Mac Sweeny
Quick answer
AI is already doing real work in accounting practices: pulling figures off invoices and statements, drafting client emails and letters, summarising long documents, and moving information between systems so nobody rekeys it. Almost everything sold to firms sits in one of four categories: document and data extraction, drafting assistants like ChatGPT and Claude, research and summarising, and workflow automation. Prove one workflow in one category before you roll anything out across the firm. AI Consulting is A$1,195 per month + GST, with a call each fortnight, clarity on what to build in Claude or ChatGPT and practical guidance on how to do it. No lock-in.
Keep the accountant in the review loop
A possible document workflow, with professional judgement kept with the accountant.
- 01
Approved inputs
Choose the source documents and confirm what may be shared.
- 02
AI first pass
Extract or summarise the material and identify items to investigate.
- 03
Accountant checks
Verify against the source and make the final professional judgement.
Do not treat generated text or extracted figures as verified accounting records.
Where accounting firms lose time today
Data entry that never ends
Bank feeds, supplier invoices, receipts, statements. Coding and rekeying is the biggest block of low-judgment time in most practices, and it is the work these tools handle best.
Client email volume
The same twenty questions, rephrased, all year. Each one still needs a correct, considered, firm-branded reply, and each one costs someone ten minutes and a change of gear.
Advisory squeezed out by compliance
The work clients value most is the work that slides to the back of the queue, because compliance has a lodgement date and advice does not.
The four categories worth knowing
Nearly every AI product pitched to accountants sits in one of four groups. Working out which group you are being sold tells you most of what you need to know: what it can realistically do, what it will cost to run, and how badly it can hurt you if it is wrong.
| Category | What it does | What to check before you buy |
|---|---|---|
| Document and data extraction | Reads invoices, receipts, bank statements and financial reports, then pushes structured data into your ledger or workpapers. | Accuracy on your document types, not on the demo file. How corrections are made, and whether the tool gets better because of them. |
| Drafting assistants (ChatGPT, Claude) | General writing and analysis: client emails, letters, plain-English explanations, file notes, first-draft commentary. | Which plan you are on and what the provider says about training on business data. Who reviews output before it leaves the office. |
| Research and summarising | Condenses long documents, contracts, trust deeds and reports into a briefing you can read in two minutes. | Whether it points you to the passage it drew from. Anything it cannot cite in the document, treat as unverified. |
| Workflow automation | Moves data and triggers steps between practice software, email, e-signing and job systems so nothing gets copied twice. | Error handling and alerting. An automation that fails quietly is worse than the manual step it replaced. |
Notice what is missing from that list. Nothing there gives an opinion, takes a position, or carries professional liability. That boundary is the whole story of AI in accounting, and it is why the sensible question is not which tasks AI takes over but which tasks it hands back to you in a more finished state.
Where to start, if you want the easy win
Start with drafting, because it has no integration risk and you can stop at any time. Give the people who write the most client correspondence a paid business plan, pick three jobs they do every week, and have them run those jobs through the tool for a fortnight with the output reviewed as usual. You learn two things: how much time it actually saves in your practice, and who on your team has the judgment to check what comes back.
Extraction comes second, because it touches the ledger and has to be measured. Run it in parallel with your current process on one client group for a month and count the corrections. If a tool needs fixing on one line in five, it has not saved you anything, it has just moved the work from typing to proofreading.
Workflow automation comes last, because it only pays once the steps either side of it are settled. Automating a process you are about to change is the most common way to waste money on this stuff.
- Drafting: paid business plan, three real jobs, two weeks, output reviewed as normal.
- Extraction: one client group, run in parallel, count corrections before you commit.
- Automation: only once the surrounding process has stopped moving.
- Everything else: revisit in six months. The market is moving fast enough that waiting costs you very little.
Client data: the part most firms skip
Accountants hold some of the most sensitive information a business has, and the tool vendors are not the ones who carry the consequences if it ends up somewhere it should not be. Before client data goes near any of this, get four things settled: which plan you are on and what its terms say about training and retention, where the data is stored, what your engagement letters already commit you to, and what your professional body currently expects. Then write the firm's rules on one page so staff are not making the call themselves at 6pm.
Free personal accounts are the real risk, not the technology. Somebody pasting a client's financials into a consumer chatbot on their own login is a problem no procurement process ever sees. That is an easier problem to fix than most partners expect: give people a sanctioned tool that works well, and the shadow usage stops. I am not a lawyer and none of this is legal advice; your own advisers and your professional body set the rules you have to follow.
What I do, and what I do not
I advise. In a session we work through which of the four categories matters for your practice, in what order, which specific tools are worth trialling, and what to leave alone this year. We can agree a written action plan as part of the work. I do not sell the software I recommend and I take no commissions from any vendor, which is the entire point of asking someone independent before you spend.
Development is separate from the consulting plan. If you decide you want something custom built, that is a separate engagement with its own scope, price and contract. Keeping those two things apart is deliberate: advice you pay for is advice that does not have to sell you a build.
Common questions
Will AI replace accountants?
No. AI is taking a growing slice of the mechanical work: coding transactions, extracting data from documents, drafting first-pass letters. Judgment, responsibility and the client relationship stay with the accountant, because someone has to sign the return and defend the position when it is questioned. The realistic change is that firms using these tools well will do more work with the same people than firms that ignore them.
Where should an accounting firm start with AI?
Start with drafting. Put a paid business plan of ChatGPT or Claude in front of the two or three people who write the most client correspondence, give them real jobs to use it on, and review the output together for a fortnight. It needs no integration with your ledger, it is reversible, and it shows you how your team actually behaves with these tools before you spend money on anything connected to client data.
Is it safe to put client data into AI tools?
It depends on the tool and the plan you are on, and it is a partner-level decision rather than an IT one. Use business or enterprise plans rather than free personal accounts, read what the provider says about training on your data and where that data is stored, check your engagement letters and your professional body's current guidance, and write the firm's rules down on one page so staff know what is allowed. This is general information, not legal advice.
Official resources
Explore the tools and check current details with the people who make them.
