The 95 Per Cent Failure Rate Is Not An AI Problem

Most Australian SMEs I speak with are already having a go at AI. Some are running formal pilots, others have a team member quietly experimenting on the side, and plenty have signed up for a handful of subscriptions just to see what sticks. The appetite is there, and so is the optimism that AI can lift both quality and efficiency.

Yet a recent MIT study of 300 enterprise AI deployments found that 95 per cent of them are failing to deliver the value the business expected. Fewer than one in twenty is producing meaningful business impact. That is a staggering number, and it lines up with what I see on the ground.

The good news is that almost none of these failures are really about the AI. They are human, organisational and strategic, which means they are entirely avoidable.

The most common mistake I see is businesses rushing to buy AI licences before they have a problem to point them at. An expensive, top-of-the-range subscription rolled out to every staff member, without any specific project or task attached, almost always ends up forgotten in a tab nobody opens.

Treat AI like any other software purchase. List the pain points in the business, rank them, and ask honestly which one is a tool that can actually solve the problem. Then pick the narrowest, lowest-risk version of the best candidate and ship it quickly. The point of the first project is not the project. It is the internal reputation and the personal confidence you build by getting one in the win column.

Sometimes the honest answer is that you don't need AI at all, and that is a perfectly good outcome. Deploying AI where a simpler tool or a process change would have done the job invites exactly the question you do not want your team asking: why are we doing this?

It is also worth remembering how sceptical your own team probably already is. A 2025 KPMG and University of Melbourne study of 48,000 people across 47 countries found that only 30 per cent of Australians believe the benefits of AI outweigh the risks. That is the lowest level of trust of any country surveyed.

Any new AI rollout therefore starts under a cloud of doubt and suspicion, and if the first one falls over, the next attempt is judged twice as harshly. Recovering from a failed pilot is genuinely harder than starting from zero. That is why the first AI project should be small, narrow and almost embarrassingly easy to call a success.

Another trap is borrowing frameworks from large enterprises. Governance committees, multi-year roadmaps and formal AI strategies have their place at organisations with thousands of staff. At twenty staff, they consume time the business does not have and slow the experimentation the business actually needs. SMEs win on AI by being faster and more pragmatic than the big end of town, not by mimicking it.

Then there are the businesses that treat AI as a one-off project rather than a capability. They deliver a first win and quietly move on. The SMEs that get the most out of AI treat that first project as a way to build internal muscle. They keep the champion in place, run the next initiative on the heels of the first, and turn AI from a one-time experiment into a permanent part of how the business operates.

It is rarely the AI that fails. It is almost always the project and people around it.