Use AI safely with client data: a beginner’s playbook
For firms nervous about client data and AI. You finish with one approved tool, a written rule the team follows, client details stripped before use, connectors set up safely, and a register you can show clients and your insurer.
At a glance
- 6 steps
- 4 min to read
- About a week to put in place
- 1 template included
What you’ll walk away with
- One approved AI tool on a business plan, with the setting that trains on your chats switched off on every account
- A client data rule in your AI policy that the whole team has seen
- Client details stripped before anyone pastes work into AI
- Any connector to your systems official, read only to start, with approvals on
- A register of AI tools, a line in your engagement letter and a written answer from your broker
Learning check · before you start
What do you already have in place?
A few quick questions, so you can skip the steps you have already done. Each one has a line explaining what it means. Not sure, or never heard of it? Choose “Not yet / not sure”: that is where most firms start, and the playbook walks you through it.
-
Step 1: Choose one tool and switch training off
Done when: Everyone uses one approved tool, training is off on every account, and the provider’s answers are filed.
Your task choose one AI tool on a business plan for the team, switch off the setting that lets the provider train its AI on your chats, and file the provider's written answers on who can see your data, where it is stored and how long it is kept.
You said you already have this covered. Skim this step, or skip to the next.
How to do it, and what to read
Pick one AI tool, such as Claude or ChatGPT, and buy its business version for the team, so the firm owns and controls the accounts. Then find the setting that lets the provider use your chats to improve its AI, and switch it off on every account, including any personal ones staff already use. Ask the provider four things in writing: who can see your data, where it is stored, how long it is kept and which other companies handle it.
- Guide Client data and AI: what happens when you press enter
- Guide Where the training setting lives in ChatGPT, Claude, Copilot and Gemini
- Template AI vendor security questionnaire
Learning check
What would you do? Pick one; every answer is explained.
-
Step 2: Set the client data rule
Done when: The firm has a written green, amber and red rule for client data.
Your task write a green, amber and red rule for what client data can go into AI, and share it with the team.
You said you already have this covered. Skim this step, or skip to the next.
How to do it, and what to read
Agree what can go into the tool and what never can, using green, amber and red. Green is your own writing and public information, such as ATO guidance. Amber is client work with names, numbers and addresses taken out, in the approved tool only. Red is tax file numbers, bank details, client files with names on them and anything sensitive, such as health or family matters, which never go into an AI chat.
- Guide What can go into an AI tool
- Article The AI data-safety playbook: which client data goes where
- Template Client data rule sheet
Learning check
What would you do? Pick one; every answer is explained.
-
Step 3: Write it into your AI policy and walk the team through it
Done when: The policy names the tool, the rule and the reviewer, and every person has been walked through it.
Your task write the approved tool, the client data rule and the person who reviews AI-assisted work into your AI policy, then walk every team member through it in person.
You said you already have this covered. Skim this step, or skip to the next.
How to do it, and what to read
Write three things into your AI policy: the approved tool, the green, amber and red rule, and the person who checks work AI has helped with before it goes to a client. Then walk the team through it in person. Training alone does not stop slips, so make the safe way the easy way: everyone has an approved account, and owning up to a mistake quickly is welcomed, not punished.
- Template AI policy template
- Guide Write an AI policy in a week
- Template 30-day AI training plan tracker
- Guide Shadow AI: find what your team already uses, without blame
Learning check
What would you do? Pick one; every answer is explained.
-
Step 4: Strip client details before anything is pasted
Done when: The team strips client details and checks the highlights as a habit, with the redaction tool on each computer.
Your task install a redaction tool (which swaps client details for placeholders) on each computer, and make it a habit to strip client details before using AI, checking every highlight before you copy.
You said you already have this covered. Skim this step, or skip to the next.
How to do it, and what to read
Before client work goes into AI, swap names, numbers and addresses for placeholders such as [CLIENT_1], then put the real details back into the answer on your own computer. A redaction tool does the swapping offline, but it can miss names and addresses, so read every highlight before you copy. Also take out details that identify a client without a name, such as the only business of its kind in a small town.
Learning check
What would you do? Pick one; every answer is explained.
-
Step 5: Connect systems safely, or not yet
Done when: Every connector is official, read only to start, approved by a named person and in the register, or you have decided to wait.
Your task decide whether to connect AI to any of your systems yet. If you do, use the software company's own connector, set it to read only, and keep approvals on so the AI asks before it acts.
You said you already have this covered. Skim this step, or skip to the next.
How to do it, and what to read
A connector, sometimes called MCP, lets an AI tool reach into another app such as Xero, your practice software, email or files. If you want one, use only the connector the software company itself provides. Start read only, meaning the AI can look but not change anything, with a demo company or one client, and leave approvals on so the AI asks you before it acts. Ask the vendor what the connector can see and change and where the data goes, and bring your IT provider in before real client data flows. Waiting is a fine answer.
- Guide Connecting AI to your systems: the beginner’s checklist
- Article MCP for accountants, explained
Learning check
What would you do? Pick one; every answer is explained.
-
Step 6: Tell clients, ask your insurer and keep a register
Done when: The engagement letter mentions AI, the broker has answered in writing, and the register has a review date.
Your task add a line about AI to your engagement letter, ask your insurance broker in writing whether you are covered, and start a register of AI tools with a review date.
You said you already have this covered. Skim this step, or skip to the next.
How to do it, and what to read
Add a line to your engagement letter saying how you use AI and that a person reviews the work. Ask your insurance broker in writing whether your professional indemnity and cyber insurance cover work done with AI. Keep a simple register, a list of each AI tool and connector with what it can reach, who approved it and when you will look at it again. Review the list every quarter.
- Guide Questions to ask any AI provider, and what to tell clients
- Template AI policy template, with a tools register
- Template AI incident response plan: the first 72 hours
- Guide AI scams and deepfakes: protect your firm and your clients
Learning check
What would you do? Pick one; every answer is explained.
Free to join
Join The Firm free to read the rest
Members get every playbook in full, plus free templates, downloads and tools across the site. Join more than 35,000 accounting professionals. All it takes is your email.
Your one-page summary
Your answers, choices and decisions, on one page
Copy this
Client AI data checklist
What client data may go into an AI tool, what must be stripped first and what never goes in. Editable Word document.
Get the templateFrequently asked questions
Is it safe to use AI with client data?
It can be, with the right setup. Use one approved tool on a business plan with training off, keep identifiable client details out of the chat unless your firm has approved the tool for them, and have a person review what AI produces.
Is MCP safe for accounting firms?
It can be, set up carefully. Use the software vendor’s official connector, start read only with a demo company or one client, keep approvals on, and record each connector in your AI policy. The risk is in what a connector can reach, so keep that small.
Does our insurance cover using AI?
Ask your broker and get the answer in writing. Tell them which AI tools you use and how, and ask whether your professional indemnity and cyber cover extend to it. Policies differ.
- Playbook How to write an AI policy for your accounting firm For firms where people already use AI, policy or not. You finish with a tools register, data rules, updated engagement letters and a team that has read it. 5 steps4 min read1 week to put in placeTemplate
- Playbook AI and the TPB Code: what tax practitioners must do For Australian registered tax and BAS agents whose firm uses AI. You finish with every tool mapped, client permission on file, a review record and a policy. 6 steps4 min read2 weeks to put in placeTemplate
- Playbook How to set up Claude Cowork in your accounting firm For firms that have decided to trial Claude Cowork. You finish with a safe setup, a two-person pilot with measured results, a one-page policy and a trained team. 6 steps4 min read6 to 8 weeks to put in placeTemplate
- Playbook Build your first internal AI tool at your accounting firm For owners with one painful process and no developer. You finish with a scoped problem, a tool built for you first, tested on real files, and a decision on scaling. 6 steps4 min read2 to 4 weeks to put in placeTemplate
- Playbook The AI data-safety playbook Which client data goes into which kind of tool, the ATO and TPB angle, and a one-page rule sheet. Written playbookTemplate
- Playbook Review AI-drafted tax work Risk tiers, where AI tax errors cluster, a 12-point review checklist and what to record on file. Written playbookTemplate