What bookkeeping tasks should you actually automate with AI?

Every conversation about AI in bookkeeping seems to ask "should we be using this at all?" I think that’s the wrong attitude. The more useful way to look at it is task by task: which parts of the bookkeeping cycle are safe to hand to AI right now, and which ones still need a human making the call. After running this in our own practice for the past year including selling one bookkeeping business and keeping the other deliberately lean, these are the five ‘departments’ we use AI for. The theme is generally the same… automate where you can and then have a human approve everything.

Payables:

Payables reconciliation is one of the most time-consuming parts of bookkeeping, especially for clients with high transaction volumes eg hospitality/ retail clients with hundreds of supplier invoices a month.

What AI can do well:

• Match supplier statements against your general ledger export

• Flag invalid or missing ABNs

• Pick up missing credit notes

• Check TPAR applicability against your rules

What still needs a human:

• Final sign-off before any payment is released
• Judgement calls on anything flagged as "doesn't reconcile"

In practice, once this workflow is properly set up, the error rate is genuinely low - on one of our recent runs of several hundred invoices, only one or two came back incorrect. But "low error rate" isn't "no error rate," which is exactly why the release step should never be automated away.

Receivables:

A receivables review is repetitive by nature - checking what's overdue, whether credit notes have been applied correctly and what housekeeping needs doing before you start chasing anyone.

What AI can do well:

• Produce an aging profile and reconciliation status
• Flag unapplied credit notes and suggest which invoice they should be allocated against
• Summarise what's been tested and what hasn't which matters more than people expect, because it stops you assuming everything's covered when only part of it was checked

What still needs a human:

• Any actual client-facing collections discussions
• Deciding whether a discrepancy is a genuine issue or just timing

Payroll:

Payroll is one of the highest-stakes areas to automate, because getting an award rate wrong has real consequences for the employee and the client. It's also one of the areas with the biggest time payoff, particularly for practices still dealing with handwritten timesheets rather than a system like Deputy or Employment Hero.

What AI can do well:

• Read timesheets (including handwritten ones) and calculate hours
• Apply award rates and classifications, including penalty rates like Saturday loading
• Flag anything unusual eg hours that trigger an overtime question under the applicable award

What still needs a human:

• Manager approval on every payroll run- without exception!!
• Interpreting genuinely ambiguous award or Fair Work questions

This is also the area where "teaching" the system properly pays off the most. Feeding it your client's actual chart of accounts, logic, and the relevant Fair Work award rates rather than leaving it to guess is what separates a workflow that saves hours from one that creates rework.

BAS / Compliance:

Pre-lodgement review is largely a checking exercise - comparing the activity statement, GST audit report, payroll activity and super against the trial balance to catch anything that doesn't reconcile before it goes to the ATO.

What AI can do well:

• Run the reconciliation checks across all the relevant reports at once
• Flag critical items that need review versus items that can be auto-cleared
• Pick up specific risk patterns eg possible duplication or sales that look overstated by the GST component

What still needs a human:

• Final lodgement decision
• Anything the workflow escalates

One thing worth building into this from day one: a rule that catches changes in how something is being paid. For instance, if super starts being paid alongside payroll instead of quarterly, that changes what "normal" looks like for the BAS check and a workflow that isn't told to look for that kind of shift won't catch it on its own.

Month end:

Month-end is really a combination of the other four areas: cash summary, P&L, balance sheet, receivables and payables all reviewed together. It's also the task most likely to get pushed down the priority list when a practice is time-poor, which is exactly the wrong task to not focus on.

What AI can do well:

• Pull together a combined report from the relevant exports
• Flag what's been tested and what's outstanding
• Surface anything unusual for a manager or practice owner to review, rather than having them dig through every file manually

What still needs a human:

• Any commentary or insight that goes to the client
• Deciding what the numbers actually mean for that specific business

Why breaking it into ‘departments’ works

Across all five areas, the same structure applies:

1. Give it a role, not just a prompt. A defined task with clear boundaries works better than a one off question.
2. Give it your standards. Your chart of accounts, your logic, your rules - not generic defaults.
3. Give it an escalation path. Anything material gets flagged for a human, NEVER auto-approved.
4. Build in a review cycle, especially early on. Treat it like bringing on a new staff member - it will miss things at first and it gets more reliable the more consistently it's corrected.

And a warning worth repeating: Spot-check everything until you've actually verified the workflow is doing what you think it's doing.

None of the above is worth doing if it comes at the cost of client data security. A few habits worth having in place before you upload anything:

• Never include tax file numbers, portal credentials, or bank logins
• De-identify first eg use client codes instead of names and check for hidden rows or tabs before uploading a spreadsheet
• Limit what any connected tool can access - a connector should only reach the specific folder it needs
• Check your platform's training settings - on individual plans this is usually off by default only on higher tiers, so it's worth confirming rather than assuming

If you don't have a written AI policy for your practice yet, it's worth doing sooner rather than later - plenty of practice owners have discovered staff were already using AI on client work without anyone having decided the rules. If you don’t have cyber insurance you should also look into this ASAP.

We cover this full topic including the specific skill setups for each of these five areas in our recent training session on building an AI bookkeeping workflow. If you want the walkthrough and the templates rather than piecing it together yourself, the recording is available here: