Accountants and doctors know more about their clients than any other profession. It’s probably safe to say they know things even their clients’ partner or closest friend are unaware of. As accountants, the reality is you’re dealing with sensitive information every moment of your working day.

The rise of AI within the profession, and the way it handles data entered into it, means choosing the right AI tools is a critical decision – one that’s less about technology than about client privacy and confidentiality.

“As CAs, we make our living off dealing with our clients’ very private and sensitive and secure information,” says Kayur Patel FCA, director of GenAI and emerging tech at PwC New Zealand.

“And our clients would not be happy if they knew we were using tools that didn’t have the required safety and security around their sensitive data.”

Investing in professional standards

Patel observes that in upholding professional standards, CAs always need to be seen doing the right thing, which is why he says using the free public versions, or paid consumer versions, of any AI model, be it ChatGPT, Gemini, Claude, or some other option, just won’t cut it for accounting professionals.

The reason these consumer public models don’t meet professional standards is they use anything entered into them to further train their models. Any sensitive client data put into a consumer public model, then, isn’t so private and confidential anymore, and there’s a risk this data could be served up as a response to another user.

“You need to be using the enterprise or business versions of these models,” says Patel. “They have security and privacy built in, and your data won’t be used to train their models.”

Yes, there are costs associated with using enterprise AI models, but CAs need to ask themselves what price they’re willing to put on client confidentiality. As a CA using AI, the cost of a business AI subscription is a price worth paying to know you’re upholding the professional standards demanded by your clients and peers.

How to choose the right AI platform for you and your practice

There are lots of generative AI tools out there, but most of them are based around technology from the ‘big three’ – OpenAI’s ChatGPT, Google’s Gemini and Anthropic’s Claude. So, how do you choose one and is there much difference between them?

Dr Ben Swift is an AI researcher at the Australian National University’s School of Cybernetics. He contends that, for most tasks, the big models are largely interchangeable and it’s impossible to say one is better than another.

“They’re all updating their models every few months and constantly leapfrogging each other,” he says. “So, that might mean a certain model is better for a certain task at a particular point in time, but even that will change.”

Patel agrees, saying the worst thing any CA can do is “spin their wheels” trying to choose which platform to go with. “If you do that, you risk being left behind,” he says. “Just pick anything, so long as it’s a safe model.”

AI outcomes depend on the user

The real differentiator isn’t the platform you choose, but the user and how they interact with the technology, observes Tristan Tan CA, commercial associate director at IFM Investors.

“It’s not the model that defines the output, because they’re all really good,” he says. “It’s more about how well you give it the instructions and context needed to have it generate the right answer.”

Swift says the variation in responses given by a model to a particular prompt are simply the way the technology works. Ask an AI tool a question in the morning and it will provide a certain answer. Ask the same question later that afternoon and the answer will be different – if not in overall content, in the tone.

What’s certain is all the models from the big three will largely give the same substantive response to an identical prompt, even if the actual text outputs differ in minor ways.

“At the top level, regardless of what model you use, you’re going to get nice quality outputs if you give it the right context,” says Tan.

However, the real skill is in having the necessary domain expertise to judge whether a response is accurate and factual. AI models are known to make things up, so any response should be evaluated with your professional knowledge and expertise, and not taken at face value.

“After that, which model you choose simply comes down to personal preference,” Tan adds.