More AI vendors are approaching building societies, and most demos look impressive. A demo shows what a tool can do on a good day. These questions are about what happens on an ordinary one.
They are a starting point for a first conversation, not a full due diligence checklist. Your own risk, compliance and procurement processes still apply.
1. Where does our data go, who can see it, and is it used to train anything?
You should know where your information is held, who at the vendor can access it, and whether it is used to improve their product for other customers.
A good answer is specific and you can have it in writing. A vague answer, or one that changes when you ask a follow-up, is worth noting.
2. What does the system decide, and what does it leave to people?
There is a big difference between a tool that helps staff find information and one that makes or shapes decisions about members.
A good answer names the things the system does not do, as well as the things it does. If a vendor struggles to say where the system stops, that's a warning sign.
3. How do you check the answers are right before it goes live, and who keeps checking when our criteria change?
Any AI system can give a wrong answer. What matters is how the vendor tests it before launch, who reviews the results, and what happens when your documents change.
A good answer describes a real process, with named roles on both sides, not just “our model is very accurate”.
4. What happens when it can't find an answer in our documents?
This is one of the most revealing questions you can ask. Some tools fill the gap with something that sounds plausible. Others fall back on general information from the internet.
A good answer is simple: it tells the user it can't find the answer, so a person can pick it up.
5. Can we test it on our own material before we commit to anything?
A demo on someone else's documents tells you little about how a tool will handle yours.
A good answer is a small, well-scoped trial on your own material, with clear success measures agreed up front and an easy way to stop if it isn't working.
Why these questions matter
None of them is technical. They're the questions a sensible board member would ask about any new supplier. AI shouldn't get a pass just because it's new.
We expect AdAstr to be asked all five, and we think we should be. If a vendor finds them awkward, that tells you something.
If you're weighing up AI for your society and want a straight conversation about it, I'm happy to talk: hello@adastr.co.uk