I’ve got a friend who runs operations at a packaging manufacturer in Western Sydney. Last year, they spent $180,000 on an AI system that was supposed to optimise their production scheduling. Today, it sits unused while staff go back to spreadsheets.

This isn’t an unusual story. Industry research suggests somewhere between 60-80% of AI projects in manufacturing don’t deliver their expected value. That’s a lot of wasted money and disappointed executives.

But here’s the thing: it’s rarely the technology that fails. The software usually works. The problem is everything around it.

The “silver bullet” mindset

The most common failure pattern I see starts with a vendor demo. Somebody sees a slick presentation, gets excited, and purchases a system without really understanding what it takes to make it work.

AI in manufacturing isn’t like buying a piece of equipment. You don’t just install it and flip a switch. It needs:

• Clean, consistent data to learn from
• Integration with existing systems
• Staff who understand and trust it
• Ongoing maintenance and adjustment

That packaging manufacturer? They had data quality issues that made the AI’s recommendations unreliable. But nobody checked that before writing the cheque.

Data problems kill more projects than anything else

Let me be specific about what “data quality issues” actually means.

Missing data: Sensors that drop out. Manual entries that get skipped. Gaps in the historical record the AI needs to learn from.

Inconsistent data: Different shifts recording things differently. Unit changes. Sensor calibration drift. The AI treats all data as equal, so garbage in, garbage out.

Wrong data: I visited a plant where their “machine downtime” field actually included scheduled maintenance, breaks, and changeovers all lumped together. The AI couldn’t distinguish unplanned failures from normal operations.

Siloed data: Production data in one system. Quality data in another. Maintenance records in a third. Energy consumption somewhere else. Getting a complete picture requires integration work that’s often underestimated.

One manufacturer I worked with spent three months just cleaning and standardising their data before the AI could do anything useful. That’s not unusual—it’s typical.

The people problem nobody wants to discuss

Technology projects fail or succeed based on people. Revolutionary insight, I know. But it’s worth spelling out how this plays out specifically in manufacturing AI.

Operators don’t trust it: If the person running the machine doesn’t believe the AI’s recommendations, they’ll ignore them. Building trust takes time and demonstrated reliability.

Middle management feels threatened: If the AI is making decisions that used to be their job, they may subtly (or not so subtly) undermine the project.

IT and operations don’t talk: AI projects often fall into the gap between IT (who understand the technology) and operations (who understand the process). Neither owns it fully.

No one’s accountable: I’ve seen projects where the vendor delivered what was specified, IT installed it, and operations was supposed to use it—but nobody was actually responsible for making it work.

The packaging company I mentioned? The operations manager who championed the project moved to another company six months in. Without his push, the project drifted.

Starting too big

Another pattern: trying to transform everything at once.

“We’re going to be a fully digitised, AI-driven smart factory!” Ambitious. But also a recipe for failure.

The manufacturers I’ve seen succeed start small. One process. One machine. One problem. They prove the concept, learn what works, build internal capability, then expand.

A food manufacturer in Geelong started with AI-assisted scheduling for just their mixing department. Took them eight months to get it working properly. Then they expanded to packaging. Then warehousing. Three years later, they’ve got AI throughout the operation—but it was gradual.

Unrealistic timelines

Vendor sales cycles create pressure to show quick results. Executive sponsors want returns within a quarter or two. But manufacturing AI typically takes longer than people expect.

Training the AI: Machine learning systems need time to learn your specific operation. That’s months, not weeks.

Integration: Connecting to legacy systems, SCADA networks, and existing software always takes longer than quoted.

Change management: Getting staff comfortable with new systems and processes is measured in months.

Iteration: The first version rarely works perfectly. You need cycles of adjustment and improvement.

A realistic timeline for meaningful results from a manufacturing AI project is 12-18 months. Anyone promising faster should explain exactly how.

So what actually works?

Based on the projects I’ve seen succeed:

Start with a real problem: Not “we should do AI” but “we’re losing $X to this specific issue and AI might help.” Clear problem, clear success metric.

Get the data foundation right first: Assess your data quality honestly. Fix issues before buying AI software.

Find an internal champion: Someone with enough authority to push through obstacles and enough operational knowledge to spot problems.

Budget for integration and change management: The software licence is often the smallest cost. Plan for the rest.

Set realistic expectations: AI in manufacturing is a marathon, not a sprint. Executives need to understand that.

Work with people who’ve done it before: Whether that’s hiring experienced staff, using consultants, or finding vendors with genuine Australian manufacturing experience. AI consultants in Sydney who understand industrial operations can help you avoid common pitfalls.

The projects that succeed

I don’t want to be all doom and gloom. When manufacturing AI works, it really works.

A steel fabricator I know cut their scrap rate by 34% using AI-assisted process control. A logistics operation in Brisbane reduced their energy costs by 22% through AI-optimised scheduling. A food producer improved their yield by 8% with predictive quality systems.

These aren’t magic—they’re the result of methodical implementation, realistic expectations, and perseverance through the inevitable setbacks.

The technology is ready. The question is whether organisations are ready for the technology. That’s a harder problem, but a solvable one.