The scenario
A six-person bookkeeping practice in Parramatta. Three accountants, a receptionist, a junior and the owner. They have read enough blog posts to be convinced they should automate something. What they do not have is a plan, so what actually happened over the last year is this: the owner signed up for a workflow tool, watched four hours of tutorials, built one automation that sorts email, then forgot the login. The receptionist built something in a second tool that she alone can maintain. The junior set up a form that captures client documents, and it works, except nobody told the clients, so two of them still email scans.
Nothing is broken, and nothing is joined up. That is the normal starting position, and it is not a sign that automation does not work for businesses like this. It is a sign that nobody mapped the business before picking the tool.
Below is the method this practice can run this week, and the one after it, and the one after that. It is deliberately boring. The boring version is the one that survives a busy month.
What good looks like
Hold the picture in your head before you start, because most automation projects fail quietly and nobody notices for months.
After a month, good looks like this. One process runs every weekday whether or not anyone thinks about it, and the owner knows exactly where to look to see what it did. The receptionist can describe the process to a new starter from memory in under a minute. There is a written list of what is automated and what is not, and it is short. And when the automation breaks, the person who notices is the machine, not the client.
Bad looks like this: three tools, one person who understands them, a spreadsheet that still exists because everyone secretly does not trust the new system, and a nagging feeling that the automation is doing something wrong that nobody has checked.
The difference between those two outcomes is almost never technical knowledge. It is the order of operations.
Step one: audit three processes, not everything
Do not audit the business. Pick exactly three processes that are repetitive and happen at least weekly. For a bookkeeping practice: client document collection, reconciling bank feeds across client accounts, and chasing clients for missing details at end of quarter.
Write each one down as a list of steps, one line each, with who does it and when it triggers. Be honest about the ugly bits, including the step that says "ring the client again because they ignored the email". The ugly bits are where the time goes.
Then mark each line as one of three types:
- Copy. Information moves from one place to another with no decision.
- Judgement. A human looks at something and decides what to do.
- Chase. Someone waits for someone else, then reminds them.
Almost all automation value lives in the first and third types. The second type is where you lose money if you automate wrong.
Step two: pick one process with a clean start and a clean finish
Of your three, pick the one where you can name the exact moment it begins and the exact moment it ends. "Client sends a document" is a clean start. "Doc is processed" is a clean finish.
This rule quietly excludes most tempting candidates. "Automate client onboarding" sounds great and has a fuzzy finish, because onboarding is a pile of separate things. "Automate reconciling" is fuzzy too, because the judgement happens mid-stream. "Collect the four documents we need before a BAS" has a clean start, a clean finish and an obvious success state.
Start with the narrowest clean-edged process you have. You will be tempted to broaden it in week two. Resist until the narrow version has run for a month untouched.
Step three: fix the process before you automate it
The practice's document collection looks like this: at the start of the quarter, the receptionist emails each client a list of what is needed. Some clients reply with everything. Some send photos of receipts over three weeks. Some send nothing and get chased at the deadline. The receptionist tracks all of this in her head and a partially filled spreadsheet.
Here is the important part. Before any software, decide what "complete" means, and make it something a machine can check. Not "client has sent everything", but "four named documents have arrived for this client for this quarter". Complete is a fact, not a feeling.
Then remove any step that only exists because the information was missing. In this case, the receptionist writes a second email every quarter explaining what a "PAYG summary" is. That explanation belongs in the first email, or on the upload page, not in a step that repeats forever.
Automating a messy process produces a messy automated process, faster. Fixing it first is the actual work, and it usually takes less time than people expect, because it is mostly deleting steps.
Step four: build the smallest thing that runs on its own
The build has four parts, and nothing else.
- A place to put the information. A form, an upload page, a shared folder. One place per process. Do not run two in parallel.
- A place to see the state. One list showing every client and where they are: not started, documents outstanding, complete. If you already have a spreadsheet, keep the spreadsheet and have the automation write to it. It is already understood, and changing two things at once is how projects die.
- A trigger. Something that fires on a schedule or on arrival. A daily check is plenty for almost every small business process.
- A comms step. One or two messages, sent from the system, that tell the client what happens next and when.
That is it. No dashboards, no analytics, no custom app, no chatbot on the website. The temptation to add a nice interface is real; the interface is the part that gets ignored while the trigger keeps working.
Send the messages in your own words, from your usual address, and keep them short. Clients do not care that the message was automated. They care that it arrived and said the right thing.
Step five: run it in shadow mode for one week
This is the step people skip, and it is the one that saves them.
For one week, let the automation do its work but still do the job manually as well. Compare. The junior still checks the spreadsheet by hand each afternoon and marks any discrepancy next to what the automation did.
You are looking for three things. Missing items, which are documents that arrived but were not recorded. False completeness, which is the system marking a client done when something is missing. And odd inputs, which are the client who sends a Dropbox link, or a photo named IMG_4821, or a reply-all thread where the document is buried in the third message.
Shadow mode is uncomfortable and feels like doing the job twice. It is one week, and it converts a guess into a check. Everything you find in that week is something that would otherwise have surfaced as a client complaint four months later.
Step six: write it down somewhere a new starter can read
One page per process, no more. What triggers it, where the state lives, what the messages say, what to do when it breaks, who to call.
The one line that matters most is the failure line. What happens when the automation stops working? The honest answer for most small businesses is "nobody notices for a while". Write down the check instead: who looks at the list, how often, and what they do if a client has been sitting in the same state for too long.
That check is not automation. It is a person, five minutes a week, looking at a list. It is the cheapest insurance you will ever buy, and it is what makes the difference between an automation that earns trust and one that quietly rots.
The failure modes
Four things go wrong again and again.
Automating the wrong stage. People automate the step that annoys them most. The receptionist hates chasing, so the chasing gets automated first, and it automates bad data. The client was never asked properly, so the chase is chasing nothing. Start where the information enters the system, not where it hurts.
Two tools where one will do. The classic is a form tool plus a spreadsheet plus a reminder tool, each holding a partial copy of the truth. Every new tool doubles the number of places a record can be out of date. If a tool you already pay for can do the job, use it and delete the plan for the second one.
The invisible failure. The automation silently stops, maybe because a login expired or a form changed, and nobody knows until a client complains. Every automation needs one visible failure signal, even if that signal is "the daily list did not update". Silence is not success.
The owner-only system. Everything works until the person who built it goes on holiday. If nobody else can describe the process or restart it, you have not implemented automation, you have added a dependency. The one-page document and the second person who knows where to look are part of the build, not the paperwork afterwards.
Week two onwards: repeat, but slowly
Once one process has run clean for a month, move to the next from your three. Keep the same method: narrow, clean edges, fix first, build small, shadow it, write it down.
Two processes a quarter is a realistic pace for a business with no dedicated technical person. That sounds slow. In a year it is eight processes, which for most small businesses is the bulk of their repeatable admin, and every one of them has been checked rather than assumed.
The other thing that happens, and it is worth expecting, is that the first process makes the second one easier. Once client documents arrive in a consistent way, the end-of-quarter chase gets simpler. Once the chase is automatic, the quarterly report has better data. Automation compounds when the processes are linked, which is another reason to fix the front of the chain first.
What to do this week
Pick three processes. Write down the steps for the one with the cleanest start and finish. Define what complete means as a checkable fact. Book one afternoon in the calendar for the build, and one week after that for shadow mode.
Everything else can wait until the first one is running.
