Every week another AI automation company launches in Australia, and every week another business owner tells us about the last one that did not work out. The pattern is familiar. A polished demo, a big promise, an invoice, and then months of silence or a half-built workflow nobody on the team can maintain. Automation is genuinely useful, but the gap between a good provider and an expensive mistake is wide, and it is rarely obvious from the sales call alone.
We build automation with Claude for Australian small and mid-size businesses, so we tend to see the wreckage of failed projects when owners arrive asking for a rebuild. This is the honest checklist we wish more buyers had before they signed. Read it as a set of questions to ask, not a list of names to avoid. The right provider will pass every one of these tests without flinching.
Red flags in the sales pitch
The sales conversation tells you most of what you need to know. A provider who understands your business spends more time asking about your work than describing their software. Watch for these signals before any money changes hands.
They quote a price before understanding your process. A fixed number in the first call usually means they are selling a template, not solving your problem.
The demo runs on their data, never yours. If they will not test against a real task from your business, the tool probably cannot handle your edge cases.
Everything is "AI" and nothing is specific. Ask which model, which tasks, and what happens when the model is unsure. Vague answers point to thin engineering.
They promise full autonomy on day one. Reliable automation starts narrow and earns trust. Anyone promising to replace a whole role immediately is overselling.
No mention of your data or where it is stored. For Australian businesses this matters under the Privacy Act, and a serious provider raises it before you have to.
Red flags once the work starts
Some companies pass the sales test and still fail in delivery. These are the warning signs that a project is heading for the $45,000 write-off pile rather than a working system.
You cannot see how it works. If the build is a black box only they can touch, you are renting a dependency, not owning an asset.
Scope keeps growing but the working demo never arrives. Real progress shows up as small pieces that function, not endless discovery documents.
They cannot tell you the running cost. Model usage has a real monthly figure in AUD. A provider who cannot forecast it has not thought about your budget.
Handover is an afterthought. Ask on day one how your team takes over. If the only answer is a permanent retainer with no exit, that is the business model, not the solution.
No plan for when the automation is wrong. Good systems flag low-confidence cases to a person. Systems that guess silently create errors you find out about too late.
What a straight answer sounds like
The contrast is simple. A provider worth hiring leads with your workflow and treats the technology as a means to an end. We build Claude-first because the model is strong at the language-heavy tasks most Australian businesses actually want handled: reading email, drafting replies, checking documents against a policy, and summarising records. Anthropic is the credibility behind Claude, but the product is what does the work. A good partner shows you a small working version against your real task inside the first week, tells you the monthly cost in plain figures, and writes down how your team keeps it running without them.
That is the real test. Not the polish of the pitch, but whether you leave the room understanding what will be built, what it costs, and who owns it at the end.
The AUD maths of getting it wrong
The cost of a bad fit is not only the invoice. A typical failed mid-size automation project in Australia lands somewhere between $25,000 and $45,000 in direct fees, then adds the quieter costs on top: the staff hours spent briefing a vendor who never delivered, the process that stayed manual for another year, and the internal trust in automation that takes months to rebuild once it is broken.
One Melbourne services firm we spoke with had spent close to $60,000 across two providers before a working version existed. The build that finally held up cost a fraction of that number, because it began with a single narrow task and proved value before anyone talked about expanding it. The lesson is not that automation is risky. The risk lives in the buying process, not the technology.
So protect yourself at the point of purchase. Ask for a paid pilot on one real task. Insist on watching it run against your own data. Get the monthly cost in writing before you commit to a bigger scope. A provider who welcomes those three conditions is almost always one worth keeping, and the ones who resist them have just told you everything you needed to know.
If you have been burned before, or you want a second opinion on a proposal sitting in your inbox right now, we are happy to talk it through. You can book a short brainstorm session to map one workflow and see what a straight answer actually looks like.
