What a health system will ask your AI product after Bayview

One spreadsheet uploaded to one unapproved AI tool is now a breach notification to 3,182 patients. Johns Hopkins Bayview Medical Center began notifying about 3,182 patients after a staff member, during an internal review on Sept. 21, 2026, uploaded a spreadsheet of patient information to an outside AI platform the hospital had not approved for that use.

According to WMAR, the spreadsheet covered patients admitted and treated between May and September 2026 and may have included names, dates of admission and treatment, medical record numbers, account numbers, the treating department, medications given, and provider names. The hospital said it did not include Social Security numbers or financial account details. The worker deleted the prompt, thread, and document from the tool's history, and the hospital asked the company that runs the tool to delete the data. Letters to patients began going out on Oct. 22. The Banner reported that the hospital described the platform as an unnamed third-party tool and said it is reinforcing policy and training on approved and unapproved tools.

This was an internal staff mistake at a hospital. No vendor was involved, and I am not going to read more into it than the reporting says. But if you sell an AI product to health systems, it changes the conversation you are about to have.

Why this lands on vendors

Every health system security and privacy team read that story and asked the same thing: where else could this happen? The honest answer includes their vendors. Your support engineer pasting a log file into a chatbot to debug it. Your sales engineer uploading a sample export to build a demo. Your product sending patient text to a model provider your buyer has never heard of.

Bayview's own response points at the controls they will look for in you: limits on unapproved apps and AI tools, and retraining staff on which tools are approved. If a health system is tightening that for its own staff, it will ask the same of anyone who touches its data.

Five questions to answer before the buyer asks

These are plain questions. Write the answers down once, keep them current, and you will reuse them on every questionnaire.

1. Where does our data go when your staff use AI tools? Name the AI tools your employees are allowed to use, and say whether any of them may ever receive health system data. If the answer is "none, ever," say how you enforce it. If the answer is "an approved enterprise tool," say which one and what its data terms are.

2. Where does our data go when your product uses AI? List every model provider and AI service your product calls, what data it sends to each, which region it is processed in, and whether the provider retains or trains on it. Include the services your subprocessors use, because the buyer will ask.

3. What stops a person on your team from doing what the Bayview employee did? Policy is part of the answer. Technical controls are the rest: which AI sites are blocked or allowed on company devices, whether exports of health system data are restricted, and who can pull a spreadsheet of patient records in the first place. The best answer to this question is that almost nobody on your team can get that spreadsheet.

4. If it happened anyway, how would you know, and what would you do? Can you tell which employees used which AI tools, and when? Do you have a process to request deletion from an AI provider, and do you know whether the provider can confirm it? How fast would you tell the health system?

5. Who owns this at your company? A named person accountable for AI use policy, the approved tool list, and the answers above. If the honest answer is "our CTO, part time," say so. A real name beats an org chart.

What a good answer looks like

The buyer wants evidence that you thought about this before they asked. A one-page AI use statement that covers the five questions, an approved tool list with dates, and a short data flow diagram for your product's AI features will carry most of the conversation.

The weak answers are easy to spot. "We take privacy seriously." "Our model provider is HIPAA compliant." "Employees sign an acceptable use policy." Each of those might be true and still leave the buyer unable to answer the question their own leadership is asking: if one of your people did what ours did, would we know?

Make it part of AI readiness

For a founder, this is a readiness project. You are deciding, at the leadership level, which AI uses are allowed with whose data, and then making the product and the team match. The same decisions that answer a health system's questionnaire also tell you where your own exposure is.

If you want a starting structure, the free templates in the CarbeneAI AI Governance Toolkit include a governance framework and maturity assessment you can adapt into that one-page statement.

Start with question three

Ask your team how many people could export a spreadsheet of patient data from a health system today, and from which systems. Then check whether any of those people have an AI tool on their laptop that is not on your approved list. Whatever you find is the first thing a health system will want fixed.