To the principal,

You subscribed to ChatGPT. You have probably tested Copilot, maybe Claude. You have sat through the conference sessions. And if you are honest about what you have to show for it, it is better-worded emails.

You have not been careless. You have been sold the same thing everyone else was sold, by people using identical language, and none of them mentioned the part that decides whether any of it works.

I have been on the wrong end of the same pitch. I have bought software on a demo that solved a problem I did not have, rolled it out to a team that quietly went back to the spreadsheet, and paid the licence for another year before admitting it.

So before I tell you what I think does work, here is what we decided when we built SYSTEMology.

That last one costs us the most, and I will come back to it.

There are a great many people talking to accountants about AI right now, and it is worth being precise about why that is a problem, because it is not the reason the marketing assumes.

It is not that you cannot evaluate a claim. You evaluate commercial claims for a living. It is most of the job. The difficulty is that every vendor in this category is making an identical promise, in identical language, and not one of them will put a number on what is left after somebody checks the work.

The part that stings

Here is the awkward bit, and I say it with respect.

You would have spotted this in a client’s business inside twenty minutes.

You can open a new client’s file and see exactly where their wheels are going to come off. Then you walk back into your own practice, where the answer to half the questions is still “ask me,” where the file noting is done a particular way because you decided it eleven years ago and never wrote it down, and where a piece of the firm leaves every time a person does.

This is not laziness. You are the busiest person in the building. CPA Australia’s senior assessor put it more bluntly than I would dare to:

A practitioner on an accounting forum said the same thing about himself, with rather more feeling:

The secret almost everyone is missing

For ten years I have been writing about business systems. Documenting processes. The unglamorous work that never felt urgent and never got done in the quiet period, because the quiet period keeps not arriving. I thought I was writing about efficiency.

That is the whole thing. Every AI tool you have been sold assumes there is a defined process underneath it to execute against. When there is, the tool performs roughly as advertised. When there is not, it produces something plausible that a human has to check line by line, and the checking eats the saving.

Which means the boring documentation job you have been putting off for a decade quietly became the highest-leverage work in your firm, and nobody sent the profession a memo.

I see it on calls every week. Firms that did the documenting first are getting multiples of what the others get from the same tools, bought in the same month, at the same price. It is not that they picked better software. It is that they had something for the software to work from.

Two things follow from getting this right, and they are worth being direct about. Your firm stops buying AI tools and starts using them, which is a different activity with a different bank balance. And the practice becomes worth something without you standing in the middle of it, which matters a great deal on the day you would like to stop standing there.

The three things you actually want from AI

When accountants tell me what they were hoping for, it comes down to three things. Each one is achievable. Each one fails for the same reason.

1. AI that automates your repetitive work

This is the one worth having, and it is real. It is also the one that most reliably disappoints, because repetitive is not the same as defined. A task can happen forty times a week and still exist only as a habit in one person’s hands.

Here is a firm that had bought roughly ten thousand dollars of licences:

Notice what is being described there. Not a technology that cannot do the work. A technology handed the work with no standard to hold it to.

2. AI that sounds like you

Almost every firm has this one already, and almost nobody rates it. Across the surveys, 77% of firms use AI for communication tasks. Emails, meeting notes, tidying a letter. Only 23% use it for anything touching reporting. Of 600 firms surveyed, 7% report a transformational impact and one in five report none at all.

So AI has arrived in your correspondence and stalled at your actual work. The way the trade press put it:

And here is the connection almost nobody makes. The reason it does not sound like you is the same reason it cannot do your work. It has no source material about how your firm actually operates. Feed it nothing and it hands back the industry average, competently written.

3. AI that adds money to the bottom line

This is the one nobody can show you, and it is worth understanding why.

Only about 18% of firms track the return on this at all. Gartner found that AI delivers real gross time savings, and then roughly 69% of the gain disappears into rework, training and new tasks that did not exist before. The time is genuinely saved. It just does not arrive anywhere you can see it.

A saving only reaches the bottom line when it comes out of a defined process, in a measured quantity, from a role somebody owns. Otherwise it disperses into the day.

The one I will not sell you yet

Since I said we would never overstate this, here is where I think it genuinely is not ready.

AI bookkeeping and transaction coding. The demos are excellent. The practitioners running it on real ledgers are not convinced:

Do the arithmetic yourself. At 97% accuracy on a client running two thousand transactions a month, that is sixty errors to find. You do not get to skip the review, and the review is most of the work.

It will get there. It is not there. If someone is selling it to you as solved this quarter, they have not run it on a messy file.

Writing the programming used to take a decade

Here is why I am writing this letter in 2026 and could not have written it in 2019.

For twenty years the honest answer to “how do I get this out of my head” was: sit down and write it, it will take months, sorry. That was the bottleneck. It is the reason your process folder has three good documents in it and a date stamp from a March you cannot remember.

That part is over. The fastest-growing use of AI in accounting firms right now is not tax work. It is writing the procedures. Roughly 43% of firms are already using it for SOPs, policies and process documentation, up sharply in a single year.

The method is unglamorous and it works. Talk through the process out loud while you do it, once. Something transcribes it, drafts the procedure, and the person who actually does the job corrects it. A two-hour writing job that never happened becomes a twenty-minute conversation that does.

So the programming that used to take a decade now takes a fraction of it. That is the entire reason this letter is worth your time.

The firm that got the AI dividend had already written the code

Shannon Smit runs SMART Business Solutions on the Mornington Peninsula. Accounting, financial planning, mortgage broking, more than twenty staff. In 2024 the firm was named Multiservice Firm of the Year at the Australian Accounting Awards.

What matters for this letter is not that Shannon fixed a broken firm. She did not have one. She has been documenting since she founded the practice in 2007 and using systemHUB to hold it since 2016. She is, by her own description, someone who fell in love with process behind a McDonald’s counter at fourteen.

That decade of unglamorous work sat there looking like overhead. Then her son Ryan joined the firm, took the Systems Champion role, and went looking for something to automate.

He found file noting. Every client conversation had to be written into three separate systems. Three to five minutes a note, around twenty notes a day, across four advisers and the admin team behind them. He built a bot for it, chosen because security posture mattered more than novelty.

That single automation reclaimed over a thousand hours a year. Roughly an hour a day, per person, on the planning team.

Then Ryan asked the question that ought to be asked more often. “So what do I do now? Have I just reduced my own work hours, and therefore my wage for the week?” The answer was no. We will find you more work, and we will automate more things.

Here is Shannon on why it worked. It is the argument of this entire letter, in one sentence, from someone who is not selling you anything:

While most firms are still trying to get work done consistently, Shannon’s was in a position to take the AI dividend the week it became available. Not because she moved fast on AI. Because she had spent a decade writing the programming, which meant there was something for the machine to run.

One more detail, and I include it because the tidy version of this story is not the true one. When the firm rolled out AI training to everybody, two team members opted out after six months. That is what real adoption looks like. Anyone showing you a case study without a number like that in it is showing you a brochure.

Read Shannon’s full story here.

You are already required to have it written down

Set AI aside entirely for a moment.

APES 320 has required every member in public practice in Australia to have a documented system of quality management since 1 January 2023. Not just auditors. Tax, compilation, bookkeeping, insolvency, valuations. Clause 3.6 is one line: a firm shall document its policies and procedures and communicate them to the firm’s personnel. Clause 3.17 requires evidence that the system actually operates.

Firms do not fail this by lacking procedures. They fail by never having written them down:

If you are small and thinking this cannot apply at your size, the standard has answered you already. Documentation for smaller firms may be “less formal and extensive,” and manual notes, checklists and forms are acceptable. It scales down. It does not switch off.

So the same work satisfies a reviewer and programs the machine. I would rather you did it once.

What I am not going to sell you

If this letter only listed things I can help with, you would be right to discount all of it.

I will not write your APES 320 manual

We are not quality review consultants and we do not sell a compliance manual. Several firms specialise in exactly that, your professional body can point you at them, and if a reviewer is due next quarter that is the call to make, not this one.

I am not going to frighten you about the 85% benchmark

I could. It works. One survey of 450 practitioners found 89% said the ATO lodgement benchmark caused them stress.

Except the ATO says, in its own words, that meeting 85% “is not a requirement to access due dates under the lodgment program.” There is no support strategy at all until you fall below 75%, and nothing beyond a phone call until you are under 40%. An assistant commissioner has said publicly, “We know the stress it provides and it is not the intent of the program.”

The most-feared number in your year is softer than almost everyone believes. Telling you that costs me a useful lever. It is still true.

If you are a sole practitioner with no staff, you probably do not need us yet

Documenting matters most when knowledge has to move between people. If there is nobody for it to move to, the return is real but slower. Read the book instead. It is free. Come back when you have hired your second person.

This does not work without one person owning it

The firms where this fails are the ones where the principal takes it on personally, alongside everything else, and it quietly dies by March. Shannon’s firm worked because Ryan held the role and had actual hours for it. It does not need to be a senior person. It rarely should be. But if you cannot name who it will be and what comes off their plate to make room, now is not the time, and I would rather say so than take the engagement.

Pick one process. Not ten.

If you take nothing else from this letter, take the order, because it is the thing almost everyone gets backwards.

Map how work actually moves through the firm, from first enquiry to invoice paid. Find the one process that causes the most pain or that only one person can do. Document that one by talking it through rather than writing it. Get your team using it. Then, and only then, look at what can be automated, because now there is something defined for the machine to run.

That is the whole method. Everything else is detail.

There is a timing argument here, and it is not a manufactured one. Your profession is buying AI tools this budget cycle. The firms that wrote the programming first are compounding on every tool they buy. The ones that did not are funding licences that produce better-worded emails. That gap widens every quarter, and it is very hard to close from behind.

You can do all of this yourself. The book is free and lays out the full framework, including the mistakes. If you would rather talk it through, book a call and we will look at your firm specifically: where your processes actually sit, which one to start with, and whether we are any use to you. If we are not, I will say so on the call.