Most of this Briefing is about the AI your own people use. The same risk runs through your suppliers. Your accountants, your payroll provider, your marketing agency, your recruiters and the software you rent all handle your information, and some of them are putting it into AI tools. When a client asks whether your subcontractors use AI on their work, as their security questionnaire increasingly will, the answer depends on what you have asked your own suppliers.
Which suppliers to ask first
Start with the ones that hold your most sensitive information or do work that goes out under your name:
- Payroll, HR and benefits providers, who hold personal data about your people.
- Accountants, lawyers and other advisers, who hold your financial and legal position.
- Marketing, creative and communications agencies, whose work your clients see as yours.
- Recruiters, who handle candidates' personal data and shape decisions about people.
- Outsourced customer service, transcription, translation and call handling.
- Software and IT providers, whose products are adding AI features, often switched on by default.
Eight questions to ask
- Which AI tools, if any, do you use on our work or our information?
- Do those tools run on business accounts your company controls or on personal accounts?
- Is our information used to train any model, and how long is it kept?
- Where is it processed, and which AI vendors appear among your subprocessors?
- Who reviews AI output before it reaches us, or reaches our clients on our behalf?
- Is there a written AI policy with a named owner?
- Will you tell us before using AI on our work in a new way, and give us the chance to refuse?
- How will you tell us if something goes wrong?
For software vendors, add three more: which AI features the product has, whether each can be switched off for your account and which model providers sit behind them.
What a good answer looks like
A good answer is specific and comes with evidence, the same standard your clients apply to you. It names the tools, says which accounts they run on, points to the vendor terms on training and retention and offers the policy. If you would accept a statement as proof from yourself, accept it from them.
Some answers deserve a second question:
- "We don't use AI", with nothing to show how they checked.
- A general reassurance about security that names no tool and no control.
- No written policy, or a policy nobody can say who owns.
- A refusal to say which tools are in use, or which model providers sit behind a product.
On its own, each of these is a reason to keep asking before the supplier handles anything sensitive.
Put it in the contract
Answers change, so the important ones belong in the agreement:
- Notice before AI is used on your information in a new way.
- No training of any model on your data.
- AI vendors listed as subprocessors under the data processing agreement.
- Notice within a set time if an AI-related incident affects you.
Your advisers decide the wording, and once it is in the agreement, each promise in an answer becomes something you can rely on.
Watch the software you already rent
Software vendors add AI features through ordinary updates, and the first sign is often a release note or a notice that a new subprocessor has been added. Whenever one of those arrives, treat it as a new tool: check the feature, decide whether it is switched on for your account and update the register.
Keep it proportionate
Because not every supplier needs eight questions, put the full set to the ones that hold sensitive information and a short version to everyone else, and ask again at each renewal. Record the answers on your supplier register, which is where ISO 27001's supplier controls expect them anyway, and next to your own tools on the AI register.
Where this fits
The supplier questions sit alongside the rest of the register and policy work in Govern, and asking again at renewal is part of the governing work in Trackside. If you need an outside reference point, the question sets that clients send you, covered in the questionnaire article, work just as well pointed the other way.