Eighteen of the most used AI models were asked 121 questions about pensions, tax, debt and savings. They got the answer wrong 57% of the time, according to research published by Saturn this month.
The number that should worry finance teams is the second one. On complex questions, the kind a client actually rings an adviser about, the error rate climbed to 88%. This is not a story about AI being imperfect. It is a story about where the failures cluster.
What the study actually tested
Saturn put 121 questions to 18 models from ChatGPT, Claude, Gemini and Copilot, repeating each question five times to see whether the answers held steady. The full run came to more than 10,000 queries, which makes this one of the larger public tests of financial accuracy rather than a handful of screenshots.
The questions covered pensions, tax, debt management, mortgages, savings and student loans. These are not trick questions. They are the everyday ground that bookkeepers, advisers and payroll staff cover before lunch.
Where the models broke down
Four failure modes showed up. Models miscalculated. They left out risk warnings that a human adviser is obliged to give. They ignored tax changes that had already been legislated but not yet commenced. And they invented rules that do not exist.
That last one is the dangerous category, because an invented rule reads exactly like a real one. In one reported case a free model answered a pension tax question in a way that would have left a British user facing a bill worth roughly $33,000. The answer was fluent, confident and wrong.

Free models did measurably worse
Paid models were better but not good. They erred in 49% of answers against 63% for the free tiers. A 14 point gap is real, and it maps onto the tiers most people actually use, since the free version is what sits in a browser tab next to the accounting software.
It also complicates the usual advice to simply buy the better model. Half of the paid answers were still wrong.
Why Australian firms should care
The study ran on British rules, so the specific answers do not transfer. The failure modes do. Superannuation caps, HECS HELP thresholds and the timing of ATO measures are exactly the kind of jurisdiction bound detail that produced the errors in the first place, and Australian rules are less represented in training data than British or American ones.
For practices already using AI in client work, this changes where the tool sits rather than whether it is used. Our guide to AI tools for accountants puts the line at drafting and summarising rather than advising, and a 57% error rate on direct questions is the argument for keeping it there.
The safer pattern is to give the model the firm’s own verified material to work from rather than asking it to recall a rule. That is the difference between a general chatbot and a finance automation workflow built on documents the firm controls.
Confidence is the actual problem
None of the models flagged when they were guessing. A wrong answer arrived in the same measured tone as a right one, which is why a separate survey found most chatbot users would act on financial guidance without checking it elsewhere.
The models themselves are improving quickly, and the gap between the frontier tiers and the free ones keeps widening, as our Claude and ChatGPT comparison tracks. What has not improved is the signal a model gives when it does not know. Until that changes, the useful question is not which model is most accurate but which tasks can absorb a wrong answer.
