AI for bookkeepers is not about letting a chatbot post your journals. The useful part sits around the books: the client emails, the chasing, the checklists, the "what changed this month?" conversation.
The big accounting vendors cover their own features well. What almost nobody teaches is how to use a general assistant like Claude or ChatGPT on your own routines, safely, with Australian obligations kept in view. Start there.
What AI should and should not do in bookkeeping
- AI sorts and drafts. You decide.
- AI groups the month's unanswered queries by client. You decide which ones block the reconciliation.
- AI explains a variance in plain English. You confirm it against the ledger first.
- AI drafts a BAS-prep checklist or chase email. You review the figures and lodge nothing without your own check.
- AI never sees client data you haven't been cleared to share. That is a permission question before it is a tool question.
On the responsibility side, the Tax Practitioners Board published guidance on AI for registered tax agents and BAS agents on 22/07/2026. Its position is that practitioners remain accountable for the accuracy of what they provide, and that you need client permission before putting client information into an AI tool, depending on how the tool is set up. Read it yourself: TPB(GS) 55/2026. If you are not a registered agent, the same habits are still sensible. This page is general information, not tax or legal advice.
Workflow 1: month-end client queries
Every bookkeeper has the same pile: forty bank lines with no receipt, five clients who haven't replied.
Export the list of uncoded or queried transactions with descriptions and amounts only. Remove names and account numbers you don't need.
Then have AI:
- group the queries by client and by type (missing receipt, unclear supplier, personal or business?)
- draft one chase email per client in your usual tone
- list what is still outstanding after a week
You check each draft, add the context only you know, and send it. The assistant has saved the typing. It has not decided anything.
Workflow 2: coding questions with a paper trail
When you are unsure how to treat something, AI is a decent thinking partner and a poor authority.
Describe the situation without identifying the client, ask for the questions you should be asking and the usual options, then verify the answer against the ATO or your own knowledge. Record your reasoning in the client notes.
Never let a confident paragraph become the coding decision. If you can't point to a source, it isn't verified.
Workflow 3: BAS prep support, not BAS preparation
Use AI to build the checklist around the BAS, not to produce the numbers you lodge.
- a per-client readiness checklist (bank feeds reconciled, payroll finalised, super and PAYG items reviewed)
- a plain-language summary of what changed since last quarter, drafted from figures you paste in
- a short note to the client explaining any unusual GST position
Software vendors are building AI directly into this. MYOB's AI BAS is described by MYOB as in beta, and its AI accounting page shows what is rolling out. Whatever you use, treat the output as a draft for sign-off. A reconciliation the software matched for you is still a reconciliation you are responsible for.
Workflow 4: SOPs and onboarding for new clients
This is the easiest win, because nothing in it touches client money.
Record yourself doing one routine once: onboarding a client into Xero, setting up bank feeds, the end-of-month close. Turn the transcript into a numbered SOP with decision points, common errors and who to call when it breaks. Then follow it yourself, and fix every step it got wrong.
A tested SOP lets a new team member or contractor work without asking you the same twelve questions.
Workflow 5: reporting that explains, not just shows
Clients rarely read a P&L. They read three sentences about it.
Paste the figures (not the client's identity) and ask AI to draft a short note on what moved, what to watch and one question for the next call. You rewrite the interpretation. If cash is tight, the client needs your judgement about why, not a tidy paragraph from a model that wasn't there.
One Wright Mode member posted this in the community: "OMG I have built my first automation using N8N (had never heard of it before here!) pulling in data from Karbon and Xero so I can see exactly what jobs and clients are taking the most time and who is the least profitable." That is one member's post, not a typical result. It does show where this goes once the manual routines are solid: your own practice data, pulled together so you can price and prioritise.
The small stack worth learning
- Xero, MYOB or QuickBooks Online: keep doing the books where they live. Xero's JAX is its AI finance assistant, and Xero says it only acts when you open it and ask, within your user permissions.
- Dext or Hubdoc: receipt and bill capture, if you already use them.
- Karbon or your practice manager: job tracking, which is also the raw data for the profitability question above.
- Claude or ChatGPT: drafting, summarising and SOP work, on an approved business account.
- n8n: later, to move data between the others once a manual routine is boring and reliable.
Where client data is allowed to go
Before any client information enters an AI tool, settle three things: has the client agreed (your engagement letter can cover it), where is the data stored, and is training switched off on that plan. My post on whether ChatGPT trains on your chats covers the settings to check. Where you are unsure, use de-identified figures and keep names, ABNs, TFNs and bank details out.
Your first build: the month-end chase
Take one client's list of open queries this month. De-identify it, run the grouping and drafting step, check every line, send. Do it three times for different clients before you change anything.
Then work out which jobs are worth automating using the what to automate first scoring test, and see the wiring in four business automations you can build this week. If you want a community that builds these alongside you, see Wright Mode.