This is a plain-English summary of published guidance, not legal advice.
What is shadow AI?
Shadow AI is staff using AI tools at work that the firm hasn't approved, can't see and doesn't control. The usual example is someone pasting a client email into a personal ChatGPT account to tidy up a reply.
It's the AI version of shadow IT, with one difference. A rogue spreadsheet stays on your server. A personal AI account keeps what you type on someone else's, and on many personal plans it can be used to train the model.
Even if nobody has mentioned it, it is probably already happening. In Microsoft and LinkedIn's 2024 Work Trend Index, 78% of people using AI at work were bringing their own tools, usually because they had too much work and found something that saved time.
Examples of shadow AI in a professional firm
Most of these happen during ordinary client work, without feeling risky to the person doing them:
- A trainee pastes a client's email chain into free ChatGPT to draft a reply.
- A fee earner uploads a draft lease to a personal Claude account and asks for a summary.
- An accountant pastes a client's payroll export into a chatbot to check for errors.
- A consultant lets a meeting-note bot like Otter join a client call, without asking the client.
- Someone installs a browser extension that reads every page they open, including the case management system.
- A team uses Copilot on personal Microsoft accounts rather than the firm's Microsoft 365 licence.
- A partner drafts advice in an AI app on their phone and sends it without checking the facts.
Whether a tool is suitable depends on the plan tier rather than the brand. We compare ChatGPT, Copilot, Claude and Otter tier by tier in Is ChatGPT GDPR compliant?
The risks of shadow AI for law, accounting and consulting firms
Client confidentiality
Solicitors must keep clients' affairs confidential under paragraph 6.3 of the SRA Code of Conduct. Accountants have the same duty under the ICAEW Code, and ICAEW's November 2026 note on AI and client data warns that entering client data into public AI tools can risk unauthorised disclosure. Consultancies usually have the same promise in their NDAs.
UK GDPR
Your firm is the controller of its clients' personal data. UK GDPR needs a written contract with anyone processing it for you, and a lawful route for data leaving the UK. A personal ChatGPT account has neither. If it's a personal data breach that risks people's rights, you have 72 hours to report it to the ICO.
Legal privilege
The SRA's August 2026 warning notice on misuse of AI says client information should only go into AI tools with proper safeguards, and warns of a "permanent waiver of privilege".
Your regulator
Using AI doesn't move responsibility. The same SRA notice says supervisors stay responsible for work done by junior staff, and that relying on AI output isn't a defence. ICAEW's generative AI ethics guide applies the fundamental principles in the same way. We summarise both in our SRA AI guidance and ICAEW AI guidance pages.
AI slop under the firm's name
AI writes confidently even when it is wrong. In Ayinde v Haringey, the High Court dealt with lawyers who put fake case citations before the court. Day-to-day mistakes are usually smaller, such as a wrong date or an invented clause, and they slip through when supervisors do not know AI was used.
How to detect shadow AI without spying on your team
The goal is to see which tools are in use rather than catch staff out. The ICO's guidance on monitoring workers says monitoring must be necessary, proportionate and transparent, so start with the least intrusive options:
- Ask. An anonymous survey with an amnesty gets you more truth in a week than any log.
- Check expenses. ChatGPT Plus, Claude Pro and note-taker subscriptions on expense claims or company cards.
- Check connected apps. Microsoft 365 and Google Workspace both list the third-party apps people have given access to their mail, calendar or files. Note-taker bots show up here.
- Look at meeting invites. Bots joining client calls are usually visible in the attendee list.
- Count traffic by domain rather than by person. Your firewall or Microsoft Defender can show how much traffic goes to AI sites without you reading anyone's chats.
Don't read people's conversations, log keystrokes or monitor in secret. It breaks trust, the ICO guidance makes it hard to justify, and people just move to their phones.
How to fix shadow AI: five steps
Banning AI on its own rarely works. People have it on their phones, and the work that pushed them to it is still there. A five-step approach works better:
- Run an amnesty and a survey. Give staff two weeks to say which tools they use, for which tasks, and what data they have entered. Promise no disciplinary action for anything disclosed during the amnesty, and stick to it.
- Approve tools. Pick business tiers that don't train on your data and come with a DPA. Our ChatGPT, Copilot, Claude and Otter comparison shows which tiers qualify.
- Write a short policy. Cover approved tools, permitted data, output review and client disclosure. Start from our AI policy template.
- Train people. An hour per team on the policy, on checking AI output, and on what good use looks like in their own work.
- Give them something better. If the approved option is worse than free ChatGPT, people go back to free ChatGPT. A secure tool that knows your templates and your clients wins on its own merits.
Once the amnesty is done, score the firm against our AI readiness checklist. It shows which of the gaps behind shadow AI your firm still has, such as no written policy, client data in personal ChatGPT logins, or a team that can't check AI's work.
A secure alternative your team will use
Step five is where we come in. We build AI workflows inside your existing systems using your precedents and client files, with human sign-off on client deliverables. You own the software and data from day one, and we put the first workflow live within 7 days.
If you want to see where to start, our AI readiness assessment shows which workflows across your services AI can already handle.
- AI policy templateA free AI policy template covering approved tools, client data rules and output review.
- AI readiness assessmentEnter your website URL to see which workflows across your services AI can already handle.
- Private AI for your firmAI workflows running inside your own systems, on your data, with human sign-off.
