Manufacturing in the U.S. is seeing its fastest growth in more than four years, driven in large part by rising demand for AI infrastructure. That expansion is arriving alongside severe supply shortages and higher input costs.

Manufacturers look to AI to increase capacity and operate more efficiently. But routing every request through the largest available model creates another capacity problem inside the technology itself: higher costs, greater compute requirements, and an architecture that becomes difficult to scale.

How general-purpose models can create inefficiencies

When companies add AI to a workflow, the default is often the largest model available. Yet many manufacturing tasks are narrow, such as finding a requirement in a technical specification, identifying a maintenance procedure, or calculating a value from a graph. A frontier model may handle each one, but using it for everything is a brute-force approach.

Think of a company where the CEO answers every email, manages the finances, and handles every HR task. The CEO may be capable of doing each job, but the company would be slow and expensive to run. Companies rely on specialists for a reason.

AI should work the same way. One capable model can understand a request and coordinate the workflow. It can then route each part of the work to a smaller model designed for that task, whether that means interpreting a table, retrieving a specification, or completing a calculation.

The case for specialized models  

Manufacturers should approach AI architecture with the same intention they bring to a production line. Different jobs require different capabilities, and an efficient system assigns each task to the right tool. A general model can manage the interaction, while specialist models complete defined parts of the process.

Industrial documents are also more complex than plain text. Technical manuals contain tables whose structure determines how information should be read. Schematics depend on relationships between lines, labels, and components. A graph might need to be reconstructed as a formula before an answer can be calculated. Flattening these documents into text can remove their meaning.

At octonomy, we use a system of models because an industrial question rarely represents one clean task. It might require interpreting a document, finding a specification, completing a calculation, and validating the result against the source. Dividing that workflow makes each step easier to test and improve.

I learned this discipline early. I started programming in 1981 on an 8-bit computer with 64 kilobytes of memory. With so little capacity, you had to consider what the computer needed before writing code. Engineers now have abundant computing power, making it easy to reach for the largest tool without examining the architecture.

That choice has consequences beyond cost. Larger models require more compute, increasing demand for GPU capacity, energy, and cooling. When a smaller model can reliably perform a narrow task, sending it instead to a frontier model consumes unnecessary resources. Throughout thousands of requests, that difference can determine whether an AI system remains practical.

Sometimes ‘I don’t know’ is the most reliable answer

The biggest risk of overreliance on larger learning models is confidence. Most AI systems are trained to produce an answer, even when they don’t have the data or context needed to make an informed decision. Rather than admit uncertainty, these systems will confidently present its guesses as fact. A system that gets most things right but occasionally states a wrong answer with total confidence has potential to be more dangerous on a factory floor than one that simply says, “I don’t know,” and requires a person to step in.

A specialist model has a defined role and clear boundaries. That makes its performance easier to test. But specialization alone doesn’t eliminate hallucinations. A reliable system must ground its response in approved source material, validate the answer, and stop when the evidence is insufficient.

In those situations, “I don’t know” is the correct response. The question can then be routed to an expert or another verified source. Manufacturers should evaluate whether AI provides repeatable, traceable answers and recognizes when human intervention is required. A 100% response rate may be a warning sign if the system never admits it can’t validate an answer.

Models made to handle real-life floor operations

Many manufacturers put their trust in larger models after seeing a handful of successful demos show what the model is capable of. But most pilots rarely integrate the real-life situations and chaos of a factory floor, and so errors aren’t caught until after manufacturers scale dozens of AI-driven workflows into their production process. Each new application increases the demands on infrastructure, increases the volume of information being processed, and creates more opportunities for a poorly matched model to slow the system down.

Using specialized capabilities across the production floor can prevent these bottlenecks and help manufacturers scale more efficiently. Instead of treating every problem as a reason to deploy another large general-purpose model, the system can direct work to the appropriate specialized model, use the resources required for that task, and bring a person into the process when the situation calls for judgment.

Manufacturers already apply this principle throughout their operations. They select machines based on their production needs, assign workers based on their expertise, and design production processes around the system’s constraints. There’s no reason not to hold AI to the same standard.

As manufacturers are expected to produce more output at a faster pace with less people, AI will become an increasingly important part of the production system. When determining which AI tools to invest in, manufacturers need to focus on building an AI system that can scale practically. Specialized models provide a way to distribute the workload across the production floor, use resources more efficiently, and reduce the risk of AI-driven bottlenecks, waste, and unplanned downtime.