AI for accountants and bookkeepers: what to use it for, and what to keep.

How are accountants using AI in Australia?

AI for accountants is mostly about the work around the advice: reviewing files, preparing working papers, chasing and sorting client documents, onboarding, and drafting letters and meeting notes. The practices getting real value use it to prepare and check, with a qualified person reviewing every output before it reaches a client or the ATO.

We see accounting firms, bookkeepers, fractional CFOs and financial planners on our sales calls and in our workshops every month. Almost none of them want AI to "do the tax". What they want is the hours back from the relay of gathering, sorting, drafting and checking that sits between a client sending something in and a partner signing it off.

The pattern is the same in most practices. Someone is already using ChatGPT or Claude on their own, usually on a personal account, and the owner has no policy covering it. The first job is getting that under control. The second is picking two or three jobs where AI earns its keep.

What AI tools do accountants and bookkeepers actually use?

Most practices end up with three layers: a business account for a general assistant like Claude, ChatGPT or Copilot; the AI already inside their accounting software, such as Xero's JAX; and, for the keen ones, small tools built for their own workflow, like an onboarding form or a document checker.

The general assistant is where most of the value starts, because it touches every job. Which one you pick matters less than whether it is on a business plan that keeps client data out of model training and gives the owner control. If you run on Xero, our guide to connecting Claude to Xero covers what a live connection can and cannot do.

Can AI do bank reconciliation for bookkeepers?

Partly, and it is getting better. Xero's JAX can now reconcile bank lines automatically, but only when it is highly confident of the match, and you can review and reverse anything it did. The realistic goal is a bookkeeper who handles the exceptions, not one who has gone.

Xero says JAX uses matches, your existing rules, memory of past transactions and prediction, and that it is switched on per bank account by the subscriber or advisor. Xero's own page describes plan availability in more than one way, so confirm what your plan includes before you count on it.

Reconciliation is the obvious one. The bigger wins we see are in the work around it: reading supplier statements, matching remittances, chasing the missing receipt, and writing the month-end note to the client.

Can AI help with client onboarding?

Yes, and it is one of the best first projects for a practice. A simple online form collects a new client's details and documents once, in the right format, and files them where your team expects. AI then helps prepare the first pass of the work from what came in.

One West Perth tax advisor came to our workshop wanting to stop loading files into Claude by hand. Since then the advisor has built a client onboarding form that creates a PDF summary of each submission, and three AI agents that split the job the way the team does: one gathers a client's information, one starts the draft, one reviews everything. The form runs on the firm's own server, under its IT manager's control, and is still being tested with dummy data before real clients see it. The full write up is in the AustAsia Group success story.

The detail that matters for accountants: the advisor chose to keep submissions inside the firm's own environment, so they follow the same storage, access and retention rules as every other client file. That decision belongs with the firm, and it is the right instinct when ID documents and TFNs are involved.

How much time can AI save on advice preparation?

For document heavy advice work, a lot. A financial planner we spoke with cut advice preparation from two weeks to two days without outside help, using a meeting note-taker and Claude. The time goes on drafting and assembling. The judgement still takes as long as it should.

If meeting notes are your starting point, our comparison of AI note-takers like Fathom, Granola and Otter covers which one suits a client-facing practice and how recording consent works. For practices drowning in PDFs, statements and receipts, AI document processing is usually the next build.

Can accountants put client data into ChatGPT or Claude?

Only with the client's permission and on a properly configured tool. The Tax Practitioners Board says entering client information into an AI tool can count as disclosing it to a third party, depending on how the tool is set up, so tax and BAS agents need client permission first.

The TPB set this out in TPB(GS) 55/2026, issued on 22 July 2026. The points that matter for a small practice:

• Get permission first. Under Code item 6, you must not disclose a client's information to a third party without their permission. The TPB recommends telling clients where their data will go, where it will be stored and whether AI tools may be used. A signed engagement letter or a signed consent can cover it.
• You stay responsible. The TPB is clear that practitioners remain ultimately responsible for their services, that AI can produce inaccurate information, and that outputs should be verified at each step and the checking documented.
• Check the tool itself. The TPB expects due diligence on commercial and in-house AI tools, so that information stays secure and the Privacy Act is met.
• TFNs have their own rule. The Privacy (Tax File Number) Rule covers tax agents and accountants who hold TFN information, and requires reasonable steps against unauthorised access.

The privacy regulator goes further. As a matter of best practice, the OAIC recommends not entering personal information, and particularly sensitive information, into publicly available generative AI tools. In plain terms: no client files in a free or personal ChatGPT account, ever. Use a business plan with training switched off, give the owner control of the accounts, and write a one-page AI policy before anyone connects email or Xero. How we handle data on every build and workshop is on our security page.

What should accountants not automate with AI?

Anything where a mistake reaches a client or the ATO without a qualified person reading it first. That means lodgements, advice, tax positions and anything signed in your name. Use AI to prepare and check those, never to finish them.

• Final sign-off. Lodging, advising and signing stay with a person. An AI tool is not a registered agent, and you carry the responsibility for what it produced.
• Judgement calls. AI is weak at your client's specific circumstances unless you give it all of them, and it will still sound sure of itself.
• Numbers you have not trained it on. In one onboarding session we ran this year, Claude priced a construction project at roughly half the market rate. Out of the box, AI does not know your rates, your clients or your history.
• Anything with no record. If the TPB asked how a figure was reached, could you show them? If not, the process is not ready to automate.

Where should a small practice start?

Start with accounts and a policy, then one job. Move everyone onto a business AI account, update your engagement letter to cover AI use, write a short AI policy, then pick the single most repetitive job in the practice and get it working end to end before adding a second.

For most practices that first job is onboarding, file review or month-end client notes. Once one job runs reliably and the team trusts it, connecting more of your systems through AI automation becomes a sensible next step rather than a leap. Owners who want to build it themselves can do it in our two-day AI Sesh workshops in Perth, Sydney and Melbourne, on their own client work.