The AI Register Your Clients Will Ask For Before Your Regulator Does

Very soon you will be asked something like this by your clients - do you use AI in delivering our services? Which tools? What happens to our data when you do? This usually means they are completing a due diligence pack or performing an audit - which means they will want an answer in a few days!

Suddenly you realise you cannot answer, not because of the time pressure, but because no-one has any of this written down anywhere.

This guide builds the two documents that answer those questions: a six-column AI register and a one-page usage policy. No code. No consultants. Ninety minutes of work, and the next questionnaire becomes a copy-paste exercise.

Who Is Asking, and What They Ask

Four groups are asking. Each asks a slightly different question, and the exact wording matters, because the wording tells you what evidence they want.

Clients

Procurement teams have added AI to their supplier questionnaires. It usually sits inside the annual due diligence pack, or arrives as an addendum to the data processing agreement. The question looks something like this:

"Does your organisation use artificial intelligence or machine learning tools in the delivery of the services? If yes, list the tools, describe the categories of personal or confidential data processed by each, and confirm whether any tool makes or influences decisions about individuals."

Insurers

Professional indemnity and cyber insurers have started asking about AI at renewal. They are pricing the risk of tools they cannot see. Expect something like:

"Do you use AI tools in the course of business? If so, confirm whether client data or personal data is entered into such tools, and describe the controls in place to prevent unauthorised disclosure."

Larger partners

If you subcontract for a bigger firm, their obligations become yours. Their regulator or their clients are asking them, so they ask you. This one usually arrives as a contract variation or a "supplier standards" email:

"As a condition of continued engagement, suppliers must maintain a record of AI tools used in connection with our work and provide it on request. Suppliers must not input our confidential information into AI tools without prior written consent."

Your own team

The fourth group is inside the building. Your team wants to do the right thing and currently has no way to know what that is. The question arrives one at a time, usually in private:

"Am I allowed to use ChatGPT for this? A candidate has emailed asking what we do with their CV. What do I tell them?"

Four audiences, one underlying request: show us you know what AI your business uses, what data it touches, and who is in charge of it. Answer that, and you can answer all four.

Why You Can't Answer It Today

The reason is not negligence. It is how the tools arrived. Nobody bought an AI platform. AI turned up one person at a time: a £20-a-month subscription expensed here, a personal login there, a free tier someone signed up for to finish a proposal on a Sunday. Each decision was small and rational. Nobody was ever asked to make the big one.

So there is nothing to point at. No approved list, because nobody was asked to approve. No record of what data goes where, because nobody was asked to record it. No owner, because ownership was never assigned. The questionnaire is not exposing recklessness. It is exposing that adoption happened without anyone deciding anything.

That is fixable, and fixable quickly, precisely because the estate is small. You do not have an enterprise AI problem. You have a dozen subscriptions and a gap in the paperwork.

The Register

The register is a table, not a system. Six columns, one row per tool. A spreadsheet is fine. What matters is that it exists, it is current, and it has an owner. Here is one for a fictional 14-person recruitment firm:

Tool. The product and the tier, because they differ. "ChatGPT" is not an answer; "ChatGPT Team workspace, business data excluded from training" is. The free tier and the business tier of the same product carry completely different data terms.

Who uses it. Named people or a named team, not "various". When the insurer asks who has access, "the six consultants" is an answer. "Some of the team" is not.

What data goes in. The honest version. Not "documents" but "candidate CVs, client job specs, interview notes". This column takes the most courage, because it is where you write down what is actually happening.

Influences a decision about a person? Yes or no. This is the column that changes your external answers. A tool that summarises CVs influences who gets shortlisted. A tool that drafts interview questions influences who gets hired. A tool that makes social posts does neither. Wherever the answer is yes, that row needs a named human review step, which is where the policy comes in.

Owner. One named person per tool, who keeps the row current and can answer questions about it. Not a department. A person.

Last reviewed. A date. When the next questionnaire lands, a review date from this quarter tells the reader the register is alive. A date from two years ago tells them it is theatre.

The Policy, Six Headings

The policy is one page. Six headings, with actual wording below, written to be lifted and adapted. It is deliberately plain: if your team cannot repeat it back after one read, it will not survive contact with a deadline.

1. Approved tools

The only AI tools approved for client work are those listed in the AI register, on the accounts named there. Personal accounts are not approved for client work. If you want to use a new tool on client work, ask the register owner first; adding it takes ten minutes.

2. What never gets pasted

The following never goes into an AI tool, on any account: client personal data not covered by our engagement terms, anything marked confidential, payroll or HR records, references, and passwords or access credentials of any kind. If you are unsure whether something counts, it counts. Ask first.

3. Human review where a decision affects a person

AI can draft, summarise, and shortlist. It does not decide. Where a decision affects a person - hiring, rejection, promotion, pricing to a named client, anything contractual - a named human reviews the AI output and makes the call. Their name goes on the decision, not the tool's.

4. Client data and confidentiality

Client data goes into an AI tool only if the register row for that tool says it does, and only on the business account listed there. Where a client's contract restricts AI processing, that client's data does not go in at all. When a client asks what we use, answer from the register, not from memory.

5. Who to tell when it goes wrong

If you think client or personal data has gone into the wrong tool, or AI output has gone to a client unreviewed, tell the register owner the same day. You will not be blamed for reporting it. You will be in difficulty for sitting on it.

6. Review cadence

The register is reviewed on the first Monday of every quarter, and the policy with it. New tools are added the week they are adopted, not at the next review. If either document is more than three months old, treat it as out of date.

How to Answer the Questionnaire

This is where the register earns its keep. Take the question types from section one. Each now has an answer you can give without a meeting, a scramble, or a survey of the team.

90 Minutes This Week

Where to Start

The next step is not a 40-page governance framework, and it is not a procurement exercise for a GRC platform. It is 90 minutes this week: one table, one page, one team briefing. If the questionnaire has already landed, do it tomorrow morning.

If you read this and realised you could not answer those four questions today, you are in the same position as most firms we work with. The difference, in a week's time, is that you will have the two documents, and they will not.