Safe AI adoption: From experimentation to everyday practice

By Emma Musgrave from Institute of Professional Accountants

Our Co-founder James McPhedran recently sat down with Emma Musgrave from Institute of Public Accountants to talk about how accountants are using AI, and where it’s heading next.

For many accountants, artificial intelligence (AI) is no longer confined to the occasional chatbot query or a quick draft of an email. It is becoming part of the way firms research tax matters, prepare advice, review documents and manage routine work.

The shift is creating opportunities to improve efficiency and free up time for more valuable client conversations. It is also prompting questions about accuracy, confidentiality and the professional responsibilities that remain firmly with the practitioner.

Jimmy McPhedran, Founder and Chief Product Officer of Elfworks, says the change has been noticeable.

“The biggest change is that AI, for some practitioners, has moved from the side of the desk into the middle of the workflow,” he says.

“Not long ago most practitioners were experimenting with a general chatbot to tidy up an email. Now AI is showing up in tax research, first-draft advice, reviewing long documents like trust deeds, and preparing workpapers and resolutions.”

For McPhedran, the next stage of adoption is less about what AI can produce and more about whether accountants can confidently use the results.

AI moves into the workflow

One of the most significant developments is the growing presence of AI within software firms already in use, sometimes without practitioners making a deliberate decision to adopt it.

More sophisticated tools are also beginning to perform multi-step tasks with less human intervention.

At the same time, the regulatory conversation has evolved. In July 2026, the Tax Practitioners Board (TPB) issued guidance statement TPB(GS) 55/2026, explaining how existing obligations under the Code of Professional Conduct apply to the use of AI.

The guidance does not create new obligations, but it reinforces the importance of competence, reasonable care, confidentiality and record keeping when technology is used to deliver tax agent services.

“So the conversation inside firms has shifted,” McPhedran says.

“It used to be ‘can AI do this?’ Now it’s ‘can I rely on it, where does my client’s data go, and how do I show I’ve reviewed it?’”

The benefits are becoming easier to identify, particularly in structured, time-consuming work.

Tax research and first-draft advice are among the clearest opportunities. AI can help practitioners organise information and identify relevant sources more quickly, while document-heavy tasks such as reviewing trust deeds or preparing workpapers can become more manageable.

McPhedran also sees benefits for developing accountants, who can learn from the reasoning and source material presented alongside an answer.

But he cautions against assuming every tool will deliver the same results.

“Can you verify the output? If a tool gives you a position without pointing to the legislation, ruling or case it relies on, your review turns into a redo and the time saving disappears,” he says.

His advice is to test new tools against a selection of previous matters where the correct answer is already known. It provides a practical way to assess accuracy before introducing AI into client work.

Building safeguards into everyday practice

As AI becomes more embedded in accounting workflows, McPhedran identifies three principal risks: inaccurate outputs, confidentiality breaches and over-reliance on technology.

AI-generated information can appear authoritative while containing incorrect references, outdated legislation or fabricated citations. Practitioners therefore need to verify outputs against primary sources rather than relying on the apparent quality of the writing.

Confidentiality presents another challenge. Under the tax practitioner Code of Professional Conduct, entering client information into an AI tool may constitute disclosure to a third party, depending on how the tool is configured and used. Where it does, client permission is required.

Firms must also consider their obligations under privacy legislation and the rules governing tax file numbers.

“The third [risk] I think is over-reliance: using AI instead of your own analysis of the client’s circumstances, when it should only be used as an aid or tool,” McPhedran says.

He believes the starting point for responsible adoption should be the quality management systems firms already have in place.

For practices subject to APES 320, this means incorporating AI into existing processes for assessing technology providers, managing risks, reviewing work and maintaining appropriate records.

“The biggest misconception is that AI needs a whole new rulebook. It doesn’t,” he says.

“A firm already running a sound APES 320 system is most of the way there.”

That does not mean adopting a tool without scrutiny. McPhedran recommends a written assessment covering its intended purpose, accuracy, limitations, data handling arrangements and the level of human review required.

Firms should understand where client information is processed and stored, whether it is used to train AI models, who can access it and how long it is retained.

McPhedran also warns against overlooking AI functions embedded in existing software or allowing staff to use personal accounts with client information.

“Banning AI outright tends to make that worse, because people just use it quietly,” he says.

A clearer approach is to establish which tools are approved, what they can be used for and how practitioners must document their review of AI-assisted work.

From policy to practice

For McPhedran, responsible AI adoption ultimately comes down to making safeguards workable within the daily operations of an accounting firm.

Elfworks has sought to address these challenges through technology designed for Australian tax practitioners.

According to McPhedran, advice drafted using the platform is supported by references to Australian legislation, ATO rulings and case law, allowing practitioners to examine the underlying sources before relying on an output.

He says client data is processed in Australia, reflecting the importance of understanding how information is handled by technology providers.

“The tension we hear from practitioners is simple: AI clearly saves time, but in tax you can’t use an answer you can’t stand behind,” he says.

“So we’ve tried to design around professional obligations rather than treat them as an afterthought.”

Drawing on Elfworks’ experience working with more than 900 Australian firms, McPhedran says successful adoption is rarely about introducing the most sophisticated technology.

“The common thread is that the firms that adopt AI well treat it as part of their quality system, not as a side project,” he says.

To help practitioners put that approach into action, Elfworks has developed an AI Review Process aligned with TPB(GS) 55/2026 and APES 320.

The resource outlines steps for assessing and approving AI tools, recording client permissions, verifying outputs, documenting reviews and monitoring tools as they change.

Get your free AI review checklist. Contact James McPhedran at james.mcphedran@elfworks.ai