Where to draw the line: AI assistance versus professional judgement
AI can draft, triage and summarise, but it cannot carry your TPB obligations. Here is where to draw the line in an Australian accounting practice.
Every practice principal has felt the pull. A client email lands at 4pm on a Friday, the inbox is forty deep, and a tool that can draft a sensible reply in seconds looks like the answer to a recurring problem. AI assistance is genuinely useful inside an accounting or bookkeeping practice. But there is a line between letting software help you work faster and letting it make decisions that only a registered agent can stand behind. Knowing where that line sits is now part of running a practice well.
This is not a hype piece. It is a grounded view on what AI should and should not decide inside an Australian practice, and how to keep your review discipline, client confidentiality and accountability intact while still getting the speed benefit.
What AI is good at, and what it is not
The honest framing is that AI is a drafting and summarising assistant, not a decision-maker. The features that earn their place are the ones that produce a first draft a human then owns. Inside Finye, that means AI ticket summaries that compress a long thread before you read it, AI triage that suggests a priority, assignee or category for an incoming request, AI reply and compose to get a draft started, and AI draft that turns a one-line brief into a structured work item. Each of these saves keystrokes and reduces the cognitive load of a busy day.
What none of them do is finalise a position. AI does not decide whether a GST input tax credit is claimable, whether a payment to a shareholder triggers Div 7A, or whether a client is better served as a company or a trust. Those are professional judgements that rest on facts, law and your read of the client's circumstances. A model produces plausible text. Plausible is not the same as correct, and the gap between the two is exactly where your value as an agent lives.
Treat AI output as a draft from a capable but junior colleague. You would never lodge a return a junior prepared without reviewing it. The same rule applies to anything a model generates.
Review discipline is the whole game
If AI changes one thing about how you work, it should be that you spend less time on the blank page and more time on review. That only helps if the review actually happens. The risk with fast, fluent output is automation bias: text that reads confidently invites a lighter check than messy human notes would.
A few habits keep the discipline honest:
- Always have a named human owner. Every AI-assisted reply, summary or work item should be reviewed and approved by an identified staff member before it reaches a client or the ATO. The model assists; a person signs off.
- Verify the facts, not the fluency. Check figures, dates, entity names and any cited treatment against source records. Do not let a well-written paragraph stand in for a checked one.
- Use structure to force the checkpoint. Running client work as work items on a board, with status workflows and assignees, means an AI-drafted item still has to move through a review status before it is sent or actioned. The workflow becomes the safety net.
- Keep an audit trail. Comments and activity history on a work item record who reviewed what and when. If a position is ever questioned, you can show the human decision behind it.
The point is not to slow down. It is to make sure the time AI saves on drafting is partly reinvested in the review that protects your clients and your registration.
Client confidentiality and what you feed the machine
Accounting practices hold some of the most sensitive data a person or business has: tax file numbers, bank details, financials, personal circumstances. Confidentiality is not optional, and it does not pause because a tool is convenient.
Before any AI feature touches client information, you should know where the data goes and how it is handled. In Finye, AI features run through a metered credit wallet, and access sits inside the same tenant boundaries and permissions as the rest of your data, so the same role controls that govern who can see a client govern who can run AI on that client's information. That matters because confidentiality breaches rarely come from a dramatic hack; they come from data flowing somewhere it should not, with no one having decided that it could.
Practical guidance: be deliberate about what you paste into any general-purpose AI tool outside your practice systems. Strip identifiers where you can, prefer features that operate inside your client management platform under your own access controls, and make sure your engagement terms and privacy practices reflect how you actually use AI. Clients are entitled to know, in plain terms, that you use these tools and that their data is protected when you do.
Accountability under your TPB obligations
This is the part that cannot be delegated to software. As a registered tax or BAS agent, you carry obligations under the Tax Agent Services Act and the Code of Professional Conduct administered by the Tax Practitioners Board. You must act honestly and with integrity, act lawfully in your client's best interests, take reasonable care to ascertain a client's state of affairs, and ensure that tax agent services you provide are competently performed.
No tool changes who is responsible for meeting those obligations. If an AI-assisted reply gives a client wrong advice, the answer to "who is accountable" is you, not the model. The Code's reasonable-care standard assumes a competent practitioner applied judgement to the client's actual circumstances. "The AI suggested it" is not a defence, and the recent additions to the Code around competence and keeping clients properly informed only sharpen that expectation.
So the line is clear in principle: AI can help you gather, draft and summarise. You ascertain the facts, you apply the law, you form the position, and you take responsibility for it. Build your processes so that the human accountability point is never ambiguous, every position has a named preparer and reviewer, and the AI sits clearly on the assistance side of the line.
Drawing the line in your own practice
A simple test helps when you are unsure. Ask: if this output were wrong, who would the TPB hold responsible? If the answer is a person in your practice, then a person must own and review it, no matter how good the draft looks. AI is a force multiplier for capable practitioners, not a substitute for the judgement that makes you one.
Finye is built on that principle. The AI features speed up the drafting and triage work, while the boards, review workflows, permissions and audit trails keep a human firmly in charge of every decision that matters. If you want to give your team faster drafting without giving up the review discipline your obligations demand, start a free trial or see the plans to find the right fit for your practice.