What should you automate with AI first?
Deciding what to automate with AI first comes down to one process: the one that is high volume, low judgement, entirely described by rules and information you already have written down, and expensive when it goes undone. Almost every business has one. Almost nobody picks it, because it is boring and the exciting ideas shout louder.
The first project matters more than its size, because it decides whether there is a second one. A first build that quietly works buys you a decade of goodwill inside your own company. A first build that is ambitious and shaky ends the conversation for years, regardless of what it might eventually have done.
What makes a process a good first candidate?
Run every candidate through four questions, and only build the one that passes all four.
Does it happen often? A task that occurs twice a year will never repay the effort of describing it precisely. You want something that happens daily or many times a week, because that is where small savings compound and where you get enough examples to know quickly whether it is working.
Is the judgement small? Good first candidates are decisions with a right answer that a competent new hire could reach from a written procedure. Bad ones require taste, relationships or reading a room.
Is the source material already written down? If your prices, policies and procedures are current and live in one place, the build is short. If they live in three inboxes, a shared drive and one long-serving employee's memory, the real first project is writing them down, and you will get value from that whether or not you ever automate anything.
Does it cost you something measurable when it does not get done? Missed calls have a number. Slow quotes have a number. Unchased invoices have a number. If nothing measurable happens when the task is skipped, automating it is a hobby.
Where do most businesses actually start?
Most start at the front door, because the front door is where undone work turns into lost money fastest.
For service businesses the first machine is usually the one that catches enquiries nobody could get to, qualifies them and books them. That is the whole reason never missing a lead is our most common opening build: the volume is high, the judgement is small, the prices are already written down, and the cost of failure is a number the owner already feels.
The second most common is the paperwork spine. Quotes going out the same hour instead of next Tuesday. Invoices chased without anyone having to feel awkward. Review requests sent at the right moment every single time rather than when someone remembers. That is the admin disappearing, and it tends to be the one that gives an owner their evenings back.
For firms that quote for a living, the estimating pile beats both. Reading vendor quotes line by line and benchmarking them against real market pricing is high volume, heavily rules-based, and expensive to get wrong, which is exactly the profile you want.
What are the wrong places to start?
Four kinds of work make poor first projects, however tempting they look in a demo.
Anything requiring taste or relationship is wrong. Sales conversations with your best accounts, redundancy decisions, creative direction. These are the things you are paid for.
Anything where a mistake is unrecoverable is wrong for a first build. You want a first project where a bad output is caught by a human, corrected in a minute, and forgotten. Not one where a bad output is a compliance breach.
Anything nobody has ever written down is wrong until somebody writes it down. This is the most common cause of AI projects that stall at eighty percent for six months.
And anything that exists mainly because a competitor announced they were doing it is wrong. That is not a project, it is anxiety with a budget.
How do you size the prize before you spend anything?
Put a dollar figure on the current cost of the process, in writing, before you take a single quote.
Count the volume: how many times a week does this happen. Count the time: how long does each one take, including the switching cost of being interrupted. Count the failure rate: how often does it not get done at all, and what does that cost. Multiply it out into an annual number.
That number does three useful things. It tells you whether any price is sensible. It gives you the success measures to hold a supplier to. And it becomes the baseline you compare against later, which almost nobody captures and everybody wishes they had. Take the baseline in week one, before anything changes, because you cannot reconstruct it afterwards.
Our own view is that a price should be justified by the value it releases rather than by the hours it took to build. That works both ways: it means a genuinely valuable machine is worth paying properly for, and it means a machine that releases nothing should not be bought at any price.
How small should the first one be?
Small enough that it is live and running in weeks rather than quarters, and narrow enough to describe in one sentence a stranger would understand.
We think in terms of one working machine at a time. It goes on paper first, gets a fixed price and named success numbers, gets built beside your existing systems rather than on top of them, then runs supervised with human approval gates for as long as you want them. Thirty days later there is a one-page verdict in numbers. If it earned its keep, it earns the next one.
That sequence is deliberately unglamorous and it is described in full on the method page. Small, then proven, then bigger is the only growth model we trust, and it is exactly how AI should enter a business.
Where to look next
The crew is a useful shopping list for this exercise: nine named specialists, each doing one job, each with a page showing what you provide and what you get back. Reading them is a fast way to recognise your own bottleneck.
For the wider frame, the Australian Government's AI Ethics Principles are worth twenty minutes before you scope anything, and business.gov.au lists current digital and adoption support you may be eligible for.
When you have your one process and its number, bring both to us. If it is not worth building yet, we will say so.