An email lands from your biggest client, raising B2B client AI data security questions. Their procurement lead has heard you are using AI in your workflow, and they have questions. Where does our data go. Who can see it. Is it being used to train anything. Can you put that in writing.
Most founders freeze at this point, because the honest answer is “I am not completely sure.” These questions are getting sharper every month, and they are being asked by people who have just been handed a vendor risk policy and told to enforce it.
This post covers the questions you will actually get, what a good answer sounds like, and the handful of things you need to have sorted before the email arrives. Not a legal template. The practical version.
Key Takeaways
- B2B client AI data security questions almost always reduce to three things: where the data sits, who can see it, and whether it trains a model.
- Most enterprise AI questionnaires are copied from software vendor forms, so answer them like a software vendor, not like a philosopher.
- “We do not send client data to consumer chatbots” is a stronger answer than any policy document, and it is a decision you can make today.
- Business API tiers at the major AI providers do not train on your inputs by default. Free and consumer tiers often do. That distinction is the whole ball game.
- Write a one-page AI usage statement before anyone asks for it. It converts a scary conversation into a document exchange.
- Access control beats encryption theatre. Clients care far more about which staff can pull up their data than about which cypher you use.
- The client asking hard questions is not a problem client. They are the one who takes vendor relationships seriously enough to keep paying you.
- If you cannot answer where a piece of client data lives, that is not a security gap. It is a systems gap, and it shows up everywhere else too.
Why This Question Is Suddenly Everywhere
Two years ago nobody asked. Now it is in the renewal paperwork.
The reason is simple. Large organisations have finished writing their internal AI policies, and those policies flow downhill to suppliers. If a bank tells its staff they cannot paste customer records into a chatbot, the next logical step is asking whether their marketing agency, their bookkeeper, and their logistics provider are doing the same thing. You are downstream of a policy you never saw.
There is also the near-miss factor. Most mid-sized companies have had at least one internal incident by now. Someone pasted a client contract into a free AI tool to summarise it. Someone uploaded a spreadsheet of customer emails. Nothing exploded, but the legal team noticed, and now there is a process.
So when your client asks, they are usually not accusing you of anything. They are ticking a box someone handed them. That matters for your tone. If you answer defensively, you sound like you have something to hide. If you answer like a supplier who already thought about this, the conversation takes four minutes and ends.
The founders who struggle here are the ones running everything out of their own head. They know what they do with client data, roughly, but nobody has written it down, and the answer changes depending on which staff member you ask. That is the actual problem. The security question just exposed it.
The Five Questions You Will Actually Get
Vendor questionnaires vary in length, but they converge on the same five things. Learn these, and you have covered most of what arrives.
1. Where is our data stored and processed?
They want a country and a company. “Our CRM is hosted in Australia, our AI processing runs through Anthropic’s business API in the US, and no client data is stored on personal devices” is a complete answer. Vagueness here reads as evasion.
2. Is our data used to train AI models?
This is the one they care most about, and it is the one most founders get wrong through simple confusion. The business and enterprise API tiers at the major providers do not train on customer inputs by default. Consumer chat products often do, unless you turn it off. If your entire AI workflow runs through a free chatbot on a personal account, the honest answer is uncomfortable, which is exactly why this question exists.
3. Who on your team can access our data?
Named roles, not headcount. “Two people: me and the account manager assigned to you” is a strong answer. “Everyone in the company” is not, and it is also a genuine risk you should fix regardless of who is asking.
4. What happens to our data when we stop working with you?
Deletion timeline and process. Thirty days, ninety days, whatever you can actually honour. Do not promise immediate deletion if your backups run on a rolling ninety-day cycle, because that is a promise you will quietly break.
5. Do you have an AI usage policy we can see?
If the answer is no, you have work to do. If the answer is a one-page document you attach to the reply, you have just outperformed most of your competitors in that client’s supplier folder.
What a Good Answer Actually Sounds Like
Short. Specific. Written down. No hedging.
Here is the shape of a reply that works:
“Thanks for asking; it is a fair question. Your data lives in our CRM, which is hosted in Sydney. When we use AI to draft or summarise, it runs through a business API account, which does not train on our inputs, and we never put client data into free or consumer AI tools. Three people here can access your account. When we finish working together, we delete your records within thirty days on request. I have attached our one-page AI usage statement. Happy to jump on a call if your team wants more detail.”
Notice what is missing. No jargon. No promise you cannot verify. No twelve-page PDF that raises more questions than it answers. It reads like someone who has thought about it, which is precisely the impression you want to leave.
The thing that makes this answer possible is not security expertise. It is knowing your own operation well enough to describe it in five sentences. Most founders cannot do that, not because they are careless, but because the knowledge lives in their head in fragments rather than in one place anyone can read.
This is where the gap between “we are fine” and “we can prove we are fine” shows up. Being fine is a feeling. Being able to answer in four minutes is a system. I have written before about the admin overload that comes from running everything manually, and this is the same pattern wearing a different hat.
Fix These Four Things Before the Email Arrives
You do not need a compliance department. You need four decisions made once and written down.
Split consumer AI from business AI, permanently. Client data goes through business or enterprise API accounts only. Personal chatbot accounts are for your own brainstorming, recipes, and holiday planning. That line is easy to draw and easy to explain, and it single-handedly answers the training question. The OWASP Top 10 for Large Language Model Applications is worth a skim if you want the fuller risk picture, but for most small businesses this one split closes the biggest hole.
Write the one-page AI usage statement. Where data goes, who touches it, what you do not do, how long you keep it, who to contact. One page. Update it when something changes. You will use it more often than you expect, and it usually ends the conversation rather than starting a longer one.
Tighten access to what people actually need. Most small businesses give every staff member access to everything, because it was simpler on day one and nobody revisited it. Clients ask about this more than encryption because they understand it intuitively. Fewer hands, fewer problems.
Keep a record of what your AI systems do. Not for the client’s benefit, for yours. When a system drafts an email, qualifies a lead, or reactivates a dormant contact, there should be a trail showing what happened and when. I covered this in more depth in why every business needs an AI audit trail. If you cannot reconstruct what the system did last Tuesday, you cannot answer any question about it with confidence.
The Deeper Thing This Question Reveals
Here is what I notice when founders get one of these emails and panic.
The panic is almost never about security. It is about not knowing. The data probably is stored somewhere sensible. Access probably is limited to a handful of people. Nothing is likely leaking anywhere. But the founder cannot say so with certainty, because nothing has been captured anywhere outside their own memory, and memory does not survive a procurement form.
That is the same reason the team cannot make decisions without them, and the same reason onboarding takes three months, and the same reason a week off produces a two-week backlog. The business runs on one person’s recall. A security questionnaire is just an unusually blunt way of testing that.
The businesses that answer these questions easily are not the ones with the biggest budgets. They are the ones where the operating knowledge sits in a system rather than a skull. Where the data sources are listed. Where the tools are known. Where somebody wrote down who has access to what and reviewed it in the last six months.
When I build an AI brain and AI workforce for a business, the first layer is always capturing what the business knows into something readable. It sounds like a productivity exercise. It turns out to be the thing that lets you answer a client’s hardest question in one email instead of three nervous days. For more on where AI fits for smaller operators, this piece on AI adoption for SMEs covers the sequencing.
Security questions from big clients are not a threat. They are free feedback on how well organised you are. Answer them well and you become the supplier who makes procurement’s life easy, which is worth considerably more than it sounds.
Wrapping Up
B2B client AI data security questions are going to keep arriving, and they are going to get more detailed. The good news is that the questions are predictable and the answers are short. Know where data sits, know who can reach it, know whether it trains anything, know your deletion timeline, and have it on one page.
The founders who fumble this are rarely careless. They are just carrying the whole operation in their head, which works right up until someone asks for it in writing. Fix that once, and the security email stops being frightening. It becomes a five-minute task with an attachment.
Big clients are not testing whether you are perfect. They are testing whether you have thought about it. Most of your competitors have not.
If you want to talk through where your client data actually lives and what to tidy up first, grab a 30-minute discovery call. No pitch deck, just a straight conversation about your setup and the gaps worth closing.
Frequently Asked Questions
Does ChatGPT or Claude train on my business data?
It depends entirely on which tier you are using. Business, team, and API tiers at the major providers do not train on customer inputs by default, and they publish this in their terms. Free and consumer plans have historically used conversations for training unless you opt out in settings. If you are handling client data, use a business account and say so when asked.
What should be in a small business AI usage policy?
Five things, on one page. Which AI tools are approved, what data may and may not go into them, who is allowed to use them, how long client data is retained, and who to contact with a concern. Anything longer will not be read by your team or by your clients. Review it whenever you add or drop a tool.
How do I answer a client security questionnaire about AI?
Answer like a software vendor: factual, specific, short. Name the tools, name the hosting location, name the number of people with access, and give a deletion timeline you can genuinely meet. Do not guess. If you do not know an answer, find out before replying rather than writing something vague that you may have to walk back later.
Is it safe to use AI tools with client data at all?
Yes, with the right account tier and sensible access controls. The real risks are rarely exotic. They are staff pasting sensitive information into free consumer tools, credentials shared across a whole team, and nobody knowing which systems hold what. Fix those three, and you are ahead of most businesses your size, including some considerably larger ones.
What does my client mean by “data residency”?
Which country your data physically sits in. Some sectors, particularly health, finance, and government suppliers, have rules requiring data to remain onshore or within approved jurisdictions. Check where your CRM, file storage, and AI provider actually process data, since these are often three different countries, and be ready to state each one plainly.
Do I need certifications like ISO 27001 to work with big clients?
Usually not at the small business end. Most mid-market clients accept clear written answers, a usage policy, and evidence of sensible practice. Formal certification matters mainly when selling into large enterprises, government, or regulated industries, and it is an expensive project. Get the basics documented first, then consider certification if a specific deal genuinely requires it.
What happens if a staff member pastes client data into a free AI tool?
Treat it as an incident, not a disaster. Find out what was shared, check the provider’s retention and training settings, delete the conversation, and tell the client if their data was genuinely exposed. Then fix the cause, which is nearly always that nobody told the team which tools were approved. One clear rule prevents almost all of these.
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.