I own a precision manufacturing firm. It is the kind of business the software industry means when it says "legacy": decades old, profitable, physical, and running on knowledge that lives in people's hands. Machines have been better than humans at the actual cutting for seventy years, and nobody on a shop floor has ever mourned that, which gives manufacturing people a calmer view of automation than most commentators manage. We have already lived the story everyone else is bracing for. The work went to whatever did it best, and the judgment moved up a level, and the trade survived by holding the part only a person could hold: the call on whether the part is right.
So when I evaluate what AI agents change for a business like mine, I am not asking whether the machines are coming for the machining. That happened before I was born. I am asking where the reading and the errands live in an industrial business, because that is what the new machines actually do, and the answer surprised me: an enormous share of a manufacturer's white-collar hours is reading.
Take quoting, the front door of every job shop. A request for quote arrives as a package: drawings, a spec, tolerances, material callouts, quantities, terms. Someone senior reads all of it, because the expensive mistakes hide in the notes, the revision history, the tolerance stack that makes an easy-looking part hard. Shops lose winnable work constantly for one reason: the package sat in a queue because reading time is scarce and quoting is unbilled. A machine reader changes the economics of that queue completely. Every package read the hour it arrives, every spec extracted, every drawing note flagged, the tolerance that kills the margin surfaced on page forty before a human spends a minute. The human still prices it. Pricing is judgment, the shop's actual product, the accumulated scar tissue of every job that went wrong. But the reading that gated the judgment stops being the bottleneck, and the shop that quotes in a day beats the equally good shop that quotes in a week more often than quality ever decides between them.
Now turn it around, because the buyers are getting machine readers too, and this is the part most manufacturers have not registered. Purchasing agents, the human kind, are starting to send the software kind to find capacity: what shops within five hundred miles hold this tolerance in this material at this volume. That agent is not going to call you. It is not going to tour the floor, notice the clean aisles, or shake anyone's hand, and every instinct of industrial sales, which is built on exactly those things, goes quiet. The agent reads your website. And here is the uncomfortable audit: most job shop websites, mine included until recently, are a photo of a building, a phone number, and the word "quality." To a machine reader that page contains nothing. Capabilities in plain terms, equipment lists, materials, tolerances, certifications, capacity signals: the shop that states these in machine-legible form is, to the buying agents, the only shop on the block with a sign out front. Everyone else is an unmarked building. This is beginning to be solved above the level of the individual shop: capacity networks are forming that hold manufacturers' capabilities in exactly the structured form the buying agents consume, a kind of directory district for an industry of unmarked buildings, and being listed in one, accurately and in detail, will shortly matter more than the trade-show booth ever did.
The same reallocation runs through the rest of the office. Certification audits are reading. Compliance paperwork is reading. Chasing a late supplier through an email thread is reading plus an errand. First-article inspection reports, customer flow-downs, the terms buried on page six of a purchase order that conflict with your quote: reading, all of it, and every hour of it currently comes out of the two or three most experienced people in the building, who are also the constraint on everything else. Freeing the senior people's reading hours is not a productivity garnish in a small manufacturer. It is the difference between the owner working on the business and drowning in its paperwork.
What I am deliberately not doing is putting the machines anywhere near the deciding. The quote still gets priced by a person. The disposition on a borderline part is a person. The call to accept a job that smells wrong is a person, informed faster and better, deciding the same way. The shop floor settled this division a lifetime ago and it holds: capability migrates, accountability does not.
And the new software worth letting onto a floor has a tell, the same tell as everywhere else in this series: it shows its work. There is a class of machining software arriving now that reads the blueprint and derives the cut from first principles, physics rather than guesswork, and hands the machinist not just a number but the formula behind it, the constants that went in, and the limit the cut was scored against. That is the bargain done right. The reading and the arithmetic move to the machine, the derivation stays visible, and the call stays a call, because a number you can audit keeps the machinist the judge, and a number you cannot is a lookup table wearing a lab coat. Follow that class of software to its destination and you reach the ambition my own companies now state plainly: machines that read the blueprint and build the part. The reading and the arithmetic are nearly there. The fixturing and the holding are still craft, and will be for a while. And the acceptance, the signature that says this part ships, is not on any roadmap to transfer, because it cannot be. The autonomy is coming for the making. It has no purchase on the answering.
If you run an industrial business and want the Monday version: pick the reading that gates your money, which is almost always quoting, and put a machine reader in front of it this quarter. Rewrite your website as if the next visitor cannot see photographs and can only extract facts, because that visitor is already coming. And leave the wisdom where it is. The oldest legacy systems in the country are its machine shops, still running, still load-bearing, and they know something the software industry is only now learning: you route the work to whatever does it best, and you never, ever route the responsibility.