A client asked about your AI controls: how to reply the same day
When a client is asking about AI controls, the reply needs four parts: a short policy statement, a list of approved tools with their tier and data terms, your data rules, and the evidence you can attach. Answer those four plainly and most supplier questionnaires are covered. If one part is missing, say so and give a date for fixing it. An honest gap with a plan reads better than a vague claim.
What the client is really asking
The wording changes. The questions mostly do not. According to Everything-PR, a client's AI questionnaire tends to ask the same five things regardless of who sends it: which tools an agency uses, what tier, what data-handling terms apply, who has admin oversight, and what happens if something goes wrong.
Behind those five sits one worry. Will our data end up somewhere we cannot see? Your reply should answer that worry directly.
The client is not asking you to stop using AI. Most clients expect you to use it. They want to know it is used with care.
How to structure your reply to a client asking about AI controls
Keep it to one or two pages. Use the four headings below. Each one maps to what the client needs to tick off.
1. Policy statement
One paragraph. Say you have a written AI policy, when it was dated, and who agreed it. Name the person accountable for it. If staff have signed it, say how many and by when.
2. Approved tools
A short table works best. List each AI tool staff may use, the plan or tier, and the data terms that apply to that tier. Say who holds admin rights on each account.
3. Data rules
State what staff may and may not put into AI tools. Be specific. Client names, bank details, personal data and confidential documents are the usual lines. Say whether the tool's provider may train on your inputs.
4. Evidence
List what you can attach or show on request. Then say what happens if something goes wrong: who is told, and how fast you act.
Rule of thumb: only claim what you could show the client tomorrow. If you cannot produce it, describe it as planned, with a date.
An example answer you can adapt
Here is a reply from a fictional firm, Ashcombe, a twelve-person agency. Change every detail to match your own setup. Do not copy claims you cannot back up.
Ashcombe uses AI tools for drafting and summarising. Client data stays out of them.
Policy. Ashcombe has a written AI policy dated 6 October 2026. Priya, our operations manager, owns it. All twelve staff have agreed it in writing. It is reviewed every six months.
Approved tools.
| Tool | Tier | Data terms | Admin |
|---|---|---|---|
| ChatGPT | Business plan | Inputs not used for training by default | Priya |
| Microsoft Copilot | Work account | Covered by our Microsoft 365 terms | Tom |
No other AI tools are approved. Personal accounts are not allowed for client work.
Data rules. Staff may not enter client names, contact details, bank details or personal data into any AI tool. Client names are replaced with stand-ins such as "Client A" before use. Outputs are checked by a person before they reach you.
Evidence. We can share our AI policy, the approved-tools register and staff sign-off records. If client data is entered by mistake, Priya is told the same day and we tell you promptly.
Notice what the example does not do. It does not promise certification it lacks. It names people. It gives dates.
Check the data terms for your own tier before you write that column. Plans differ a lot, as our guide to ChatGPT Team, Enterprise and free plans explains. Copilot has its own points to check too.
Evidence worth attaching
- Written AI policy. Dated, with an owner.
- Approved-tools register. Tool, tier, terms, admin, date added.
- Staff agreement records. Who agreed the policy, and when.
- DPIA, where needed. A data protection impact assessment is a written check of the risks before you process personal data a certain way.
- Records of use. Admin or access logs from the tools, if your plan provides them.
On the DPIA point, the AI Policy Tracker summary of ICO guidance says organisations must complete a DPIA for high-risk AI processing. The ICO treats most AI involving profiling, large-scale processing or novel technology as meeting that threshold. Using a chatbot to tidy a press release is unlikely to qualify. Using AI to screen people might.
One naming point. On 30 September 2026 the ICO became the Information Commission, though it is still commonly known as the ICO, as the AI Governance Institute notes. Either name is fine in a reply.
If the client mentions ISO 42001
Some larger clients ask about ISO/IEC 42001. Konfirmity describes it as the first international certifiable standard for an AI management system, with 38 reference controls grouped into nine areas. Certification is issued by an accredited body and lasts three years.
Most small firms do not hold it. That is fine. Say so plainly. Then explain which of its themes you already cover: a policy, approved tools, data rules. Do not imply you are certified or aligned in full.
Gaps to close before you send
Read your draft against this list. Each item is a common gap.
- No dated policy. Write a one-page policy today. Date it. Name an owner.
- No tool register. Ask staff which AI tools they use. List them. Approve or remove each one.
- Free or personal accounts in use. These rarely carry business data terms. Move client work onto business plans.
- No upgrade path. If a tool's tier falls short of the client's needs, Everything-PR advises saying so directly and describing the upgrade path. Most vendors offer an Enterprise tier for this.
- No proof of staff agreement. A policy nobody signed is weak evidence.
- No DPIA where one is needed. Check any use that touches personal data at scale or makes decisions about people.
Worked example: Ashcombe finds Jo uses a free ChatGPT account. The reply says: "One user is moving to our business plan by 31 October 2026." That is honest, dated and easy to check.
Where software helps
A policy says what staff should do. It does not show what they did. That gap is what clients and insurers increasingly probe. Insurers now ask similar questions, covered in our piece on what cyber insurers ask about AI use.
Helio Guard is built for this. It checks pastes, messages and file uploads on AI sites, on the PC. Sensitive items are taken out and the question still goes. Client names become stand-ins such as "Client A". It records each person's dated agreement to your policy and produces a monthly record you can share. We never see what your team types. Windows only for now. It is £1 a person a month, from 5 people, with 14 days free and no card needed. Details are on the pricing page.
Common questions about replying to AI questionnaires
How long should a reply to a client AI questionnaire be?
A reply to a client AI questionnaire usually works best at one or two pages. Cover policy, approved tools, data rules and evidence under clear headings. Attach supporting documents rather than pasting them in. If the client sent a form, answer each question in its own box, then attach the same evidence pack.
Should I admit we use free AI accounts?
Yes, if free AI accounts are in use, say so and give a dated plan to move client work onto business plans. Clients tend to check claims later, and a discovered gap damages trust far more than a declared one. Pair the admission with the change you are making and who owns it.
Do I need ISO 42001 to answer a client's AI questions?
No, ISO 42001 certification is not needed to answer most client AI questions. Few small firms hold it. A dated policy, an approved-tools register, clear data rules and records of staff agreement answer what most questionnaires ask. State honestly that you are not certified if the client asks directly.
Related: How to find out which AI tools your staff are using, Microsoft Purview for AI without E5: what are the options?, ChatGPT Team vs Enterprise security controls: a control-by-control checklist