For a few years I sat on the buying side of the table, running mobile and digital commerce at a bank. Vendors came in with decks. Now I sit on the other side, building systems that use these models for real organizations. Here is what I would ask if I were back in that chair and the pitch had “AI” in the title.
- Where does our data go? Don’t stop at “is it secure.” Ask which company’s servers, in which country, and whether that company trains its models on what you send it. The answer decides whether the pitch is even possible for a credit union or a clinic. Some of the best work I’ve done keeps the data on the device. A medical device app I built reads a cartridge’s fill level from the phone’s camera, on the phone, and nothing leaves it.
- What happens when it’s wrong? Every model is wrong some of the time. The question is who catches it. A state site I worked on puts one line under its AI assistant, “Powered by AI, please verify important details.” That’s honest. If a vendor promises it’s never wrong, walk away.
- Which model, and how long does it take to change it? The models leapfrog each other every few months. If the product only works with one of them, you’re stuck with the vendor’s choice. In the platforms I build, the model is a part we can swap. The data, the workflows, and the permissions are the product.
- Who owns the prompts, the logs, and the outputs? If you leave in two years, what walks out the door with you? Get it in writing before the pilot starts.
- What does it cost in month twelve? Per-call pricing is cheap in the demo and surprising at your real volume. Ask for the bill at the number of members, patients, or documents you actually have.
None of these are technical questions. They’re the questions a board asks about any vendor. Where is our data, who is accountable, what does it cost, and what happens if we leave?