Eleven questions to ask an AI vendor before you sign the BAA

A procurement checklist for healthcare organizations evaluating AI vendors, written from the vendor side of the table.

Christian Armstrong, MD, MHSA Chief Executive Officer, MedPAC Technologies July 30, 2026 6 min read

The short answer

Before signing a Business Associate Agreement with an AI vendor, a healthcare organization should establish: whether protected health information is used to train models, which subprocessors receive PHI, where data is hosted, how long data is retained after termination, whether outputs are traceable to source records, who is liable for a clinical decision influenced by the tool, and whether the vendor will support an audit. Get each answer in the contract rather than in a sales call.

Ask explicitly whether your PHI trains models — for the vendor and every subprocessor.
Get the subprocessor list in writing, with notice requirements for changes.
Define the post-termination export window before you need it.
Establish whether outputs are traceable to source records; untraceable output is unauditable.

We sit on the vendor side of this table. Here is the list we would want a customer to bring, including the questions that are inconvenient for us.

Data

Is our PHI used to train or fine-tune any model, yours or a subprocessor’s?
Who are your subprocessors today, and how much notice do we get when that list changes?
Where is our data hosted, and does it leave the country?
What is the retention period after termination, and in what format do we get an export?

Clinical

Is every output traceable to the source record that produced it?
How is the tool labeled to end users — advisory, or decision-making?
What is your process when a clinician reports that an output was wrong?

Accountability

Who is liable when a clinical decision is influenced by your tool?
Will you support an audit, and at whose cost?
What is your breach notification timeline, in hours?
What happens to our data if you are acquired?

A vendor who answers all eleven in writing is not necessarily the right vendor. A vendor who will not is definitely the wrong one.

Questions from this piece

Should an AI vendor be allowed to train on our patient data?

That is a decision for your organization, but it must be explicit in the Business Associate Agreement rather than assumed. Many healthcare organizations prohibit training on their protected health information entirely, and a vendor unable to accommodate that in contract language is telling you something.

What happens to our data when we terminate an AI vendor?

Whatever your contract says. Establish the export format and the window before signing; a thirty-day default is common and is often too short for a health system to complete a migration.

Keep reading

Clinical AI Sep 2, 2026 7 min

What post-acute teams get wrong about clinical AI

Most clinical AI pilots in skilled nursing fail on workflow, not accuracy. What actually determines whether a model changes care.

AI Build Aug 19, 2026 9 min

We build software with AI agents. Humans orchestrate.

How MedPAC actually ships: AI agents do the typing, human staff direct the work, and every change is traceable from prompt to production.