2026-08-26 · 4 min read

Customer security questionnaires: answering the new AI-usage sections

Which AI questions show up in questionnaires?

The questions repeat across questionnaires because they all stem from the same fear: the customer's data reaching a third-party AI vendor without control. These are the common ones:

  • Do your employees use generative AI tools? Which ones, and for what?
  • Is customer data entered into AI tools? If so, under what contractual basis (DPA, processing terms)?
  • Do your AI vendors appear on your subprocessor list?
  • Is our data used to train models?
  • What technical controls prevent sensitive data from entering unapproved tools?
  • Is there a written AI policy, and are employees trained on it?

Why is "we don't use AI" a bad answer?

Because it is almost never true — and the customer knows it. Employees adopt AI tools with or without approval, and a blanket answer that turns out inaccurate undermines the credibility of the entire questionnaire. Denial also signals you have no AI governance at all: whoever won't admit to usage certainly isn't managing it.

The customer isn't looking for a vendor without AI — they're looking for a vendor in control of its AI. An honest answer with controls beats a polished denial.

What does a strong answer to the AI section look like?

A strong answer is built from three components: policy, technical control, and proof. In practice it reads like this:

  1. Written policy — "Generative AI use is governed by an approved policy: permitted tools, permitted data types, and a mandatory process for sensitive documents." If you're unsure what counts as sensitive, start with what should never be pasted into ChatGPT.
  2. Technical control — "Documents undergo consistent anonymization before any AI input: identifiers are replaced with tokens (PERSON_001), including metadata cleaning."
  3. Data flow — "Anonymization runs locally in the browser; the original file is never sent to any server, so no new subprocessor is added to the chain."
  4. Training and records — "Employees are trained on the policy, and use of approved tools is documented."
  5. A standards anchor — teams that fold AI controls into an existing ISMS answer from a position of strength; see ISO 27001 and AI usage.

How does anonymization turn a weak answer into a strong one?

Most of the hard questions in the AI section evaporate once the data entering the tools is no longer identifiable. "Does our data reach AI vendors?" — no: what reaches them is an anonymized version with no names, IDs or customer details. "Is our data used for training?" — even in the worst case, there is nothing identifying in it. An answer that used to sink deals becomes one that supports safe AI adoption across the account.

Frequently asked questions

A customer asks if we use AI — what do we answer?

Honestly: yes, with approved tools under a written policy, and mandatory anonymization before any sensitive data enters a tool. A governed answer is stronger than denial.

Do AI tools have to appear on our subprocessor list?

If identifiable customer data reaches them — usually yes, depending on your customer agreements. When only anonymized content is entered, customer data is not disclosed to the vendor — say so explicitly in the answer.

We have no AI policy yet — what do we do before a questionnaire?

Start small: one page defining permitted tools, forbidden data types and an anonymization step for sensitive documents, then train the team. A short, true policy beats a long one nobody follows.

Should we just ban AI to simplify the answer?

Not recommended. Blanket bans are almost always violated in practice, which makes the questionnaire answer false — a contractual and reputational risk. Controlled use is both credible and a sign of maturity.

How do we answer "is our data used to train models"?

AI vendor policies vary between tools and plans, so the durable answer is minimization: when only anonymized content is entered, even the worst case exposes nothing identifying about the customer.