AI is good at reading, drafting, summarising and sorting. It is not yet something to trust, on its own, with decisions that commit money, numbers that must be exactly right with nobody checking, decisions that significantly affect a person, or legal and compliance sign-off. The reason is simple: current AI tools produce answers that look right whether or not they are, and they can't tell you reliably when they've got it wrong. So the safe pattern for a business is "AI drafts, a person approves". AI does the first pass on the paperwork, a named person checks and makes the decision, and the system records who approved what. This guide sets out the four areas to keep AI out of (for now), why, and how to set up the review step so the time saved isn't lost to the risk.
For the jobs where AI does help, department by department, see AI use cases by department.
Why AI needs a person checking it
Today's AI tools, including the large language models behind ChatGPT, Copilot and similar products, work by predicting a likely answer. That's why they're good at drafting an email or reading an invoice laid out in a way they've seen before. It's also why they can be confidently wrong:
- They can produce a figure, a clause or a reference that looks plausible and isn't real.
- The same question can get a slightly different answer each time.
- They don't know your business rules unless you give them, and they may apply them inconsistently.
- They rarely say "I don't know". A low-confidence answer often reads exactly like a high-confidence one.
None of that makes AI useless. It means the output is a draft, and somebody who understands the work has to look at it before it counts. That person is what people mean by "human in the loop" (a person checks and approves before anything happens). Our post on what ChatGPT can and can't do with your own systems covers the technical side of this in more detail.
1. Final financial decisions
Keep AI out of: approving spend, releasing payments, setting or changing credit limits, writing off debts, approving refunds or credit notes, and posting journals without review.
Why: these decisions commit the business's money and are hard to undo. A wrong payment to a supplier, or one changed bank detail accepted without question, can't be pulled back by correcting a prompt. Your approval rules (your delegation of authority) exist so a named person is accountable for each commitment. An AI tool can't hold that authority. See approval matrix and delegation of authority for how those rules work.
What AI can do instead: prepare the decision. For example, read the invoice and suggest the PO match, list the bills due this week with any exceptions flagged, or summarise an overdue account before someone decides on credit. The approver then decides, with the evidence in front of them.
2. Numbers that must be exactly right without anyone checking
Keep AI out of: producing figures that go straight into a BAS, payroll, a tax return, a board report, a customer invoice, a tender price or a bank reconciliation with nobody checking them.
Why: AI tools are not calculators and not a ledger. They can misread a digit, transpose a column, add up wrongly when asked to do arithmetic in a chat, or invent a total that "looks about right". For finance and operations numbers, the rule is that the figure should come from the system that holds the data (the accounts, the payroll system, the stock system), calculated by ordinary software you can audit. That one place everyone takes the number from is what's meant by a single source of truth (glossary).
What AI can do instead: explain the numbers, not produce them. For example, it can turn a variance report into a short plain-English summary, draft the commentary for a month-end pack from figures the reporting system has already calculated, or highlight lines that look unusual for a person to look at. If AI reads figures from a document, such as invoice totals, the totals should be checked against the lines and against the PO before anything is posted. Our three-way matching guide shows how that check works.
3. Automated decisions about people that use their personal information
Keep AI out of (as the final decision-maker): screening or rejecting job applicants, ranking staff performance, setting pay or bonuses, refusing a customer's account or credit application, deciding a claim or hardship request, or suspending a contractor, where the decision uses personal information about that person.
Why: these decisions can significantly affect someone's life, and from 10 December 2026 organisations covered by the Privacy Act must be open about them. The Privacy and Other Legislation Amendment Act 2024 added new parts to Australian Privacy Principle 1 (APP 1.7 to 1.9). According to the OAIC's APP 1.7–1.9 transparency obligation fact sheet, an organisation must explain in its privacy policy where it has arranged for a computer program to make a decision, or to do something substantially and directly related to making one, that could reasonably be expected to significantly affect an individual's rights or interests, and personal information about that individual is used. The policy must describe the kinds of personal information used and the kinds of decisions involved. The OAIC's announcement of the new resources and its updated APP 1 guidelines set out the detail.
Two points matter for anyone using AI here:
- It isn't only about AI. The OAIC fact sheet includes rule-based software and spreadsheets as well as AI.
- A person glancing at the result doesn't automatically take it out of scope. If the program's output is a key factor in the decision, it can still count. Genuine human review means the person has the information, the authority and the time to decide differently.
The rule is about transparency, not a ban, and small businesses under the turnover threshold are generally exempt, with exceptions. Our detailed post on the Privacy Act automated decision-making rules from 10 December 2026 covers who it applies to and how to list your decisions. It's general information, not legal advice; check your position with the OAIC or a privacy lawyer.
What AI can do instead: the paperwork around the decision. For example, check that an applicant's documents are complete, pull expiry dates from a contractor's licences, or summarise a claim file for the person who decides.
4. Legal and compliance sign-off
Keep AI out of: signing off contracts, confirming that you comply with an award, a licence condition, safety law, building codes or a reporting obligation, lodging regulatory returns, or answering "are we allowed to…?" questions that someone will act on.
Why: AI tools can quote legislation or award clauses that are out of date, from the wrong state, or that don't exist. They can't take responsibility, and the responsibility doesn't move to the software vendor because you used their tool. When a regulator, auditor or court asks who approved something, the answer has to be a person with the right knowledge.
What AI can do instead: the reading and the first draft. For example, compare a new supplier contract with your standard terms and list the differences for your lawyer, summarise a long compliance guidance note so the right person knows which parts to read, or pull the figures a return needs from your systems into a draft for checking.
How to set up "AI drafts, person approves"
The review step is what makes AI safe to use. It has to be designed, not assumed. For each AI task:
- Name the approver. One role is responsible for checking and approving the output, and it's in their job description, not "whoever's around".
- Show the evidence. The approver sees the source document next to the AI's draft: the invoice beside the draft bill, the docket beside the timesheet entry. Checking a draft without the original isn't checking.
- Flag low confidence. Set the tool or the integration to mark results it's unsure of (an unreadable field, a supplier it hasn't seen, a total that doesn't match the lines), and send those to the front of the queue.
- Nothing happens without approval. The draft doesn't post, send, pay or update a record until the approver accepts it. Build this into the workflow, so it can't be skipped on a busy day.
- Record who approved what. Store the approver, the time and any changes with the transaction. That's your audit trail (glossary), and it's what you'll need if a decision is questioned.
- Count the corrections. Track how often people change the AI's draft and what they change. If corrections aren't falling, the task, the data or the tool needs fixing.
- Keep the data rules. Decide which information may go into which AI tool. Our AI policy for staff template covers this in a page.
- Review the list of AI tasks. Keep a simple register of where AI is used, who approves and what decisions it touches. The same register helps with the Privacy Act transparency work.
For example, if your accounts payable officer approves AI-read supplier invoices, they see the PDF beside each draft bill, invoices from new suppliers or with mismatched totals are flagged first, nothing posts until they click approve, and each month you look at how many drafts needed changes.
When the rule can relax
Over time, some steps can need less checking. That should be earned with evidence, not assumed:
- the task is low-stakes and easy to reverse (sorting emails into folders, not paying bills),
- you've tracked corrections over months and they're consistently rare,
- the results are still spot-checked, and the audit trail still records what happened.
Even then, the four areas above (money, exact numbers, decisions about people and compliance sign-off) are where we'd keep a person making the final call. Plain rule-based automation is often a better fit than AI for the routine parts anyway; our page on automating routine decisions explains the difference.
Frequently asked questions
What should a business not use AI for?
Don't rely on AI alone for final financial decisions, numbers that go into tax, payroll, invoices or reports unchecked, decisions about people that use their personal information, or legal and compliance sign-off. Use it to read, draft, summarise and sort, with a person approving the result.
What does "human in the loop" mean for AI in a business?
It means a person checks and approves the AI's output before anything happens: before a bill posts, an email is sent or a record changes. The person needs the source document, the authority to change the answer and enough time to check properly. A rubber stamp isn't a human in the loop.
Can AI code invoices accurately enough to trust?
It can do a good first pass, especially for regular suppliers with consistent past coding. It shouldn't post bills unchecked. Have AI suggest the coding and the PO match, flag the ones it's unsure about, and have a person approve each bill. Track how often the suggestions are changed.
Does the Privacy Act ban AI decisions about people?
No. From 10 December 2026, organisations covered by the Privacy Act must explain in their privacy policy where computer programs make, or substantially help make, decisions that could significantly affect individuals using their personal information. It's a transparency rule. The OAIC's fact sheet is the place to start, and a privacy lawyer can confirm your position.
If a person approves every AI decision, are we safe?
You're much safer, but only if the review is real. The person needs to see the evidence, be able to decide differently and actually do so when the AI is wrong. For the Privacy Act, the OAIC says a person's involvement doesn't automatically take a decision out of scope if the program's output is a key factor.
Where is AI lowest risk to start with?
Reading documents such as supplier invoices and delivery dockets, drafting routine emails, and summarising long reports or threads, with a person checking each result. Our AI fit check helps you test a task before you start, and AI document processing covers the reading jobs.
If you want AI to take paperwork off your team without handing it decisions it shouldn't make, book a free initial chat. We'll help you pick the task, design the review step and connect it to the systems you already use. Our AI consultant for small business page explains how we work.