What happens to your data when you use AI
Not everything, and not always. Here is the specific traffic, and the five questions that get you a straight answer from whoever is selling.
Every practice owner asks this eventually, usually quietly and usually late: if we start using these tools, where do our client records end up?
Right question. The honest answer is narrower than either "it is all fine" or "never touch it".
The actual traffic
Most business systems that use AI do not ship your database anywhere. They send one thing, at the moment it is being worked on.
Read a scanned referral so it becomes searchable, and that document goes to a reading service. Ask an assistant a question, and the question plus the material it is allowed to answer from goes to a model. That is the traffic. Appointments, clinical notes, invoices and everything else stay where they are unless somebody deliberately wires them in.
So the question is not "do you use AI". It is which specific things leave, and when. A supplier who cannot answer that in two sentences has not thought about it.
The five questions
What exactly gets sent, and can I see the list? Short answers are good. Vague ones are the warning.
Is any of it used to train anything? For business tools this should be no, and it should be written down, not said on a call.
Can I switch it off and keep working? This is the one that matters most. In the systems I build, document reading is a switch. Turned off, documents are stored and viewed exactly as uploaded and searchable by name. The practice loses a convenience, not a record. If switching a feature off breaks the product, it was never a feature, it was a dependency.
Where does the data physically live, and can I take a copy today? Not the marketing answer. Which server, in whose account, and can you export the lot this afternoon without asking permission.
What happens when we stop paying? Ask now. Nobody has ever got a better answer to that question during the argument that prompted it.
What good design looks like from the inside
One database per client business. Not everyone's records in one pile with a column marking whose is whose. Separation should be structural, not a filter someone could forget to apply.
Least data, not most. A system that sends only the document being read is safer than one that ships your whole file store off to be indexed, and it is cheaper to run.
Sensitive things off by default. Anything that sends a client document outside the building should arrive switched off and be turned on deliberately, in writing. I have a medical client whose contract says exactly that, and the switch in their system genuinely is off.
A plain page you can hand over. Practices holding medical or financial records often have their own obligations upwards. Being able to give a regulator or a board a page saying what leaves, to whom, and how to stop it, is worth more than a certificate.
The thing people have backwards
The risk is usually not the AI service. It is the WhatsApp thread where client documents sit on three staff phones forever. The shared login four people use. The spreadsheet emailed to a personal address. The folder anyone in the building can open. The laptop that left with the person who resigned.
Most businesses worrying about a reading service are leaking far more through habits nobody has examined in years. Fix the habits first, then ask the five questions, get specific answers, and write down which switches are on and who turned them on.
That is a real answer to a real concern, and it takes an afternoon rather than a policy project.