HY Accounting — AI & Tax
AI is pretty good at tax questions.
Until it isn’t.
Here’s what it gets right, where it quietly falls apart, and why the gap between those two things matters more than most people realise.
Let’s be honest about this from the start. We use AI in this practice. Every day. It saves time, it surfaces things faster, and it’s genuinely useful for a wide range of tasks.
And if you’ve been using ChatGPT or Claude or Gemini to answer tax questions, we’re not going to tell you to stop. For a lot of things, it’s fine.
Understanding what a term means. Getting a general sense of how something works. Figuring out what questions to ask before a meeting. All legitimate. All reasonably reliable.
AI is good at giving you the shape of an answer. General concepts, broad frameworks, standard definitions. If your situation fits neatly inside the general rule, you’ll probably get something useful.
The problem is the other 30%.
Tax is a domain where the specific facts of your situation determine the outcome. Not the general rule. Not the broad principle.
The actual facts — the entity type, the timing, the purpose of the transaction, the history behind it.
When you ask AI a tax question, it answers the question you typed. It doesn’t know what it doesn’t know about your situation. It can’t ask the follow-up questions that change the answer. And because the response sounds complete and confident, it’s easy to assume it is.
That 30% gap isn’t random. It shows up in the same places, the same ways, every time.
Where it breaks down
The specific failure modes.
Jurisdiction drift
Most AI models were trained predominantly on US content. Australian tax law — Division 7A, franking credits, superannuation, trust rules — has no real equivalent in most other countries. The model draws on both without telling you which one it’s using.
Recency gaps
Tax law changes constantly. The 2026-27 Federal Budget introduced material changes to CGT, negative gearing, and trust rules. AI tools have training cut-offs. They’ll answer confidently about rules that may no longer apply the same way.
Fact pattern sensitivity
Small details change the answer entirely. The purpose of an expense. The timing of a decision. Whether an entity has a prior year loss. AI answers the general description you gave it. It has no way of knowing which detail it’s missing.
Confident wrong numbers
Ask AI for a specific threshold or rate and it will give you a number with complete confidence. Sometimes it’s right. Sometimes it’s slightly off. Sometimes it’s made up. The tone is identical in all three cases.
70% right still produces a wrong answer
when the 30% is the part that applies to you.
That’s not a caveat. That’s the whole problem.
The prompt problem nobody talks about.
The quality of what AI gives you depends heavily on how well you ask. Someone who knows the right questions, who knows when to push back, who understands enough about the topic to interrogate the answer — they get materially different output.
Most people asking tax questions don’t have that context. They type a question, get a confident response, and have no way of knowing whether they’ve just received solid general guidance or a jurisdiction-drifted, outdated, fact-insensitive answer presented in the same authoritative tone.
The output looks the same either way. That’s the part that’s genuinely dangerous.
How we use it
AI produces the direction. Legislation confirms it.
In our practice, technical questions don’t get answered from general AI output alone. We use legislative research tools — including Law Cyborg, which is built specifically around Australian law — to test the answer against the actual source. That’s not the same as asking ChatGPT. It’s anchored to the legislation as it currently exists, applied to the specific facts in front of us.
The offer
Get It Reviewed.
If you’ve used AI to answer a tax question — and you’re making a real decision based on that answer — it’s worth having someone check it against the actual facts of your situation.
Not because AI is always wrong. Because it’s right enough, often enough, that people act on it without knowing which category they’re in.
That’s the conversation. Bring us the output. We’ll tell you whether it holds up.
📋
You bring the output
Whatever AI told you. Screenshot, copy-paste, the actual question you asked.
🔍
We test it properly
Against your actual facts. Against current Australian law. Against the questions AI didn’t think to ask.
✓
You get a clear answer
It holds up, it doesn’t, or here’s what’s missing. No ambiguity.
This is for you if:
- AI told you something was deductible and you’re not 100% sure
- You asked about a capital gains event and got an answer you’re planning to act on
- You used AI to figure out how to structure something — a business, a property purchase, a loan
- You got an answer about trust distributions, director loans, or super contributions
- The answer felt right but something in the back of your mind isn’t settled
Get It Reviewed
Not sure if your AI answer holds up?
Let’s find out.
Book a short call. Bring what you’ve got. We’ll tell you whether it’s solid, where it’s shaky, and what to do about it.
Or email us directly at enquiries@hyaccounting.com.au
HY Accounting. This article is general in nature and does not constitute tax advice. Your specific circumstances should be assessed by a qualified practitioner before any decision is made. H Youssef Accounting & Taxation Services Pty Ltd. Liability limited by a scheme approved under Professional Standards Legislation.