What your AI policy has to cover
A guide to building one your people can actually use.
In brief
- Your AI policy has to reflect your tools, your clients and your obligations, so it has to be yours. Build it, don't download it.
- Cover six areas well: scope, confidential information, accountability, record-keeping, disclosure and agentic AI. Then add two things that policies often miss: training, and what to do when something goes wrong.
- Write it so your people understand it. The method, with prompts for every step, is the companion download: Build your AI policy: a step-by-step method.
By Inbal Rodnay
This is not like your other policies
Think about what your other policies actually do. The confidentiality policy, the code of conduct, the one about company property.
They write down what your people already know. Nobody reads the confidentiality policy to find out what confidentiality is. They arrive knowing not to leave client papers on public transport, and the policy makes the expectation explicit and enforceable.
That is codification, and it works because the common sense came first.
AI is not there yet. We have not grown up with it.
Is it fine to let a notetaker join a client meeting?
Does it matter which one?
If the AI is built into our office suite, can I put a client's tax file number or IRD number in it?
If you are not certain about the answer to these questions, your team is not certain either.
That uncertainty has two costs. Some people will make mistakes. Others will avoid the technology altogether or use it in ways that are more restricted than necessary.
What an AI policy has to do differently
AI policy has to do three things your other policies do not.
- It has to teach, not just state. "Don't put confidential information into AI" is not helpful if people cannot tell which tools are safe and what to do with the AI that shows up in tools they’ve been using forever.
- It has to stay current. Your confidentiality policy barely changes year to year. AI capability changes every few months, so the review rhythm is not housekeeping, it is real policy work.
- It has to make calls on ideas that are new to us. Agentic AI, disclosure to clients, whether a notetaker can sit in on a meeting and where the transcripts get saved.
There is no established practice to codify, so your policy has to make a call (and possibly change that call as you learn more and as the technology evolves).
This is not theoretical. In 2023 a child protection worker at a Victorian government department used ChatGPT while drafting a report for the Children’s Court. The privacy regulator found, on the balance of probabilities, that personal information, including names, had been entered, and that the department had breached its privacy obligations. Among the reasons: there were no specific rules on when and how staff could use generative AI tools, and the department’s AI training had reached only managers and leaders.
That staff member likely had no malicious intent. A good policy could have saved them the whole ordeal.
Why a list of rules is not enough
Uncertainty makes some people overshare and others freeze. A good policy fixes both, by educating in ways that a one-page list of rules doesn’t.
"Never share confidential information with AI" sounds safe, but AI is already in your office suite and, with the right licence, can see your files and emails. So how does that work?
"Don't use public AI tools" sounds firm until someone asks what a public AI tool is. Is ChatGPT public? What if I use a work account?
"Keep a record of AI use" sounds sensible until someone asks what needs to be recorded. Every prompt? Every draft? Every time Copilot summarises an email
Understand first, decide second
You cannot write a good AI policy until you understand what AI can do. Every decision in the policy flows from capability. Once you understand the capability, you can see the risks. Once you see the risks, you can make informed decisions.
Take AI meeting assistants. The notetaker joins the call, records it, transcribes it, and stores the transcript somewhere. Some tools store the audio and video too. Where is that data stored? For how long? If the transcripts land in your own file system, who can read them, and could they be subpoenaed or required in legal proceedings
You cannot answer "can we use notetakers?" until you have understood all of that. Once you have, the decision almost makes itself. See: AI meeting assistants: more than a note-taker.
Then there are AI tools that can now drive your browser and click around for you, such as Claude in Chrome and Copilot in Microsoft Edge. They work in websites you are already signed into, so they can act in your systems as you. Both vendors warn about prompt injection: hidden instructions on a web page that try to take over the AI. A sensible starting point: allow them only on sites you choose, and keep them away from banking, payroll, and tax or government portals until your policy covers them.
This is also why the review rhythm has to be much more frequent than for any other policy. AI is evolving quickly. New capabilities and new risks emerge all the time. Every six months is a reasonable review cadence for an AI policy.
What every AI policy needs to cover
Whatever process you follow, a workable policy for an organisation handling client and internal confidential information needs to cover at least these six areas. For each, here is what it can look like in practice.
1. Scope. Say plainly which types of AI the policy covers. Then make clear that it covers AI wherever it appears, not just the obvious AI chatbots. That includes the AI embedded in your accounting or ERP system, your practice management and other business apps, your document tools and your inbox, which is where most AI use in your organisation is already happening. It also includes the sub-processors those apps use, and the new features, connectors and agents that keep arriving inside tools you have already approved. Approving a tool does not automatically approve everything it can do. See: Is my data safe to use with this AI tool?
In practice: list the AI you have approved, including the AI features inside your main software, and say that any new AI feature, connector or agent needs approval before use, even inside a tool you already have.
2. Confidential information. What can and cannot go into AI, and into which tools. Supplying one AI tool and banning all others sounds clean, but it also bans everyday features such as AI answers in search engines. Decide instead which tools may be used with confidential information. Retire the phrase "public AI tools". Instead, talk about work-supplied vs personal accounts, and specify what information is approved for use in each. Be clear that paying for a personal plan does not make it a business plan. See: Is my data safe to use with this AI tool?
In practice: a one-page table. One row for each approved tool and account type, such as your work Copilot or Gemini, your practice or finance software, and personal accounts. One column for each kind of information, such as public, internal, client confidential and personal or sensitive. Tick what is allowed. Anything not on the table needs approval first. Tax file numbers carry extra rules in Australia: the Privacy (Tax File Number) Rule 2015 applies, and the Notifiable Data Breaches scheme covers tax file number information even in small businesses. In New Zealand, treat IRD numbers as confidential client information.
3. Accountability when AI did the work. Under the professional standards that apply to you, accountability stays with you, not the tool. The policy should say so clearly.
Reviewing AI output is not optional, and "the AI did it" is not an answer you want to give a client or a regulator. Some teams now add a second check on top of the human one: another AI reviews the work before it goes out. See: How to check AI's work.
In practice: “The person who signs off the work is responsible for it, whatever tools helped produce it. Check AI output in proportion to the risk before it is used.”
4. Record-keeping. Accountability asks who is responsible. Record-keeping asks whether you can prove what happened. A generic “keep a record” is not enough. People use AI in small tasks and large tasks. When do they need to keep a record? How much detail? And where does it get stored? It also covers your decisions about tools. Keep the vendor terms you relied on and a short note of why you approved each tool. If something goes wrong, that note is your evidence that you did the due diligence.
In practice, a simple test: keep a record when AI output goes to a client or regulator, informs advice or a decision, or ends up in a document you must retain. Note the tool, what it did and who reviewed it. Drafting an internal email usually does not need a separate AI-use record. In the public sector, AI outputs can themselves be public records (the National Archives of Australia and Archives New Zealand both say so), so involve your records team.
5. Disclosure. When and how do you tell clients and internal stakeholders that AI was involved? Your engagement letters, where you use them, can explain relevant AI use in providing services and record any client permission required. Your privacy policy may also need to reflect the data handling that follows, including any overseas disclosure (in Australia, APP 1 and APP 8; in New Zealand, IPP 12). Sometimes disclosure is required for a specific document. In the NSW Supreme Court, for example, an expert who has the court’s leave to use generative AI must identify which parts of their report it was used for, and the program and version used. That matters for forensic accountants. The Practice Note is reviewed periodically, so check the current version, and other courts, including in New Zealand, have their own guidance on generative AI. For Australian registered tax and BAS agents, the Tax Practitioners Board's guidance, TPB(GS) 55/2026, notes that, depending on how a tool is configured and used, entering client information into it may be a disclosure to a third party that needs your client's permission. That permission can be given in a signed engagement letter. In New Zealand, the confidentiality principle in the NZICA Code of Ethics applies.
In practice: one standard sentence for engagement letters or, in-house, for board and committee papers, saying where AI was used and who reviewed the output. In government, also check any AI transparency requirements or expectations your jurisdiction sets.
6. Agentic AI. Agentic AI is different because it is about action, not just information. Custom agents and increasingly our workplace AI tools (Copilot Cowork, Claude Cowork, ChatGPT Work) can work with files, draft and send messages and, in some cases, act in a browser. That brings new risk, because software that can act can also make decisions and behave unpredictably. Your policy needs to say who may build or switch on agentic capability, what has to happen before an agent goes live, and whether a register is required. The same goes for skills, the saved instructions an AI follows, and for connectors that link it to your other systems. A skill from an unknown source is a set of instructions your AI will follow without question, so decide who may install one and what review it needs first.
In practice: a register with one line per agent: what it does, who owns it, what it can read and change, who approved it and when it was last tested. The Australian Government publishes a free AI register template you can start from.
Agency is the blind spot. Many AI policies were written when AI could only answer questions. If your policy was written before your tools could take actions, it needs updating.
Two things policies often miss
Training. Make AI training a condition of access, and refresh it when tools change. A policy no one has been taught is a policy no one follows. Remember the Victorian case: training had reached only managers and leaders.
What to do when something goes wrong. Say what people should do if client or personal information goes into the wrong tool: stop, tell a named person, and record what happened. Name who decides whether it must be reported. Some breaches must be notified: in Australia under the Notifiable Data Breaches scheme or, for state and territory agencies, any scheme your jurisdiction runs; in New Zealand under the Privacy Act 2020. A policy that makes it safe to own up gets you the early warning you need.
Build, don't download
Now the good news. AI can be very helpful here, you just need to be smart about how you do it. Type "write me an AI policy for my accounting firm" into a chatbot and you get something generic, confident, and not yours.
The better way has four steps: research, understand, draft, review. AI can help at each step, including helping you learn, and you end up making informed decisions and approving the final policy.
For the method, with the prompts for each step, see: Build your AI policy: a step-by-step method.
If you want a checklist to test your draft against, the Australian Government’s National AI Centre publishes a free AI policy guide and template, and New Zealand’s MBIE publishes Responsible AI guidance for businesses. Use them to check your coverage, not to replace your own thinking.
A policy you can trust, not a box you ticked
What you end up with does the three things an AI policy has to do. It teaches, because it was written by someone who understood the tools before they wrote the rules. Your people can follow it, because it builds the judgement it asks them to use. And it stays current, because you built the review rhythm in from the start.
Teaching does not have to make the policy long. Some organisations keep the policy itself to a few pages and put the explanations in a separate guidance note, or in a set of one-page charts such as what can be used for what, with a clear path for anything that is not on it.
Building it this way has one more benefit. Once the policy exists, the same AI that helped write it can review it. Every six months, give it the policy and ask it to check current vendor and regulatory guidance for changes that affect it. A policy that reviews itself, with a human checking the sources and deciding what to change.
Your next step: see Build your AI policy: a step-by-step method and follow the steps to create, review or update your AI policy.
Sources
Office of the Australian Information Commissioner, About the Notifiable Data Breaches scheme
Office of the Privacy Commissioner (New Zealand), What is a notifiable privacy breach?
Archives New Zealand, Artificial intelligence and public and local authority records
Department of Industry, Science and Resources (National AI Centre), AI policy guide and template
Department of Industry, Science and Resources (National AI Centre), AI register template
Anthropic, Use Claude in Chrome safely
Microsoft, Browse with Copilot in Microsoft Edge
Tool capabilities and external source links correct as at November 2026. This guide points to external policy resources; it does not reproduce them. It is general guidance on developing an AI policy and is not legal advice. Align your policy with your obligations under APES 110 or the NZICA Code of Ethics, as applicable; in Australia, the Tax Agent Services Act 2009 and its Code of Professional Conduct, the Privacy Act 1988, state or territory privacy laws and the Privacy (Tax File Number) Rule 2015; in New Zealand, the Privacy Act 2020; and any sector regulator requirements that apply to you, such as ASIC or APRA in Australia.
Disclaimer
This guide provides general guidance only and is current as at October 2026. Tool names, features and capabilities were correct at that date but may change. The publication therefore focuses on general principles rather than specific feature lists.
This guide is not intended to be comprehensive and is not a substitute for professional, legal or other advice tailored to your circumstances. Use it alongside your organisation’s artificial intelligence policies and your applicable legal, regulatory, professional and ethical obligations. Check current vendor terms, product information and your own obligations before acting on or relying on the information.
While Chartered Accountants Australia and New Zealand (CA ANZ) seeks to ensure that the information is accurate and current, it does not guarantee its accuracy, completeness, currency or suitability for any particular purpose. To the fullest extent permitted by law, CA ANZ disclaims all warranties and liability for any direct, indirect, special or consequential loss or damage arising out of or in connection with any inaccurate statement, opinion, advice or information contained in this publication, any omission from it, or the use of or reliance on this [guide/document] or any material it contains.
© 2026 Chartered Accountants Australia and New Zealand ABN 50 084 642 571