Case study · Two UK tax practices
Bringing AI into two tax practices, with the guardrails built in
Rolling out Claude at two tax practices: partners first, rules set in the tool, file access limited to one folder, and an audit trail that keeps no prompts.
Published
- AI-native software
- Security
- Managed IT
Two tax practices, one of about 85 people and one of about a dozen, wanted their teams to use AI without client data going somewhere nobody had approved. We’re rolling out Claude at both, with the rules built into the tool and the way it’s set up, not just written into a policy.
The challenge
Partners were being asked what the practice was doing about AI, and some people were already trying free tools on personal accounts. The worries were the obvious ones: client data going into a service nobody had checked, and AI output reaching a client without anyone reviewing it.
Banning AI wasn’t realistic, and a policy on its own doesn’t stop a mistake. The safeguards had to sit in the tool and the devices it runs on.
What we set up
- A checked tool on business terms. Before anyone used it, we reviewed Claude’s business plan for data protection: that it doesn’t train on the practice’s data, the data processing agreement, and where data is processed.
- Rules set in the tool itself. Instructions that apply to every chat across the practice: name the specific legislation rather than invent a reference, never fabricate case law, say so when unsure, treat itself as a productivity tool whose tax output a qualified person must review before it reaches a client, and don’t repeat back sensitive client identifiers.
- The desktop app, deployed centrally. Installed through Intune on the practice’s managed devices, with its security settings applied from one policy rather than laptop by laptop.
- File access limited to one folder. For the features that work with files on the computer, each person’s access is limited to a single folder, enforced centrally, so the AI can’t reach their OneDrive or mapped drives.
- An audit trail that never stores prompts. Usage is recorded centrally for oversight and reporting, but the text of prompts isn’t kept, so the audit trail can’t become a second copy of privileged client material. It records what was produced, not what it said.
At the smaller practice the same approach runs on its Cloud PCs, with the audit records kept in the practice’s own cloud account and web search switched off.
How we did it
Partners first. At the larger practice, nine partners were set up before anyone else, with a short safe-use note on what can and can’t go in, while the acceptable use policy caught up with AI.
The rest followed in stages: the audit trail and reporting, the desktop app, the security settings on every device, and then the file features, switched on once each device was ready for them.
Why it matters
Most of the risk with AI in a practice sits in the gap between what has been approved and what people are quietly using. An approved tool that’s set up properly closes that gap better than a ban does. Our post on using AI with client data covers the questions to settle first, and AI-native software & agents shows what we build on top.
Know someone who’d find this useful?