Key takeaways
What this article covers, in order:
- A Tuesday morning with 63 transactions
- What AI tends to get right
- Where you still need to check
- A worked example of a "wrong but reasonable" suggestion
- A checking routine that takes ten minutes
- Should you trust the AI more over time?
AI categorisation reads each bank transaction and suggests which account it belongs in, so you confirm instead of typing. It's very good at the repetitive stuff, like the same fuel station, software subscription or supplier every month. It's weakest where the bank line alone doesn't tell the full story, so you still need a quick human check on a handful of transaction types.
A Tuesday morning with 63 transactions
Liam has a landscaping business in Ipswich. Ten days without opening his books and there are 63 new transactions waiting. Before AI suggestions, that meant 63 dropdowns. Now most of them arrive with a category already filled in: fuel to Motor vehicle, the nursery to Materials, his phone plan to Phone and internet.
He goes down the list. 54 look right, so he confirms them. 9 need a second look. That's the realistic shape of AI categorisation for most small businesses: it shifts your job from data entry to review. (Your own split will differ. Liam's numbers are illustrative, not a benchmark.)
What AI tends to get right
AI is good at patterns, and most small-business spending is patterns.
- Recurring suppliers. The same merchant name for the same amount every month is the easiest case there is.
- Obvious merchants. Fuel stations, telcos, software companies and insurers usually have descriptions that say what they are.
- Your own history. Once you've categorised a supplier a few times, a good system learns how you code it, even if your choice differs from what someone else would pick.
- Bank charges and interest. These have predictable, bank-generated descriptions.
Where you still need to check
Here's the part that matters. These are the transaction types where an AI suggestion should never be confirmed on autopilot.
| Transaction type | Why AI struggles | What to check |
|---|---|---|
| Mixed business and personal purchases | The bank line shows one total, not what was in the trolley | Split the personal part out |
| Big-box stores | One merchant sells everything from screws to barbecues | What you actually bought |
| Transfers between your accounts | Can look like a payment to a person | That it's coded as a transfer, not an expense or income |
| Customer payments | Might be suggested as income when an invoice is already open | Match to the invoice instead |
| Owner drawings | A transfer to "J Smith" could be a wage, a contractor or you | Who it really went to |
| Large one-off purchases | Could be equipment (an asset) rather than an expense | Whether it belongs on the balance sheet |
| Refunds | Money in from a supplier isn't sales income | Code it against the original expense |
| GST treatment | Same merchant can sell taxable and GST-free items | The GST code on the line |
Look at that list and you'll notice something: none of these are hard because the AI is bad. They're hard because the information simply isn't in the bank description. A human looking at the same line would have the same problem without the receipt or the context.
A worked example of a "wrong but reasonable" suggestion
Liam's line reads: "BUNNINGS 7421 ... $384.60". The AI suggests Materials, which is what he usually codes Bunnings to. Reasonable. But this trip included $298.00 of pavers for a client job and $86.60 for a new letterbox at his own house. The correct entry is $298.00 to Materials and $86.60 to Owner drawings. Check: $298.00 + $86.60 = $384.60.
If he'd hit confirm without looking, his expenses would be overstated by $86.60. Small on its own. Repeated across a year of Saturday trips, it adds up and your accountant has to unwind it.
A checking routine that takes ten minutes
You don't need to re-examine every line. Focus your attention where the risk is.
- [ ] Confirm the routine ones (regular suppliers, fees, subscriptions) in a batch
- [ ] Open every transaction over a threshold you choose, say $500
- [ ] Open every big-box and supermarket transaction
- [ ] Open every transfer to a person's name
- [ ] Match money in to open invoices before accepting any "income" suggestion
- [ ] Scan the GST codes on anything that isn't routine
- [ ] Correct a wrong suggestion instead of skipping it, so the system can learn your preference
That last one is easy to skip. If Liam leaves the letterbox in Materials because "it's only $86", the pattern stays wrong next time too.
Should you trust the AI more over time?
Somewhat. The more transactions you've confirmed and corrected, the better the suggestions usually get for your recurring suppliers. What doesn't change is the list in the table above. A mixed purchase will always need you, because the information isn't on the bank line.
A good rule of thumb: trust the AI with repeats, check it on anything new, large, mixed or personal.
What happens if you stop checking?
Books still balance. That's the dangerous part: nothing breaks visibly. Profit drifts, personal spending sits in expenses, a transfer double-counts as income, and you find out when your accountant sends a long list of questions near 30 Jun 2027. If your books have already drifted, it's worth a once-over with your bookkeeper before you build on top of them.
How HelloBooks helps
On the HelloBooks Pro plan (A$30 a month), AI auto-categorisation suggests a category for transactions as they arrive from your bank feeds or CSV imports. They land in a review list where you confirm or change each one, so you stay in control of what goes into the books. The Free plan includes free AI credits to get started. When AI credits run out, AI categorisation pauses and your books keep working; you just categorise by hand. Pro also adds AI Analysis on every report. See our AI bookkeeping software page and plan details.
FAQs
Does AI categorisation mean I don't need a bookkeeper?
Not necessarily. It cuts the data entry, but someone still needs to review exceptions, reconcile and keep the books sensible. Many owners find their bookkeeper's time shifts toward review and advice.
Can AI categorise GST correctly?
It can suggest a GST code, but check anything that isn't routine, especially merchants that sell both taxable and GST-free items. Your BAS agent or accountant handles lodgement; your job is to get the codes right.
What if the AI keeps getting the same supplier wrong?
Correct it each time rather than skipping. If your software learns from your choices, the suggestion should improve. If a supplier genuinely sells you different things, you'll always need to check it.
Is my data used safely?
Bank feeds are read-only and you approve access through your bank. Ask any provider how they store and process your data before you connect.
Do CSV imports get reviewed the same way as feed transactions?
In HelloBooks, yes: imported CSV transactions land in the same review list as feed transactions, where you confirm or change each category before it's final.
Let the AI do the repeats, keep your eyes on the odd ones, and your Tuesday mornings get a lot shorter.
Start free, no card needed. Try HelloBooks Free
