Is this task worth automating? A five-question test.

§ 01 — How do I know if a business task is worth automating with AI?

Run the task through five questions. If it fails two or more, stop, you're about to spend money solving a problem AI isn't well suited to.

1. Is it repetitive and rule-describable? Could you write down the steps clearly enough for a new hire to follow them without asking a question? If the task involves genuine judgement calls that change from case to case, it's a harder build and a worse bet.
2. Does it eat real hours you can count? Not "it feels like a lot" — an actual number, tracked for a week or two. Guesses run high. I've had clients estimate four hours and clock closer to ninety minutes once they timed it.
3. Is the input digital, or easily made digital? Emails, spreadsheets, PDFs and scanned forms all count. A process that lives entirely in someone's head, or depends on a phone call, is a different and usually bigger project.
4. What does an error cost when the automation gets it wrong? A false positive is the automation flagging something as fine when it isn't (or the reverse: rejecting something that was actually fine). Rate that cost honestly. Low-cost errors (a report needs a manual glance) are fine to automate around. High-cost errors (a wrong number goes to a client) need a human checkpoint built in, which changes the economics.
5. Is there a boring, non-AI tool that already solves it? Half the time, the honest answer is a spreadsheet macro, a Zapier rule, or a setting that's already sitting unused in the software you're paying for. This is the question that saves people the most money, and the one AI consultants have the least incentive to ask on your behalf.

This is close to the same test I run in the first few days of an audit — it's just faster and rougher when you do it yourself first.

§ 02 — What does a worked example look like?

Take a hypothetical: a bookkeeper spends six hours a week manually matching supplier invoices to purchase orders, chasing down the ones that don't line up.

Run it through the test. Repetitive and rule-describable — yes, the matching logic is close to identical every time. Real hours — six a week, confirmed by a timesheet, not a guess. Digital input — mostly, the invoices arrive as PDFs or emails. Error cost — moderate: a missed mismatch means a supplier gets paid for something that wasn't delivered, so the build needs a human review step for anything the system isn't confident about, not full autonomy. Boring non-AI tool — worth checking first, since several accounting platforms already do fuzzy invoice matching as a built-in feature nobody turned on.

If it clears all five, the math usually looks like this: six hours a week across roughly 48 working weeks is 288 hours a year. At a fully loaded admin labour cost in the AUD $40–$70 an hour range, that's somewhere between $11,500 and $20,000 a year sunk into a task a machine can do most of. Automating even 70% of that matching (the rest stays manual because it's the genuinely ambiguous cases) frees up close to four hours a week. Against a build cost in the low thousands, the payback period lands well inside a financial year, often inside a quarter, which is why "average payback under 90 days" ends up being a realistic bar rather than a marketing number.

§ 03 — When does this framework not apply?

It breaks down fastest on tasks that look automatable but aren't. Three patterns show up often.

High judgement, low pattern: negotiating a price with a supplier, deciding whether an unhappy client gets a refund, triaging which of ten fires to put out first. These involve weighing context an AI system doesn't have and shouldn't be trusted to invent.

High exception rate: if more than a third of cases don't fit the standard pattern, you're not automating a task, you're automating the easy third of it and still doing the rest by hand, plus now maintaining a system. The exception rate is simply the share of cases that fall outside the normal pattern and need a person to step in. A high one is the single most reliable sign a task looks simpler from the outside than it is.

Low volume: a task that happens twice a month doesn't clear the bar no matter how annoying it is. Building and maintaining an automation has its own ongoing cost, and under a certain frequency a checklist beats a system every time. There isn't a universal cutoff here — it depends on how expensive the task is per occurrence — but if you're doing the arithmetic and the annual hours saved are in the low tens, it's usually not worth it yet.

If your task lands in any of these three, the honest move is to say so and leave it alone. That's most of what an audit is for, in fact.

If you've got a task you're not sure about, send the brief and I'll run it through the same five questions with you.