AI vendor questions are the formal checks a company runs before they let any AI product touch their data, and there are 5 that really decide things.
People ask us these at ChatFuse and we write the answers down for them. That's the only kind of answer you should accept. If they won't put it into a contract, then it doesn't count, no matter what the sales rep tells you.
Most security questionnaires run 40 items long, and they still miss at least 2 of these. A long list doesn't mean it's got more rigour.
Why do these 5 questions matter more than the rest?
Because if you get these wrong, there's no fixing it later. A bad UI just annoys people. But if they trained on your client data, they can't unlearn it. If they don't keep an audit trail, you'll never know what went wrong. The harm is already done.
Everything else on your list is negotiable. These 5 items are different. You can't solve them by changing vendors later. If you're still figuring out whether to build or buy, build vs buy AI breaks down what's worth doing yourself.
What is the right answer to "do you train on my inputs"?
The answer's a written no for everything, your prompts, files, and anything you save. And it has to flow all the way down to the model companies behind the product. That last part is the one most buyers don't think to check.
Almost every AI tool runs on models from OpenAI, Anthropic, Google, or Meta. So the vendor's promise is only part of the story. You need to know what their agreement with those providers actually says. At ChatFuse, we don't train on your inputs, and our contracts require the same from our providers. That's how we can say it clearly, no conditions.
What does data isolation actually mean?
Data isolation means your information stays inside your account. It doesn't just sit in a separate database, it never reaches other people's sessions. You should always ask whether memory, files, and history are tied to your account, not just stored apart. The 2 ideas sound similar, but how they're enforced makes all the difference.
- Separate database rows or schemas
- Encrypted at rest and in transit
- Says nothing about retrieval
- Your memory only reachable by you
- No cross customer retrieval, ever
- Enforced in code, not policy
Why does physical location matter?
Because the law that protects your data is the law of the physical place it's stored and every country it travels through. A company might be based in one country but use vendors in others, and that's what surprises people in a security review.
Get that vendor list in writing. If they can't give you one, it means they haven't done the work or they don't want to tell you. Either way, you have your answer.
You might also be surprised by how long that list gets. One AI app can need a model host, a server provider, a database for embeddings, something for analytics, and a tool for logging errors, and each one could be in a different legal zone. At ChatFuse, we keep that list very short on purpose. Every new name on it means another set of laws to follow and another contract to understand.
What should an audit trail contain?
An audit trail needs to record who did what and when, and from which location. It must be stored long enough to be useful when you discover a problem, not just while it's happening. Most incidents aren't found for weeks, so keeping logs for just 7 days isn't enough.
You also need to be able to pull these records yourself. It's important that you don't have to ask your vendor for the data. Getting it on your own counts, since you can't wait for a support ticket during an investigation.
Who at the vendor can see my data?
Most vendors will admit that some of their engineers can sometimes see your data, and the good ones are clear about when that happens. Don't trust a simple 'nobody can,' because somebody has to fix problems in production. The difference is whether that access is there all the time or only for specific reasons.
At ChatFuse, our people don't have any standing access to customer data. The technical specifics are laid out in zero trust AI data security. The key question for any provider, including us, is what steps are required for a person to look at your data and if that action gets recorded where you can check it.
Do these questions apply to free AI tools too?
They apply even more. The free tools you see are usually funded by something besides your subscription. Their terms often change in the exact spots these questions hit. And when staff use a personal free account for company jobs, that's how this shows up most.
It won't be on any procurement list because nobody bought a thing. We wrote about that blind spot in shadow AI. We also covered why business and consumer tiers split: see one AI platform for consumers and enterprise.
What should I ask an AI vendor about training on my data?
Ask them to put it in writing: nothing from your account gets used to train anyone's models. That means your prompts, your files, anything you save. Get that promise to also cover the model companies they use. A sales rep saying it or a web page isn't enough; you need it in the contract.
Is my data safe in a consumer AI tool?
It depends on the plan you pick and the contract that comes with it. The deals for personal use often skip the protections that business customers need the most. If you can't get clear answers to all 5 of those points for a service your team is using, that's not a secure setup, it's a risk you haven't measured.
What is a subprocessor list and why does it matter?
A subprocessor list shows which outside companies handle your data for a supplier, right down to the AI models. Your information might end up under a different country's laws without you knowing. The only place this gets documented is that list, so you need it in writing.
How long should AI audit logs be retained?
You need to hold onto AI audit logs for a full year. That gives you enough time to look into any issues that might not show up for months. Don't settle for just a week. You should also ask if you can run your own queries on the logs without having to file a support ticket.
When you're in the middle of an investigation, waiting on someone else is the last thing you want.
Does using multiple AI models make vendor assessment harder?
It can, unless you're dealing with a single platform that puts every model under one contract. ChatFuse handles over 100 models, all running through our set of terms and one data policy. That means your security team only has to assess one vendor, rather than a dozen. This consolidation is frequently the fastest path to getting through a security review.
Take the 5 questions to your next vendor call
Take these 5 questions with you when you talk to a vendor. What you're really listening for isn't the answer itself, but how fast they can get it and if they'll put it on paper. ChatFuse gives you all 5 answers in writing. This is what to ask for, and what a good answer sounds like.
If you're in a regulated industry, you might want to see what those questions are when the law gets involved; patient data in ChatGPT covers that.
Start free with ChatFuse to try it, or check out the details on our security page first.
Comments
Loading comments…