Six questions to ask before you let an AI tool near client data

Data security is the reason accountants give most often for not using AI more: 30% overall, and 36% of those at C-suite level, in the 2025 study for Chartered Accountants Worldwide (Ipsos for Chartered Accountants Worldwide, April 2025). The worry is sensible. The problem is that it tends to stop at worry, because nobody in a small practice knows exactly what to ask.

ICAEW gave the answer its clearest form in July 2026. Firms should understand "where data is stored, who can access it, whether it may be used to train models, and how outputs are monitored", and confidential client information should not go into a system "without appropriate due diligence, contractual protections and internal approval" (ICAEW Regulatory News, 28 July 2026). The six questions below turn that into something you can send a vendor.

What should you ask an AI vendor before using it for client work?

Where the data is stored and who else processes it, whether it's used to train models, what the written contract says, how accuracy is checked, what happens when something goes wrong, and how the tool keeps up with changes to the rules. Get the answers in writing, from the vendor, before any client data goes in.

1. Where does our data go, and who else touches it?

Ask for the storage location and the list of sub-processors: the other companies that handle your data on the vendor's behalf, which for an AI tool usually includes whoever runs the underlying model. Bright Software's vendor guide, which is a marketing document from a software company and says so, puts this first for a reason: you remain the data controller whatever the vendor does under the bonnet (Bright Software Group, May 2026). A good answer names the countries, names the sub-processors and tells you how you'd be told if they changed.

2. Is our data used to train the model?

Yes or no, in writing. Many consumer AI products train on what users type unless told otherwise, and some business tiers don't by default. What you want is a contractual commitment that your data isn't used for training, not a setting someone on your team has to remember to switch off.

3. What does the contract say?

UK GDPR expects a written contract with any processor, setting out what they do with the data and the security measures they take. Ask for the data processing agreement and read the parts about sub-processors, retention and what happens when you leave. If the vendor can't produce one, the tool isn't ready for client data, however good the demo was.

4. How do you check it's accurate?

An AI tool can be confidently wrong, and in tax and accounts work that's expensive. Ask how the vendor tests accuracy, whether the tool shows its sources, and whether it will say "I don't know" rather than guess. For anything touching figures or deadlines, the answer you're looking for is that the tool works from an authoritative source it can point to, and that a person in your practice checks the output before it goes anywhere. About 80% of accountants are comfortable with AI extracting data, but only 38% with AI signing anything off (DataSnipper, 2026). That instinct is right.

5. What happens when it goes wrong?

Ask whether you can see what the tool did and why, after the event. If a client asks why a figure was wrong, or HMRC opens an enquiry, "the AI did it" is not an answer. You need to be able to reconstruct what went in, what came out and who checked it. Ask too who you contact, how quickly they respond, and what their contract says about liability.

6. How does it keep up?

Rates, thresholds and filing requirements change every year. Ask when the tool was last updated for UK rules and how you'll know when it's updated again. A tool that was accurate when it launched and never touched since is a liability with a nice interface.

What does a bad answer sound like?

"We use ChatGPT" with nothing in writing about where data goes or whether it's used for training. "Our customers haven't had any problems." A data processing agreement that's "being finalised". Any answer that arrives as a conversation rather than a document. None of these means the vendor is dishonest. They mean the tool isn't ready for client data yet.

I lead a live AI rollout inside a regulated business, so these are questions I've had to put to vendors myself rather than questions I've read about. The vendors worth using answer in writing. The ones to avoid change the subject.

What should a practice do this month?

Send the six questions to every vendor whose tool touches client data, including the AI features inside software you already use. Put the answers in your approved tools register, and don't approve anything for client data until all six are answered in writing.

• List every AI tool and feature in use, including the ones that arrived inside existing software.
• Send the six questions, by email, so the answers are in writing.
• Record each answer against the tool in your register.
• Approve for client data only when all six are answered. Everything else stays approved for drafting only, or not at all.

My AI governance kit for accountancy practices includes a vendor checklist built on these questions, the approved tools register and the policy to go with them. If you're being sold something and can't tell whether it's any good, the vendor check answers that in writing.

Sources

1. 01Ipsos for Chartered Accountants Worldwide, AI and the Future of the Chartered Accountancy Profession, April 2025.
2. 02ICAEW Regulatory News, Protecting client confidentiality: internal controls and the ICAEW Code of Ethics, 28 July 2026. icaew.com/regulation/regulatory-news/regulatory-news-2026-07/protecting-client-confidentiality
3. 03Bright Software Group, Six Questions Every Accountant Should Ask Before Buying AI, May 2026. A vendor white paper.
4. 04DataSnipper, AI Report for Audit and Finance, 2026. A vendor survey of 200+ professionals.
5. 05ICO, guidance on AI and data protection, and on contracts with processors. ico.org.uk

This article is general information, not professional or legal advice. Check current ICAEW and ICO guidance for your practice.