What to Automate in Your Finance Function (and What to Keep Human)

Scroll LinkedIn at the moment and you'll see a wave of posts that go something like this - "I built my own accounting system with AI. Here are the prompts to automate your finance function." I understand the appeal, but it's a trap, and I'm seeing more business owners walk into it

General AI tools are fast, cheap and capable. Ask one to categorise a month of transactions or build a cashflow model and you'll get something that looks right in about 30 seconds

Yet looking right is the problem

A finance function is more than reports and reconciliations

Reports and bank reconciliations are the visible part of the work. Underneath them sit Goods and Services Tax (GST) rules, Single Touch Payroll (STP) reporting, super obligations, award rates, accounting principles like accruals, and record-keeping rules that mean most business records have to be kept for five years under Australian Taxation Office (ATO) requirements. None of that shows up in a prompt

If you don't know those rules exist, AI won't tell you what you've missed. It'll give you a confident answer and move on. You find out when the Business Activity Statement (BAS) is wrong, a super payment lands late, or your accountant asks a question nobody can answer

A homemade system also breaks the moment the person who built it leaves, goes on holiday or forgets how it works. There's no audit trail, no lock dates and no approval step unless the builder knew to add them

What to automate in your finance function first

The better bet is the AI already built into the systems you use. In tools like Xero, people who understand the work have done the thinking for you - the tax codes, controls and audit trail. When the rules change, they update the software. Payday Super is a good example. From 1 July 2026 employers have to pay super with wages rather than quarterly, and the payroll platforms did the heavy lifting to make that possible. A prompt you wrote in March wouldn't have known

Start with the work that's high volume, repetitive and easy to check -

• bank feeds and coding for the transactions that look the same every month
• receipt and bill capture
• chasing missing paperwork and overdue invoices, which nobody enjoys doing by hand
• first drafts of report commentary (a person still edits them)

Gartner's 2026 research on finance teams found these kinds of jobs (data extraction, automating accounts payable and receivable, and report creation) pay back in about nine to 10 months. Forecasting and insight projects take much longer to show value

We build tools at Digit too. They accelerate what's possible with Xero and fill the gaps between systems, so we can automate how people work. None of them replace the ledger. That stays in Xero

It's still not set-and-forget

This is what the LinkedIn posts don't explore. Even the AI inside your accounting software needs someone watching it

More of our clean-up enquiries now come from businesses that switched on auto-reconciliation and let it run with nobody checking. The software can't tell that one payment to a supplier is a loan repayment and the next is stock. It matches, codes and moves on, month after month. By the time anyone notices, the mess runs through the whole file, and cleaning it up takes far longer than doing the work properly would have

Auto-reconciliation is a bit like a new junior who's keen, quick and never asks a question. Brilliant on the easy stuff. Left alone for six months, they'll confidently put the director's loan in cost of sales and you won't find out until year end

As a Xero Platinum Partner, we saw this first-hand when our team tested Xero's auto-reconciliation across client files earlier this year. On simple files it saved real time. On a file with intercompany transfers, it coded transactions to several suspense accounts and had to be undone every week. On our own books, it coded a subscription correctly one month and wrong the next

Automation doesn't fix a problem. It repeats whatever it's been given, faster. If your accounts are clean and your rules are right, AI saves you hours. If they're not, it amplifies the mess

What to keep human

Some tasks shouldn't run without a person signing off. They're the ones where money is hard to get back, legal responsibility can't be handed to software, or a real person gets hurt by a mistake

The test we use is whether a mistake can be undone. If it's reversible and easy to check, let AI run. If it moves money, pays someone or carries your name to the ATO, AI prepares it and a person decides

Shadow AI - pick the tools before your team does

If you don't tell your team which AI tools to use, they'll try whatever's in front of them. A 2026 PagerDuty survey of office workers at large companies across Australia, the UK, the US and Japan found 66% had used AI tools they believed weren't allowed at work. More than a third had entered customer data into public AI tools

The risk with shadow AI is the data. Before anyone uses a tool with business data, think through where the data is kept and processed, whether it's used to train the model, and what people should never paste in (tax file numbers, bank details, payroll records, anything that identifies a client). Business-tier accounts that don't train on your data are the minimum. Free personal accounts shouldn't go near company information

We went through this ourselves and wrote about how we stop AI from handling data it shouldn't

Guardrails from day one

The best place to start is with the practical foundations -

• A short list of approved tools, and a clear line on what data never goes into them
• AI prepares, a person approves. Payments, payroll, lodgements and anything leaving the business. Keep your approval limits exactly where they are
• A call-back rule for bank detail changes, however convincing the email or the voice on the phone sounds
• One owner. Someone has to know what AI is doing in the finance function and check it
• An honest look at your privacy obligations. If your business is covered by the Privacy Act, new rules about automated decision-making start on 10 December 2026

Start small and hand over trust in stages

Pick one painful, repetitive task where a mistake is easy to undo. Bank coding, receipt capture and debtor reminders are good candidates. Payroll and payments are not

Fix the data first. A messy chart of accounts or out-of-date supplier details will be copied faithfully by any automation you add. Then run the new process beside the old one for a month or two and compare the results before you rely on it

Trust gets handed over in steps. First the tool suggests, then it drafts, then it acts with your approval, and only then does it act on its own and report back. Only move up a step when the last month's results back it up. Our guide to automating your bookkeeping without losing control walks through the Xero setup in more detail

One question worth asking this week

If someone else looks after your books, ask them how they use AI on your data. The Tax Practitioners Board's July 2026 guidance makes it clear that registered tax and BAS agents remain responsible for AI-assisted work, and that client information going into AI tools needs the client's permission. A good bookkeeper will have a clear answer. If yours hasn't thought about it yet, ask them to

AI is going to take a lot of the grind out of finance work, and I'm glad about that. I'm less sure we've worked out, as an industry, how much judgement we can safely hand over. For now my answer is not much, and not yet

Current as at September 2026 per the ATO, Fair Work Ombudsman, Office of the Australian Information Commissioner and Tax Practitioners Board