ChatGPT, Claude, Copilot or another AI subscription does not by itself make that service appropriate for protected health information.
If an outside cloud provider creates, receives, maintains or transmits electronic protected health information on behalf of a HIPAA-covered organization, HHS says the organization must have the required business-associate arrangement with that provider and otherwise comply with the HIPAA Rules. [hhs.gov]
Some AI vendors offer specific enterprise or healthcare services that can operate under a Business Associate Agreement. That does not mean every account, product tier, AI feature, personal account or employee workflow is automatically covered.
If you cannot identify the BAA, the covered service and the approved workflow, patient information should not be going into the AI tool.
A paid subscription is not a BAA. A vendor's general compliance claims are not a BAA. And deleting one obvious identifier does not automatically make patient information de-identified. HHS recognizes two methods for HIPAA de-identification: Expert Determination and Safe Harbor. Safe Harbor involves removal of specified identifiers, far beyond simply deleting an SSN.
WHAT HIPAA ACTUALLY REQUIRES
An appropriate Business Associate Agreement when the AI provider is acting as a business associate.
HHS states that a cloud service provider that creates, receives, maintains or transmits ePHI on behalf of a covered entity or business associate is itself a business associate, even in circumstances where the provider possesses only encrypted ePHI and does not hold the encryption key. [hhs.gov]
Only appropriate, authorized users should have access to systems and data containing PHI.
The practice needs appropriate visibility into access and activity.
Knowing what the organization approved isn't necessarily the same as knowing what employees actually use.
A platform capable of supporting a compliant environment does not automatically make every employee workflow compliant.
The agreement matters. So do the services being used, their configuration, access controls, devices, workflows and the behavior of the people actually using AI.
Your professional obligations did not disappear when AI arrived.
KNOW WHAT YOU'RE ACTUALLY USING
“We use ChatGPT” is not enough information to determine whether patient information belongs in it. Neither is “we use Claude” or “we have Microsoft Copilot.”
The same AI brand can include different products, accounts, features and configurations. Some may be available under a Business Associate Agreement. Others may not.
Before PHI goes into any AI tool, you need to know exactly which version your employee is using and whether that use is actually covered.
OpenAI identifies specific HIPAA-eligible products and functionality available under a Business Associate Agreement. An employee simply having a ChatGPT account does not establish that your clinic's use of it with PHI is covered.
Verify the actual workspace, account, BAA and functionality before PHI enters ChatGPT.
Anthropic offers a HIPAA-ready Enterprise configuration, but standard Enterprise use is not automatically covered. The appropriate BAA and HIPAA configuration must be in place, and not every feature is necessarily covered.
“We have Claude Enterprise” is not the end of the compliance question.
Microsoft offers multiple Copilot products and experiences. Microsoft provides HIPAA/HITECH support and Business Associate Agreements for applicable in-scope enterprise services, but “Copilot” alone does not identify the service or data path.
Verify the exact service, Microsoft 365 tenant, applicable agreement, licensing and configuration before PHI enters the workflow.
Don't ask only, “Is ChatGPT HIPAA compliant?”
If you can't answer those questions, don't put PHI into the AI tool.
AI services, product terms and covered functionality change. Verify current vendor documentation and applicable agreements before allowing PHI into an AI workflow. Last reviewed: November 2026.
THE RISK IS ALREADY INSIDE
“Can you clean up this patient note?”
Someone pastes clinical notes into an AI assistant.
“Summarize this referral.”
A referral gets uploaded because reading it manually takes ten minutes.
“Draft an appeal from this denial.”
Insurance correspondence containing patient information becomes AI prompt context.
“What does this report mean?”
A lab result, diagnostic report or other patient document gets uploaded for explanation or summarization.
“I'll Just Use My Own Account.”
The approved process is inconvenient, restricted or nonexistent. The employee already knows that personal ChatGPT or another AI service can perform the work in seconds. So the employee uses it.
And now we've introduced Shadow AI.
Most employees aren't trying to create a security incident. They're trying to finish their work.
Shadow AI often begins when an employee discovers that an AI tool can eliminate ten minutes of tedious work, while the organization's approved way is slower, more complicated or doesn't exist at all.
The biggest AI risk in a small clinic may not be the AI you approved. It may be the AI you don't know someone is using.
THE RISK OF DOING IT WRONG
Your staff may already be using AI to summarize patient notes, review referrals, draft appeals or make sense of clinical documents. The productivity is real. So is the responsibility for where that patient information goes.
HIPAA does allow cloud services to handle electronic PHI, but when a cloud provider is acting as a business associate, HHS requires an appropriate BAA and continued compliance with the HIPAA Rules. The healthcare organization also remains responsible for understanding the environment and managing its risks. That is very different from employees copying patient information into whatever ChatGPT, Claude or Copilot account happens to be open.
"We scrub the patient information first” is not a safe shortcut. Removing a patient's name or Social Security number does not automatically make the information HIPAA de-identified. HHS recognizes two formal methods: Safe Harbor and Expert Determination. Safe Harbor requires removal of a defined set of identifiers and no actual knowledge that the remaining information can identify the individual.
Instead of asking every employee to make a compliance decision every time they type a prompt, consider a simpler architecture.
Give employees the AI capability they are already asking for, but bring the AI to the sensitive data instead of continually sending sensitive data to external AI services.
A private AI factory can keep approved AI workloads on infrastructure controlled by the practice. Private AI does not eliminate HIPAA responsibilities. Access controls, security, risk analysis, auditing and workforce policies still matter. But, private AI eliminates an enormous risk: sending patient data to an external AI provider just to use AI.
Keep the AI close while keeping patient data private.
PRIVATE AI, RIGHT-SIZED
Private AI no longer has to mean enterprise-scale infrastructure. A small practice can run useful AI models, document search and automated workflows on compact infrastructure inside the office.
Incoming Records
Condense outside records to the information staff or clinicians need to review.
Referral Correspondence
Create first drafts based on documentation kept within the practice's private environment.
Prior Authorization & Appeals
Draft correspondence using information kept within the private AI environment.
Internal Knowledge
Ask questions across practice policies, procedures and internal reference materials.
Clinical Drafting Assistance
Prepare first drafts from clinician-provided notes for human review and approval.
Bring the AI to the work. These workflows can run inside the practice's private AI environment rather than requiring sensitive work to be submitted to public AI services.
Venturi's Micro AI Factory is designed to bring private AI capability into a small office rather than requiring the office to send its sensitive work to public AI services.
Keeping AI on premises can remove important third-party data risks, but the practice still needs appropriate administrative, physical and technical safeguards.
A PRIVATE ALTERNATIVE
AI can give a small clinic valuable time back. The challenge isn't stopping employees from benefiting from AI. It's giving them a useful, approved way to work without creating a new blind spot around patient information.
Venturi's Micro AI Factory brings private AI into your office, sized for your people and workflows, so sensitive work can stay on infrastructure you control.
Size the infrastructure to the work, the users, and the busiest hour, not to a generic hardware recommendation.
Venturi TS, an NVIDIA Inception program member.
