The AI that got the tax wrong didn’t show its working.

When AI gets a tax question wrong, it rarely says so. The answer can read as confident and complete while resting on a fact nobody supplied, two rules blended into one or an assumption the user never sees. A wrong answer that sounds right is harder to catch than one that sounds unsure.

Since 1 October 2026, the stakes for that kind of error have changed. The Treasury Laws Amendment (Strengthening Accountability for Tax Adviser Misconduct and Other Measures) Act 2026 gives the Tax Practitioners Board new powers and makes a failure to comply with the Code of Professional Conduct a civil penalty in its own right.

The Act does not mention AI. It regulates the service the agent provides, whatever tools helped produce it. For agents using AI, that puts the focus on one feature in particular: whether the tool shows how it reached its answer.

Language models are built to produce a plausible answer. Complex tax scenarios expose the gap between a plausible answer and a correct one in three ways.

The first is missing facts. A tax outcome often turns on details the client hasn’t provided. A model asked for an answer will usually give one anyway, filling the gap with an assumption.

The second is blended rules. Tax law is full of rules that look alike but apply in different circumstances. A model can merge two of them into an answer that matches neither.

The third is invisible assumptions. Even when the conclusion is right, the agent may not be able to see what the model took as given to get there.

In each case, a confident wrong answer and a confident right answer look the same on the screen. The model can’t reliably tell them apart, so the agent has to.

Schedule 1 of the Act amends the Tax Agent Services Act 2009. It commenced on 1 October 2026 and applies to conduct from that date.

The most significant change for day-to-day practice is new section 50-31. A registered tax or BAS agent contravenes it if they “fail to comply with the Code of Professional Conduct”. The maximum civil penalty is 2,500 penalty units for an individual and 50,000 for a body corporate. The Board must investigate before it can apply to the Federal Court for a civil penalty order.

The Act also gives the Board new powers to suspend registration. That includes an interim suspension of up to 90 days, which the Board can extend, where the Board is satisfied on reasonable grounds of serious misconduct and a significant risk to clients or the public interest. Infringement notices now apply to some Code breaches, such as failing to maintain knowledge and skills or to keep proper client records. Several existing civil penalties have also been raised, and the Board can now accept enforceable undertakings. New criminal offences cover providing or advertising tax agent services while unregistered.

The standards themselves are familiar. Code item 7 already required agents to ensure a service “that you provide, or that is provided on your behalf, is provided competently”. Code item 10 requires agents to “take reasonable care to ensure that taxation laws are applied correctly” to a client’s circumstances. The 2026 Act widens the consequences when those standards are not met.

For an agent relying on AI, the risk sits in the parts of an answer that can’t be seen. Four design features bring those parts into view.

Visible assumptions let the agent check what the tool took as given before relying on its conclusion. Code item 9 asks agents to take reasonable care in ascertaining a client’s state of affairs, and an assumption the agent can see is one they can test against the client’s actual circumstances.

Inspectable reasoning lets the agent trace an answer back to the rules that produced it. A blended rule is much easier to spot when each step is laid out.

Calculations that give the same result for the same inputs take guesswork out of the numbers. If a figure changes, the agent can see which input changed it.

Human judgement on interpretive calls keeps those decisions with the person accountable for them. The tool can set out the options and the agent makes the call.

Under the amended Act, these features carry real weight. If the Board investigates, an agent who can show how an answer was reached is in a stronger position to demonstrate the work was done competently and with reasonable care. An agent relying on a tool that hides its working is left defending a conclusion they can’t explain. Understanding how a tool reaches its answers is also the kind of skill Code item 8 asks agents to keep current, and a failure to maintain knowledge and skills can now attract an infringement notice.

The Act leaves the choice of tools to the agent. What it holds the agent to is the quality of the service. Tools that expose their assumptions and show their working make that standard easier to meet and easier to demonstrate.

Knowing that an AI tool should show its working is one thing. Seeing what that looks like in a tax product is another, and that is what Andrew Noble’s Summit session covers. Agents will see how one system sets out its assumptions, its reasoning and its calculations so a practitioner can check each step.

They will also see how practitioners can help test and improve the rule logic themselves. That gives agents a direct role in getting the rules right, before they rely on the answers.

When AI Gets Tax Wrong: Guardrails for Complex Tax Reasoning

Andrew Noble, Founder of LodgeiT Labs, has built his approach around making tax reasoning inspectable, challengeable and fit for professional review. He shows how LodgeiT combines Tax Genii, Logical English, deterministic calculators, reasoning trees and expert adjudication. If your firm is deciding which AI tools to trust with tax work, his session is a chance to see what that standard looks like in practice.

Wednesday 11 November 2026, 12:00pm AEDT

Register for the session

Part of The AI in Practice Summit, 11 and 12 November 2026. Free, online, every session recorded.