Start with a complaint, not a strategy
The worst way to start with AI is a company-wide strategy workshop. The best way is to ask your team which task they most resent doing, and then look at that.
Resented tasks are almost always repetitive, rule-based and high-volume, which is exactly the profile of a good automation candidate. Your team has already done the analysis for you. They just called it complaining.
What a good first project looks like
- Lead hand-off. Form to CRM to follow-up, without anyone touching a spreadsheet. The classic, and still the most valuable.
- Quote and proposal assembly. The information exists in three places. Someone spends an hour assembling it. Every time.
- Reporting. Pulling numbers from five tools into one view, weekly, forever.
- Review requests. Nobody remembers to ask. A machine always does.
- Inbox triage. Sorting, tagging and routing what comes in, so the humans only see what needs a human.
Each of those is small, measurable, and finished in weeks rather than quarters. That matters: a first project that lands quickly buys the trust to do the next one. A six-month "transformation" that stalls buys nothing but skepticism.
What it should cost, and how to judge it
Judge every automation on one number: hours returned per month, multiplied by what those hours are worth. If nobody can estimate that before the project, nobody will be able to prove it afterwards.
A sensible agency will scope the first project small enough that the payback is obvious within a quarter. If the proposal is a platform build with a six-figure number and a discovery phase measured in months, you are being sold a transformation program, not an automation.
If the payback needs a spreadsheet to explain, it probably is not there.
What not to automate
The rule that governs all of this: automate a broken process and you just get to be wrong faster.
If your lead follow-up is bad, automating it produces bad follow-up at scale, reliably, forever. If your data is inconsistent, automation will make confident decisions on inconsistent data. Fix the process first. It is usually cheaper than the automation and it is always faster.
And leave the judgment with people. Machines are excellent at moving things between systems and poor at deciding what matters. Automate around the human decisions, not through them, that is the shape of our AI automation work, and it is why we usually start by fixing the data before anything gets automated at all.
What the first eight weeks actually look like
Concretely, because "AI transformation" is a phrase that means nothing and costs a fortune.
Weeks one and two: watch. Someone sits with the person who does the task and maps what actually happens, including the workarounds, the copy-pasting, and the bit where they check a spreadsheet nobody else knows exists. This step is where the real findings are, and it is the step most vendors skip because it does not look like progress.
Weeks three and four: fix and design. The process gets tidied before it gets encoded. Steps that exist only because of an old constraint get removed. What remains is split into data movement (automate) and decisions (keep human, but give them better information).
Weeks five to eight: build, test, hand over. Small, working, in production. Not a platform, a specific thing that does a specific job and can be measured.
- You should be able to see it working by week eight. If the timeline is six months before anything runs, that is a warning.
- The hours saved should be countable. If nobody can count them, nobody can defend the project.
- The team should be relieved, not nervous. The point is leverage, not layoffs, nobody was hired to be a copy-paste machine.
Then, and only then, do it again with the next process. Automation compounds when each project is small enough to finish and clear enough to prove. It stalls when it starts as a strategy.
What to avoid when you start
Most first automation projects fail in one of four ways, and all four are avoidable.
- Starting too big. A company-wide program is a way of never finishing anything. Pick one process, finish it, prove it.
- Automating a broken process. Fix it first. Encoding a mess in software gives it a version number and makes it permanent.
- Choosing AI where plain automation would do. A great deal of this work is classic process automation, deterministic, testable, cheap. Use a model where the input is genuinely messy, not because it sounds more impressive.
- Skipping the data. An automation acting on inconsistent data makes confident, consistent, wrong decisions. Get the tracking and data right before you let anything act on it automatically.
The businesses that get real value from this are rarely the ones with the most ambitious strategy. They are the ones who automated one annoying thing, measured the hours it gave back, and then did it again. It compounds quietly, which is the opposite of how it gets sold.
Frequently asked questions
Where should a Melbourne business start with AI automation?
With the task your team most resents doing. Resented work is almost always repetitive, rule-based and high-volume, which is exactly the profile of a good automation candidate. Start there rather than with a company-wide AI strategy workshop.
What is a good first automation project?
Lead hand-off from form to CRM to follow-up, quote assembly, weekly reporting pulled from multiple tools, automated review requests, or inbox triage. Each is small, measurable and can be finished in weeks, which builds the trust to do the next one.
How do I judge whether an automation is worth it?
Hours returned per month, multiplied by what those hours are worth. If nobody can estimate that before the project starts, nobody will be able to prove it afterwards. Payback should be obvious within a quarter for a sensible first project.
What should I fix before automating?
The process itself, and the data underneath it. Automating a broken process just produces the wrong outcome faster and more reliably. Inconsistent data leads to confident automated decisions built on nonsense. Both fixes are usually cheaper than the automation.