Machine Teaching: Transforming Expert Knowledge into Million-Dollar AI Solutions

‍

I've hardly ever worked on a use case that wasn't worth at least a million dollars in a singular manufacturing line, or even a single machine for a year, if you can improve the KPI by just a single percentage point.

In manufacturing, those KPIs are typically things like throughput (how much you can make in an hour), efficiency (reducing resource usage), or waste management. Small improvements yield massive returns.

‍

This isn't an isolated opportunity. The industrial world is filled with these high-value problems waiting for the right solution.

‍

‍

The Challenge: Capturing Complex Expertise in Industrial Processes

‍

Real, high-value problems often deal with proprietary information - like you're talking about somebody's KFC special recipe. That makes it hard to talk about any real examples.

‍

One example I can share from past work involves helping PepsiCo R&D's Snack Foods division build an expert AI operator for making Cheetos. As it turns out, making Cheetos is actually pretty hard to do.

‍

They're made on specialized equipment called an extruder. This extruder had about 25 different control parameters that require adjustments every few minutes. It takes an operator 10 to 12 years to become an expert because of the adaptations and nuances you must learn about the machine.

‍

When we first approached this problem, we tried traditional AI techniques. But like many complex industrial processes, the solution wasn't simple.

‍

The real breakthrough came when I interviewed the expert operator. They explained that becoming an expert on this extruder required mastery of five very distinct skills that have to be used in very specific ways.

‍

These five skills also had to be arranged in a very specific workflow. It wasn't until we approached the problem through the lens of these distinct skills that we started making real progress.

‍

Think about it like learning to fly a plane. You're not going to learn takeoff by landing, and you're not going to learn to land by taking off. These are discrete skills that require different approaches, different expertise, and different mental models.

‍

That's the fundamental principle of machine teaching – breaking down complex tasks into discrete, teachable skills. It's not just about data. It's about structured knowledge transfer.

‍

It's like when you teach someone a jump shot. You don't say "there's only one way to shoot a basketball." There's a lot of different ways you can launch a ball towards a basket, but there's a very specific way that, if you practice within those confines, you're most likely to succeed.

‍

Real expertise doesn't come from simple pattern recognition. It's about knowing which skill to apply and when. That's what we did with the agent, and we were able to train the AI – in a few weeks – what would take a human 12 years to learn how to do at that level.

‍

The success we achieved with the Cheetos extruder wasn't just about the specific application. It revealed a fundamental limitation in how most companies approach AI for complex industrial problems.

‍

The Limits of Single-Technology AI Approaches

‍

Agentic AI companies tend to tell industrial firms, "We're the AI experts. We're building this specific AI system and then you can rent it or access it to do something interesting." Most of these AI agent systems specialize in one and only one kind of technology.

‍

One company builds its agents on large language models (LLMs). LLMs are great at language, but they're not great at other aspects of decision making in the same way that a hammer is great at driving nails, but it's not so great at turning screws.

‍

Another company builds its agents on optimization algorithms. Optimization algorithms are great when there's lots and lots of options, but there are certain things that they're not great at.

‍

But true success in this space requires a paradigm shift.

‍

The only effective way for high-value expertise to be enhanced with AI is to flexibly mix multiple technologies. This ensures you're using the best technology to teach each specified skill within the expertise.

‍

I don't participate in any conversation that claims one technology is better than another. I like all technologies in the same way I like hammers, nails, and screwdrivers – they're just different tools that are good at different things depending on the application.

‍

Needing distinct skills to successfully operate the Cheeto extruder isn't a scenario unique to making Cheetos. Almost every factory setting has similar situations.

‍

And the solution is similar as well. You need the right tools and methodology that enable you to build multi-agent systems that will revolutionize your processes, with your people, under your control.

‍

Thousands of Million-Dollar Use Cases

‍

This use case with the Cheeto extruder is just one example of thousands of similar opportunities across industry.

‍

I talked to a company that brought me a use case for intelligent autonomous agents in its processes. It was a perfect use case for intelligent autonomous agents using the right methodology.

‍

As I learned more about what the business did, I said "You must have a hundred use cases like this." Their response? "No. We have a thousand."

‍

And that was just one holding company with a set of manufacturing companies.

‍

The economic impact is massive, and it multiplies when you start deploying modular AI across these environments. But this requires a different paradigm than what's commonly pitched in AI today.

‍

Industry doesn't need more dashboards or simple automation – it needs intelligent autonomous systems that can adapt, make decisions within constraints, and operate with the expertise of your best people. And most importantly, they need to deliver measurable business value.

‍

In industrial settings, even a single percentage point improvement in a critical KPI translates directly into millions of dollars of value. Hundreds or even thousands of these high-value opportunities exist in almost every industrial organization.

‍

The technology exists today to capture this value by digitizing your best operators' expertise. The question isn't whether these million-dollar opportunities exist in your business – it's how quickly you can start transforming them into reality.

‍