Most jobs sold as AI are automation wearing a better jacket, and automation is the cheaper, steadier thing to buy.
"AI" has become the word people reach for whenever a computer should be doing something a human currently does. That is expensive, because the two jobs underneath are priced very differently.
The split Penny uses is simple enough to apply yourself before you speak to anybody.
If the hard part is copying information between systems, checking it against a list, and making sure something happens on time, that is automation. The rules are steady, the data is already on a screen somewhere, and nothing has to be interpreted.
This is plumbing, and plumbing is well understood. Our published guide for it is A$2,500 to A$12,000 to build, most likely around A$6,000, and A$80 to A$300 a month to run. A builder can usually connect these systems in days rather than months.
The giveaway is that you can explain the whole job as a series of "if this, then that" sentences without ever using the word "depends".
If the hard part is messy email, documents nobody has standardised, working out what a customer actually means, or producing writing that has to make sense, then there is real language work in the middle and rules alone will not carry it.
That is where AI earns its keep, and it costs more. Our guide is A$8,000 to A$30,000 to build, most likely around A$15,000, and A$180 to A$600 a month to run, because something has to host it, watch it and pay for every run.
Worth noting: needing AI is not the same as needing a build. If one person uses it a handful of times a day while sitting at their desk, a paid ChatGPT or Claude subscription probably covers it. AI becomes a build when it has to run inside your systems, on its own, reliably.
The costly shape is a job that reads something messy, makes a call about it, then moves the result through three systems, with a person approving part of it.
That is not one build, it is two stitched together, plus an approval screen, plus permissions, plus a record of who approved what. The happy path is quick. The awkward fifteen per cent is most of the invoice, and it is why we never score a process as fully removable.
Four questions. Answer them about the job as it is done today, not as you wish it were done.
Now read them back together.
There is one answer that overrides the other three questions. If the job would decide or rank people, or affect someone's rights, safety or livelihood, stop. Hiring, credit, health, legal matters, visas, who gets a place in a program.
Work like that carries obligations, and AI Builds will not scope it automatically. The right first step is an adviser who knows those obligations. It is also worth reading what the Australian Government expects of you: the National AI Centre replaced the Voluntary AI Safety Standard with its Guidance for AI Adoption on 21 October 2025, which sets out six essential practices for safe and responsible AI (industry.gov.au).
Where a person is still making the call and the software only sorts or summarises, a build can be scoped. The scope has to say plainly that a human stays in charge, and the quote has to include the screens that make that real.
Plenty of people answer "not sure" to the hard part, and that is fine. Penny treats an unknown on what to build as the cheaper option, so nobody gets talked into a build by their own uncertainty. She treats an unknown on risk the careful way instead, and says so on the verdict either way.
That is the whole bias of this site in one sentence. If it is genuinely a toss-up, buy the smaller thing.
Written by the AI Builds team. AI Builds is operated by Media Mantra, ABN 89 594 851 152.
She checks the job, writes the scope, and never sells you a build.