Before you approve an AI tool, AI vendor data processing questions should be part of the conversation. Every AI demo ends the same way. The tool does something impressive in ninety seconds, the founder nods, and nobody asks where the data actually goes once the call ends. I have watched that happen in rooms full of sharp operators.
The answers you get before signing carry far more weight than anything in the demo. Features you can swap out in a month. A supplier quietly feeding your customer list into a training set is something you find out about much later, usually from someone else.
Here are the nine questions I ask every vendor before anything touches a client’s system, what a good answer sounds like, and what to do when you get a shrug instead.
Key Takeaways
- The nine AI vendor data processing questions below take about fifteen minutes to ask and tell you more about your real risk than any product demo will.
- Most AI tools are a thin layer over someone else’s model, so ask which model, which provider and which country before you ask about features.
- Sub-processors are where the exposure hides. One vendor usually means four or five companies touching your customer data.
- If a vendor cannot tell you plainly whether your data trains their model, assume it does and decide accordingly.
- Under the Privacy Act 2020, you stay responsible for your customers’ personal information even when a supplier is doing the processing.
- Retention and deletion answers matter most at the end. Ask what happens to your data ninety days after you cancel.
- A vendor who answers in writing within two working days is a completely different risk profile to one who answers vaguely on a call.
Why AI Vendor Data Processing Questions Beat a Slick Demo
A demo shows you what a tool can do. It tells you nothing about who handles your data on the way through.
Most AI products sold to small businesses are not the model. They are an interface, some workflow logic and a database sitting on top of someone else’s model. That is not a criticism; it is how the market works, and it is why these tools ship so quickly. But it means the vendor in front of you is rarely the only company handling your information.
Take a customer service assistant. One enquiry can pass through the vendor’s application server, a cloud host, a model provider’s API, a logging service, a vector database and an analytics tool. That is six companies for a single message. The vendor signed agreements with five of them. You signed one agreement with the vendor.
Under New Zealand’s Privacy Act 2020, you are the agency responsible for the personal information you collect, and handing the processing to a supplier does not hand over the responsibility. The Office of the Privacy Commissioner’s information privacy principles are worth twenty minutes of your time before you sign anything, particularly principle 11 on disclosure and principle 12 on sending personal information overseas. Australian operators sit under a similar arrangement with the Australian Privacy Principles.
None of this means AI vendors are careless. Most are perfectly fine. The problem is that you cannot tell the careful ones from the careless ones by watching a demo, and the questions that separate them are quick to ask. If you want the contractual side as well, I covered that separately in the AI automation contract checklist.
The First Three AI Vendor Data Processing Questions: Where Your Data Goes
These three establish the physical path. Ask them in this order, because each answer changes what the next one means.
1. Which model is behind this, and who operates it?
A good answer names the provider and the model family, and tells you whether it runs on the provider’s infrastructure or the vendor’s own. A bad answer is “our proprietary AI” with nothing behind it. That phrase almost always means a commercial model with a prompt wrapped around it, and it matters because the model provider’s terms govern a large chunk of what happens to your inputs.
2. Which countries does my data physically sit in?
“The cloud” is not an answer. You want named regions for two separate things: storage and processing. A vendor might store your records in Sydney and still send every prompt to a US endpoint for inference. Both are disclosures. If you hold client contracts or work in health, finance or government-adjacent sectors, data residency may already be a term you have agreed to elsewhere.
3. Who are your sub-processors, and can I see the list?
Every serious vendor maintains a sub-processor list, usually on a public page, along with a policy for notifying customers when it changes. If the answer is “just us”, they either do not understand the question or they have not read their own stack. Ask for the list in writing before you sign, not after.
Questions Four to Six: Who Sees It and What They Do With It
The first three tell you where your data travels. These three tell you what happens to it once it arrives.
4. Is my data used to train your models or anyone else’s?
You want an explicit no, with a clause you can point to. The business and enterprise tiers of the major model providers generally exclude customer inputs from training by default, while free consumer tiers have historically been looser. Watch for the phrase “we may use aggregated or anonymised data to improve the service”. Ask what aggregated means in practice and who decides. If the answer wanders, treat it as a yes and put nothing sensitive through the tool.
5. Which of your staff can see my data, and under what conditions?
The answer you want involves role-based access, audit logging, and support staff needing an approved ticket before they can view a customer account. The answer that should worry you is “our team has access as needed”. Every vendor needs some level of support access. The difference between a well-run one and a risky one is whether that access is logged and time-bound.
6. What happens when something goes wrong?
Ask for the breach notification commitment in hours, not in vague language. In New Zealand, privacy breaches likely to cause serious harm must be notified to the Privacy Commissioner and to affected people, and you cannot meet that obligation if your supplier tells you three weeks late. A vendor who has thought about this will have a written incident process. A vendor who has not will improvise on the day, using your customers’ information.
Questions Seven to Nine: What Happens When It Ends
Most founders evaluate a vendor on the assumption the relationship works out. Half the risk sits at the other end.
7. How long do you keep my data, and where is that written down?
Ask about three separate things: the records in the product, the prompt and response logs, and the backups. Model providers commonly retain API logs for a defined window, often around thirty days, before deletion. Vendor backups can persist far longer. “We delete on request” means nothing without a stated timeframe and a confirmation you can file.
8. Can I get everything out, in a format I can actually use?
Test this during the trial rather than taking their word for it. Export the contacts, the conversation history, the call recordings, whatever the tool holds. If the export is a PDF of a table, that is not an export; that is a screenshot with extra steps. If the export is behind a paid tier or a support request, you have found a lock-in you did not agree to.
9. What happens to my data if you are acquired or you shut down?
Customer data is an asset in an acquisition, and plenty of small AI companies get bought or fold. You want a change of control clause and a wind-down commitment: notice period, export window, deletion confirmation. Most vendors have never been asked this. The good ones give you a straight answer anyway, and the ones who bristle at the question have told you something useful.
How to Ask Without Sounding Like a Lawyer
Send the nine questions as a plain email before the second call. One page, numbered, with a line saying you need written answers before you can proceed. Give them two working days.
The content of the answers matters. The behaviour around the answers matters just as much. A vendor who replies within a day with specifics, including the parts that are not flattering, is a different proposition to one who wants to “jump on a quick call to walk you through it”. Written answers create a record. Calls do not.
Red flags worth acting on: refusing to name the underlying model, requiring an NDA before showing a sub-processor list, answering data residency with a marketing phrase, or telling you those protections are available on an enterprise plan only. That last one is a pricing decision dressed up as a security architecture.
Weight the questions by what you are actually feeding the system. A tool that drafts your social posts barely needs this scrutiny. A voice agent answering your phone needs all nine. When I set up an AI receptionist for Dr Claire, a dental practice owner who was losing close to half her inbound calls, every one of those recovered calls became a recording holding a patient name, a mobile number and a health-adjacent reason for ringing. Bookings went up 44%, and the reason I could put that system live at all is that the data path had been answered in writing first.
Worth reading alongside this: how founders end up with tools nobody vetted in AI automation failures in small business, and the comparison approach in AI automation companies in NZ compared.
The Bigger Pattern Underneath This
Nine questions will not make you an expert in data protection law. They will do something more useful, which is to force a vendor to describe their system in specifics instead of adjectives. That is what AI vendor data processing questions are designed to uncover. Vendors who cannot do that usually have not thought it through, and you do not want to be the customer who finds out first.
The deeper issue is that most businesses buy AI tools one at a time, each with its own data path, its own retention policy and its own copy of your customer list. Fifteen tools later, nobody can answer where anything lives. The alternative is designing the thing properly from the start: one AI brain that holds your business context, an AI workforce doing the actual work, and a small number of deliberate connections you can name. Fewer suppliers, fewer copies, fewer surprises.
Ask the nine questions. Then ask yourself why you needed a fifteenth tool.
Take the Next Step
If you are weighing up AI suppliers and want a straight second opinion on where your data would end up, book a 30-minute Discovery Call. No pitch deck. Bring the vendor’s answers, or the silence you got instead, and I will tell you what I would do in your position.
Frequently Asked Questions
What are the most important AI vendor data processing questions to ask before signing?
The three that reveal the most are: which model sits underneath the product and who operates it, which countries your data is stored and processed in, and whether your data is used for model training. Those three answers will tell you whether the vendor genuinely understands their own stack. The other six sharpen the picture, but a vague answer on any of the first three is enough to pause.
Does the Privacy Act 2020 apply if my AI vendor is based overseas?
Your obligations do not disappear because the supplier is offshore. If you collect personal information in New Zealand, you remain responsible for it, including when you disclose it to an overseas provider. Information privacy principle 12 sets out conditions for sending personal information outside New Zealand. In practice, that means checking whether the receiving country or the vendor’s contract provides comparable protection before you send anything.
Can an AI vendor use my business data to train their models?
Only if the terms allow it, which is why you ask. Business and enterprise tiers from the major model providers generally exclude customer inputs from training by default, while free consumer tiers have been less consistent. Watch for wording about aggregated or anonymised data used to improve the service, because that phrase is often left undefined. Ask for a specific clause reference rather than a verbal assurance.
What is a sub-processor and why does it matter?
A sub-processor is any third party your vendor uses that touches your data: the cloud host, the model provider, the logging service, the analytics tool. It matters because your contract is with the vendor, not with those companies, so their practices become your exposure. Any vendor operating properly will keep a published sub-processor list and notify customers before adding a new one.
Do I need a data processing agreement with a small AI vendor?
If personal information is involved, yes. A data processing agreement sets out what the supplier may do with the data, how long they keep it, what happens on termination and how quickly they tell you about a breach. Most established vendors have a standard one ready to send. A supplier who has never produced one is a supplier who has never thought seriously about the question.
What should I do if a vendor refuses to answer these questions?
Treat the refusal as the answer. Any supplier handling customer data should be able to describe their data path in writing without drama, and the good ones do it quickly because they get asked regularly. If they defer, stall or insist on a call instead of an email, either limit the tool to non-sensitive data or move on. There are plenty of alternatives.
Is it safe to put customer data into ChatGPT or Claude for business use?
It depends entirely on the tier and the settings. Business and team plans from the major providers offer stronger commitments around training and retention than free consumer accounts, and both are configurable. Check the specific plan you are on rather than assuming. The bigger risk in most small businesses is not the provider; it is staff pasting client information into personal free accounts nobody knows about.
About Octavius
Titus Mulquiney is the founder of Octavius AI, where he builds AI brains and AI workforces for founder-led businesses stuck running everything out of their own head. Twenty years in marketing, ex-Sony product manager, ex-GM Zeal NZ. Based in Auckland, working with operators across NZ, Australia, and the US. Connect on LinkedIn.