What if the missing ingredient in industrial AI isn’t more data, but better teaching?

‍

That’s the question Composabl CEO Kence Anderson answered in a recent episode of Manufacturing Happy Hour. He pointed to a fundamental disconnect between how AI is typically built and how manufacturing actually operates: most AI systems are designed to perceive and predict – not to act.

‍

Check out the full episode: AI Agents for Manufacturing 101

‍

Action – making decisions in real time, under constraint, with imperfect information – is where value is created on the manufacturing floor. And when AI models are trained only on historical, system-level data, they miss an entire set of inputs that’s never been captured in a database.

‍

While historian data and machine learning models can recognize patterns – and even forecast outcomes – they can’t explain why a seasoned operator adjusts a process when a machine sounds different, or why inputs react differently on humid days. They don’t understand the edge cases, the exceptions, or the rules-of-thumb that keep production on track.

‍

These gaps point to a different kind of data: human expertise that isn’t recorded anywhere but governs how things actually get done.

‍

Imagine bringing in a talented engineer to manage a new process. Smart, experienced, technically capable. But without an expert to teach them the nuance of that process, they’ll spend weeks relearning what others already know. The raw intelligence is there, it just needs to be properly guided.

‍

The same holds true for machines. The fastest path to intelligent decision-making isn’t more exposure. It’s better instruction.

‍

The true impact of AI in manufacturing comes from giving engineers the ability to transfer what they know into multi-agent systems that can reason and act in real time. Intelligent AI won’t emerge from data alone. It must be purpose-built and taught by experts.

‍

The knowledge is already there. The challenge now is to capture and codify it before it disappears.