Not into a public, personal chat account. ICAEW guidance says confidential information should not be loaded into public generative AI tools. An accounting firm can use AI on client data in a setup the firm controls: an approved business tool or a private deployment, with agreed hosting, no training on your data, logging and a written staff policy.
#Key takeaways
- Pasting client data into a personal chatbot account is the confidentiality risk professional bodies warn about.
- Your staff are probably already using AI at work; a ban without an approved alternative pushes it out of sight.
- An approved setup answers four questions: where data is processed, who can see it, whether it trains a model and what is logged.
- Engagement letters and client contracts may say more than the law does; check them.
- Most firms need a policy and an approved tool first, and a built workflow only where volume justifies it.
#What do professional bodies say?
ICAEW has warned members that putting client data into an AI prompt will very likely breach client confidentiality, and that AI output should not be trusted blindly (ICAEW, february 2024). Its generative AI guide adds that confidential information should not be loaded into public generative AI tools, even if it has become publicly available (ICAEW generative AI guide, checked november 2026).
If you practise elsewhere, check your own institute and the IESBA code your body adopts. On top of professional rules sit data protection laws, such as the UAE's Federal Decree-Law No. 45 of 2021 (UAE government portal) and Saudi PDPL, and the confidentiality terms in your engagement letters.
#What is your team doing with AI today?
Assume someone is already using it. The Microsoft and LinkedIn work trend index found that 78% of AI users bring their own AI tools to work (Microsoft and LinkedIn work trend index, may 2024). In an accounting firm that looks like a junior pasting a trial balance into a free account to fix a formula, or a tax senior asking a chatbot to summarise a client's letter from the tax authority.
The useful response is not a ban that people quietly ignore. It is an approved tool that is good enough that nobody needs the personal one, plus a short policy on what can go in.
#What does an approved setup look like?
- A business or enterprise plan, or a private deployment, where the provider does not train on your data.
- Processing in a region your clients' contracts and local law allow, written down.
- Single sign-on, so access ends when someone leaves.
- Logs of who used which client data, kept where your quality function can see them.
- A one-page policy: which data classes are allowed, which never are, and who to ask.
- Protection against instructions hidden in uploaded documents, which the OWASP top 10 for LLM applications lists as prompt injection.
#When is an off-the-shelf business plan enough?
For most small and mid-sized practices, an enterprise version of Copilot, ChatGPT or Claude, configured properly and paired with a policy and training, covers everyday drafting and research. A private deployment or a built workflow makes sense when client data has to stay in a specific region, when you need logs on the engagement file, or when the same task runs across many clients.
#Where does Stennir fit?
Stennir's AI consulting engagement includes the governance work: which tools are approved, which data can go where, who approves AI output and how it is logged. It also ranks the workflows worth building, so the policy and the first use case arrive together.
Where a workflow needs client data to stay in-region, AI implementation builds it in the hosting region your rules require, with your firm owning the accounts and the off switch. AI training teaches staff what can go into which tool, on their own client work. More on the data side in can you use ChatGPT or Claude with customer data under PDPL.
We are
Stennir sets up approved ways to use AI on client data, with agreed hosting, logging and a named person approving output.
We aren't
Stennir is not a law firm and does not give legal opinions on your confidentiality obligations; your counsel and your institute do that.
#What should a managing partner do next?
Ask three people in different teams which AI tools they used this week and what they put in. That is your real starting point. See how compliance teams approach AI, then take the free readiness assessment or book a 30-minute call.
#Common questions
Is ChatGPT Enterprise or Copilot safe for client data?
They can be, if configured for it: no training on your data, processing in an acceptable region, single sign-on and logging. Whether that satisfies your clients' contracts and your institute's rules is a question for your firm and its counsel. Stennir helps you check the setup against those requirements.
Do we need client consent to use AI on their files?
That depends on your engagement letters, the client contract and local data protection law, so it is a question for your counsel. Many firms add a clause on approved tools and processing locations to new engagement letters. Stennir can set out the technical facts your counsel needs to draft it.
What stays with us if we work with Stennir?
Everything: the accounts, the client data, the logs and the off switch. Stennir works inside the setup your firm approves, and if you stop working together, the tools keep running under your control. Nothing from your client files is used to train a public model.
Does it work for Arabic client documents?
Yes. Arabic and English are supported for reading, summarising and drafting, and other languages on request. The same confidentiality rules apply whatever the language, and reviewers see the original document next to any summary or translation the AI produces.
