Australia will spend $172 billion on IT this year. The projects that fail won't fail at the technology.

Gartner reckons Australian organisations will spend A$172.3bn on IT this year, and that more than 40% of agentic-AI projects will be scrapped by 2027. What actually kills the ones that die isn't the technology.

Australian organisations are on track to spend A$172.3bn on IT this year. For the first time, software is the biggest single line of that, ahead of IT services. A fair slice of it is wearing an "AI" label.

That figure comes from Gartner's latest forecast , and it's worth sitting with for a second. A$172 billion is not experimentation money. It's commitment.

Here's the part that didn't make the headline. The same firm reckons more than 40% of agentic-AI projects will be scrapped by the end of 2027 . Not paused. Cancelled. The reasons they give are unclear business value, rising cost, and weak controls.

I've sat in enough of these rooms to believe it. And I'd add something the analysts are too polite to say plainly.

The failures are rarely technical

When an AI project gets quietly shelved, the post-mortem almost never reads "the model couldn't do it." It reads like this. Nobody agreed what "done" looked like. The work was handed to a tool instead of owned by a person. The pilot dazzled everyone in the demo, then had nowhere to go, because no one had changed the actual process around it.

Gartner has a blunt phrase for part of the problem: "agent washing." Vendors rebranding the chatbot or the automation script they already sold you as an "agent." By their count, of the thousands of companies claiming to do agentic AI, perhaps 130 genuinely do. So a portion of that A$172 billion is buying old software with a new noun on the box.

Buying is the easy part

Signing the licence is easy. It feels like progress. You can put it in a board update.

The hard part is the decision underneath it. Which specific piece of work should change. What "good" looks like once it does. Whose name is against it when it breaks. None of that comes in the software, and none of it is on the price list. It is also the thing that separates the projects that survive from the ones written off in eighteen months.

The engagements that work tend to follow a pattern. They don't start with a platform. They start with one workflow somebody actually hates: the month-end scramble, the inbox that swallows three people, the report nobody can produce before the meeting starts. They define what a good outcome looks like in plain language. They put one owner on it. Then, and only then, do they reach for a tool, and usually a smaller one than they expected.

If you're running a mid-market business

You don't win this by matching the spend. You can't, and you shouldn't try. Your advantage isn't budget. It's that you can change a process this quarter without a steering committee and three rounds of sign-off.

So before the next AI line item lands on your desk, ask three questions:

• What is the specific piece of work we're trying to change, in one sentence?

• What does "done well" look like, and how will we recognise it?

• Who owns this? Not the vendor, not "the team." A person.

If you can't answer those, more software won't save you. You'll just have a more expensive version of the same problem, and you'll be in the 40%.

The businesses getting something real out of this aren't the ones that bought the most. They're the ones that changed one thing on purpose, and knew who was responsible when it wobbled. Spending is a number. Deciding is the work.