Manufacturers Lead in AI Adoption. Most of It Isn't AI.

Manufacturers lead the mid-market in one category of AI adoption, according to Kaufman Rossin's latest survey: automation and RPA. The same survey has them trailing on generative AI, the chatbots and drafting tools that get the stage time at every conference. Repeat the first part in a boardroom, and people nod. It sounds like manufacturers have quietly gotten ahead of everyone else.

Look closer at what that leading category is actually doing, though, and most of it isn't AI at all.

Classic robotic process automation (RPA), software that runs a fixed process the same way every time, is built to mimic a person's repetitive clicks and keystrokes across systems that don't talk to each other. It covers a lot of ground: moving data between screens, generating reports, routing a routine request to the right queue. One familiar version in finance is matching a purchase order to a receiving report to a vendor invoice and posting the result. Whatever the specific task, the mechanism underneath tends to be the same: rules somebody wrote down. No model, no learning, no probability, anywhere in the chain.

Calling that "AI" is a stretch, and "AI-powered" prices better in a proposal than "RPA" does, but the industry has two real hooks to hang the label on. Most RPA platforms bundle in one genuinely AI-driven piece, usually a model that reads a document and pulls structured data out of it, alongside the purely rule-based workflow built around it. Vendors sell the whole bundle as "AI-powered automation," and the name technically points at something real. It's just pointing at the small piece that involves a model and letting it cover the much larger piece that doesn't.

The second hook sits inside the buyer, not the vendor: whichever team and budget line owns a project decides how it gets reported, and an automation initiative run by an "AI and digital transformation" group gets counted as AI adoption the next time somebody surveys the industry, no model required anywhere in it. That's likely part of what's sitting inside the Kaufman Rossin numbers themselves. Not fabrication. Two real things, a model doing real work in one narrow piece, a reporting category built around a team's name, both lending their credibility to a much larger deterministic system.

What matters more than catching the industry in a labeling trick is why this category delivers real value and finishes what it starts. It's earning that outcome because it's deterministic, not in spite of it. A rule that behaves identically every time is auditable. You can trace exactly why it did what it did, which is precisely what you want from anything touching money. It's also nearly the only kind of project many manufacturers can actually pull off: the same survey found fifty-five percent name legacy ERP integration as their top barrier to AI adoption, against a forty-one percent mid-market average, and a contained, one-task rule clears that bar in a way a sweeping platform never does. That's a different bar than "intelligent." In some ways, a higher one.

That extraction piece, the one doing the real lifting inside the "AI-powered" bundle, earns its place. Pulling structured data off an invoice that doesn't arrive in a consistent format, a clean PDF from one vendor, a scanned image from another, an email body from a third, is a job no fixed rule can handle. A model belongs there. Everything downstream of it, the matching and the posting, goes back to being deterministic the moment the data lands in a structured field.

That split is the real filter, and it's sharper than "is this AI." For any task you're considering handing off, ask whether you're looking at a fixed rule or a judgment call on messy input. A fixed rule needs automation, not AI. Don't let a vendor price scripted, rule-following work as if it were intelligence. A judgment call on messy input is where a real model belongs, scoped narrowly, with someone still checking its work, because probabilistic tools are wrong sometimes in ways a deterministic rule never is.

Your skepticism of the vendor pitch was already earning its keep. Keep pointing it at the real question: not whether you should be "doing AI," but which of your own high-volume tasks are fixed rules waiting to be automated, and which are messy enough to need a real model with a person watching it. The label on the proposal tells you what the vendor wants to sell. It doesn't tell you which one you should be looking at.