I’ve been following a manufacturer in Geelong for the past two years as they implemented AI for process optimisation. They recently shared their results publicly, so I can finally write about it.

The headline number is impressive: 34% reduction in scrap rate, worth about $1.2 million annually. But the journey to get there is more instructive than the result.

The starting point

The company manufactures plastic components for automotive and industrial applications. Injection moulding, with about 40 machines running various products.

Their scrap rate had been stuck around 4.5% for years. They’d tried the usual approaches—better materials, equipment upgrades, operator training. Improvements were marginal and often temporary.

Quality depended heavily on which operator was running which machine. Experienced operators knew the subtle adjustments needed for different conditions. Less experienced operators followed procedures but couldn’t match results.

“We had 30 years of knowledge walking around in people’s heads,” the production manager told me. “When those people retire, that knowledge leaves.”

Why they chose AI

The trigger was a particularly bad quarter where scrap spiked to 6.8% due to a combination of new products, material variations, and operator turnover. The cost was obvious; so was the need for a different approach.

They looked at several options:

• More automation (expensive, didn’t solve the knowledge problem)
• Better documented procedures (already tried, limited impact)
• AI-assisted process control (seemed promising but unproven)

After visiting several facilities that had implemented AI solutions and speaking with AI consultants in Brisbane, they decided to pilot AI on two machines.

The implementation

Phase 1: Data collection (4 months)

Before any AI work, they needed data. Their machines had basic monitoring, but not enough for AI.

They added sensors tracking:

• Process parameters (temperatures, pressures, speeds, times at 1-second intervals)
• Machine status (cycle times, faults, mode changes)
• Material batch information
• Environmental conditions (ambient temperature, humidity)
• Quality outcomes (good/scrap/rework, defect type)

The quality data was the tricky part. They needed operators to record defect information accurately and consistently—not just “scrap” but what type of defect. This required new procedures and training.

Phase 2: Analysis and model development (3 months)

With six months of data, they worked with a technology partner to develop predictive models.

The analysis found patterns that weren’t obvious:

• Certain temperature combinations correlated with specific defect types
• Material from one supplier behaved differently than others, requiring different settings
• Morning start-ups had higher scrap rates that could be reduced with adjusted procedures
• Some defects correlated with environmental humidity—a factor they’d never considered

The first models predicted quality outcomes from process parameters with about 80% accuracy. Not perfect, but useful.

Phase 3: Pilot deployment (3 months)

They deployed the AI system on two machines in advisory mode—it recommended settings, but operators decided whether to follow recommendations.

Initial results were mixed. Operators were sceptical. “The computer says to reduce temperature, but that doesn’t feel right for this material.” Sometimes they were correct; sometimes the AI was.

The breakthrough came when an experienced operator tested AI recommendations systematically over two weeks. The AI-recommended settings outperformed his intuition about 60% of the time. That operator became an advocate, which shifted attitudes.

Phase 4: Expanded rollout (6 months)

After the pilot showed promise, they expanded to 15 machines (their highest-volume products).

Key learnings from the rollout:

• Training needed to be extensive—not just how to use the system, but why it made certain recommendations
• The AI needed to handle material variation better; they added material testing data as an input
• Some machines behaved differently enough that models needed adaptation
• Operators needed to understand when to override the AI (novel situations, equipment issues)

Phase 5: Optimisation and expansion (ongoing)

Today, 28 of their 40 machines run with AI assistance. The remaining 12 are lower-volume products where the payback didn’t justify implementation.

The system continues to improve as it learns from new data. They’ve gone from 4.5% scrap to under 3%—the 34% reduction.

What they spent

The project was not cheap. Approximate costs:

Sensors and hardware: $180,000 (across all implemented machines) Software and AI development: $220,000 Integration with existing systems: $75,000 Internal labour (project team): $150,000 (over two years) Training: $45,000 Ongoing licensing and support: $60,000/year

Total initial investment: roughly $670,000 Ongoing cost: roughly $60,000/year

With $1.2 million annual savings, payback was under 8 months. That’s excellent ROI.

What they learned

Start with data infrastructure

“We should have started the data project 18 months earlier,” the production manager said. The data collection and quality work took longer than expected and was essential to everything else.

Operator involvement matters

The most successful machines are those where operators engaged with the AI system, not just followed it passively. They encouraged operators to challenge recommendations and provide feedback.

Not all machines are equal

Some machines and products saw dramatic improvement. Others saw little. The AI worked best where process variation was high and the relationship between parameters and quality was complex. Simpler products on well-tuned machines had less room for improvement.

Continuous improvement required

The AI isn’t a set-and-forget solution. New products need model updates. Material changes require adaptation. Performance monitoring is ongoing.

Culture change is real

“The biggest change wasn’t technical,” the production manager told me. “It was how people think about data and evidence. We’re a more analytical company now.”

Would they do it again?

Without hesitation, yes. Even knowing the difficulties, the ROI is clear.

Their advice to other manufacturers considering AI:

1. Start with your most painful quality problems. That’s where the payback is and where motivation is highest.

2. Invest in data before AI. If you can’t measure it, you can’t improve it.

3. Involve operators from day one. They know things about your process that won’t show up in data.

4. Be patient. This took two years from start to current state. Quick wins are rare.

5. Pick partners carefully. They went through two technology partners before finding one that understood manufacturing, not just AI.

The bigger picture

This one company’s story illustrates what’s possible for Australian manufacturers. They’re not a multinational with unlimited resources. They’re a regional manufacturer with 180 employees who decided to invest in AI and made it work.

It required real investment—money, time, effort. It wasn’t plug-and-play. But the results are tangible and ongoing.

Stories like this are becoming more common. Australian manufacturing AI isn’t theoretical anymore; it’s operational and delivering value. The question for other manufacturers isn’t whether it’s possible, but whether the opportunity fits their situation.