Many company policies are nodded at during onboarding and forgotten thereafter. Take the humble meeting-room policy: how many employees actually remember to cancel a booking when plans change or treat the space as their own personal office?

But some directives can have a material impact on how the business operates and the risk it accepts. A policy on AI use is one of them.

A 2026 Thomson Reuters report found more than a third of professionals (34%) use unapproved AI tools for work. Microsoft's Work Trend Index puts it much higher, estimating about 80% of workers use AI independently – a habit it calls ‘bring your own AI’.

Personal use of AI at work is the new shadow IT, says Gavin Heaton, co-CEO of Disruptors Co, an innovation consultancy in Sydney. “Easy access trumps everything,” he says.

The need for rules on how employees use AI has never been greater. Apps such as ChatGPT and Claude can make an individual much more productive. However, without any guidelines a company risks important information ending up in the public domain. Yet only about half of Australian and New Zealand businesses have any policy for AI use at all, according to Datacom research.

So, the question for a practice is no longer whether staff use AI. It is whether anyone has told them how.

Define acceptable AI use

The best approach is to write the rules down in two parts, says Kayur Patel FCA, director of GenAI and emerging tech at PwC New Zealand. First, staff need to know what they are allowed to do – which tools they can use and for what. Second, clients increasingly ask how their accountant uses AI safely. They need something ready to hand over, Patel says.

A good policy starts with a list of approved tools. Patel recommends always using the enterprise versions – ChatGPT Enterprise, Copilot and the like, not the free consumer apps. PwC New Zealand uses ChatGPT Enterprise, he says.

The second pillar is human oversight and here accountants have an advantage: they already have it.

“Nothing goes to a client without being signed off by someone with a CA [accreditation],” Patel says. A manager reads and reviews the work whether or not AI touched it, and that does not change.

“Someone has to read every single word – and check every number. The review process a firm already runs is an excellent control for AI.”

Manage your AI budget

For smaller firms, the harder question is how to provide access at a reasonable cost. A tiered approach is more effective than a blanket yes or no, Heaton says. Give everyone the free tier of a tool like Copilot, which cannot upload documents or run agents. Then allow your staff who need more features or capacity to access the paid version, with its agents and custom assistants.

Any IT tools or functions designed for more than a single team – a department or the whole firm – go to IT to build and deploy, Heaton adds. Staff can build and test locally but cannot push a tool to everyone, or save it into shared drives, until it is proven and approved.

Your power users are the biggest danger to your budget. Once you start using more powerful models on Claude and ChatGPT, you can switch to a consumption model that provides credits for chunks of computing power, measured in tokens. It is not possible to assign an exact number of tokens to a task, so running agents and sub-agents on various missions becomes an exercise in juggling token allocation per month.

Application programming interfaces (APIs) and unsupervised agents that call a model in the background “can kill your budget pretty quickly”, Heaton says.

The temptation is to buy more tokens when you run out. There are many stories of eager employees blowing their monthly allowance in three days and returning for more.

A good AI policy can use budgets to enforce discipline. Capped accounts cut themselves off at a monthly limit.

Protect client data

Next comes client privacy. Staff can paste unencrypted source data – a client P&L, financials, names – into a personal account, possibly onto servers offshore, without stripping anything out. ChatGPT and Claude may use data from consumer accounts to train their models. Human reviewers at Anthropic and OpenAI could see client data added to these accounts.

A policy that names what data can and cannot go into a tool heads that off before it happens.

One of the most important additions to a policy for a fast-moving technology? A date showing when it was last updated.

“The principles stay the same,” Patel says. “What changes are the specifics.”

A model's data-retention terms can move it in or out of scope week to week. Claude’s most powerful model, Fable, has 30-day retention policies in the US. Firms must set a review cycle and hold to it.

What to include in an AI policy

A practical AI policy should spell out the following:

• Purpose and scope: why the organisation uses AI, who the policy applies to and which tools or systems are covered.
• Approved and prohibited uses: what AI can and cannot be used for in day-to-day work.
• Data rules: what information staff may enter into AI tools, including limits around confidential, personal, client or commercially sensitive data.
• Roles and accountability: who owns the policy, who approves AI use cases and who is responsible for monitoring compliance.
• Risk and approval process: when higher-risk AI use needs review, escalation or formal sign-off.
• Human oversight: when people must check, verify or override AI-generated outputs.
• Transparency and records: when AI use should be disclosed, and how decisions, tools and approvals should be documented.
• Issue reporting: how staff should report errors, misuse, bias, privacy concerns or unexpected AI behaviour.
• Training and expected behaviour: what staff need to know before using AI at work.
• Review cycle: how often the policy will be updated as tools, laws and business needs change.

CA ANZ AI resources

CA ANZ offers a range of practical courses and tools designed to build your confidence and capability with AI in your workplace. Discover the resources