In a metal fabrication shop AI arrives in six places before anyone decides anything: the estimator who turns a drawing package into a price, the programmer who writes the cutting and bending programs and nests the sheets, the scheduler who sequences the jobs across the laser, the brake, the weld cells and the paint line, the quality group that writes up inspection and keeps the material certificates, the buyer who chases steel and outside processing, and the shipping desk that answers where the order is. The owner's job is not to pick the tool. It is to decide three things before a tool touches a customer's drawing, a price or an inspection record: who in the shop owns the question, what may never go into an outside tool, and what the shop will never let a machine do alone. This article is for the owner of a job shop, a contract fabricator or a family-owned metal shop with a few dozen to a couple of hundred people, and it is a child of the broader piece on AI for a manufacturing company, which covers the owner's decisions for any maker.
I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them. I am a lawyer by training. Now I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies through 013 Labs, and fabricators are a group where a tool's mistake lands on a customer's part, in a customer's assembly and in a customer's audit file at the same time.
Where AI already is in your shop
Ask the estimator what the free tools on the office computers are doing this week and the honest answer is: reading the customer's drawing package, pulling the material, the thickness, the bends and the tolerances into a list, and drafting the quote letter with the lead time. Ask the programmer and it is the first pass at a nest and the work instruction for the welder. Ask the scheduler and it is the narrative to the customer who wants the date moved. Ask quality and it is the first draft of the inspection report and the corrective-action write-up. Ask the buyer and shipping and it is the supplier chase and the packing list. The quoting software and the machine vendors each switched on an assistant with their last update. None of it was approved. Some of it has already sent a customer a price with nobody's name on it, and some of it has carried a customer's drawing, a specification or a controlled file onto a vendor's server under no agreement at all.
The three decisions, in the order a shop makes them
Who owns the question. Not the IT provider, the software vendor or a new hire. In most shops it is the estimating manager, the operations manager or the quality manager, as they are the people the floor already asks when something new shows up. Give that person a written scope of half a page: what they may approve alone, what comes to the owner, when the page is reviewed. The article on whether your IT company can handle AI draws the line.
The data rule. One sentence. For a fabricator it usually reads: no customer drawings, specifications or files of a part, no file that is export-controlled or marked customer-confidential, no unit costs, routings, material prices or quote history and nothing about an employee go into a tool the shop has not approved, and no tool is approved to hold customer information without written vendor terms that say where the data is kept, who owns what the tool produces and what the vendor may do with what it is given. The owner writes it. Customer drawings under a nondisclosure agreement can end a relationship in one email, and export-controlled work is a matter for counsel, not a free tool; the shop needs to know which customers send such files and keep them out of anything unapproved.
The never-build list. The three things the shop will never let a machine do alone. For a fabrication shop the list writes itself: no quote or price is sent to a customer by a tool, or without a named estimator having checked it against the drawing; no first-article report, inspection record or material certificate is signed without a named inspector who did the measuring or the checking; and no customer drawing is uploaded to an outside tool without a written rule that says it may be. Everything else is a question of fit. These three are not.
What each department can do with it, and what it cannot
Estimating. A tool can list the parts, materials, thicknesses and finishes from a drawing package and draft the quote letter from a price the estimator has built. It cannot know the steel price moved last week or that the customer's tolerance is tighter than the laser holds. The estimator prices; the tool prepares. Programming and nesting. A tool can draft work instructions and setup notes. The program that runs on a machine is checked by the programmer, proven on the first part and owned by a person; a tool's guess at a bend sequence can scrap a sheet. Scheduling. A tool can write the status narrative to the customer. It cannot know about the machine that went down at ten or the rush job just accepted, so a person who knows the floor promises the date. Quality and certificates. A tool can lay out an inspection report from measurements already taken. A first-article report or a material certificate says a qualified person checked something, and a statement nobody checked is a certification handed to a machine. Purchasing and shipping. Requests for quote, the chase of a late shipment and the packing list are repeatable work; the choice of supplier and the purchase order stay with the buyer, and anything that reaches a customer's dock is looked at by a person.
What fits at fifty people, and what does not yet
The first project is work your people already do by hand, the same way, every week: the quote letter from a price already built, the schedule narrative, the work instruction for a repeat part, the supplier request for quote. The article on picking your first AI project has the five tests. What does not fit yet is anything that decides: pricing from history, promising a date against a floor the tool cannot see, releasing a program, signing off a lot. A tool can prepare any of those for a person. It should not finish them. For the schedule piece, can AI do production scheduling explains what a tool can prepare and what the planner must still decide.
The bank, the insurer, the big customers and a buyer will ask
The AI question reaches a fabricator from four directions that have nothing to do with technology. The largest customers' supplier questionnaires and quality audits now carry a line about how the shop controls tools that touch drawings, quotes and quality records. The bank's annual review carries one too; the article on what to tell your lender about AI covers that meeting. The insurer's renewal form asks whether automated tools produce work or records in the shop's name, because a part that fails in a customer's assembly is a claim, and the carrier wants to know who checked it. And a buyer, if the shop is ever sold, asks it in diligence, because owner-led fabricators are bought into platforms; the article on what buyers ask about AI in due diligence has that version. One dated page answers all four, and it is the page the three decisions above produce.
The article on AI for a construction company covers the contractor who installs what you make.
Picture a family-owned fabrication shop of eighty people with a laser, a bending line, a welding department and two large customers in regulated industries. The estimator has been running customer drawing packages through a free tool on his own login to pull the materials and bends into the quote, and a quality technician has been using it to tidy inspection reports. The owner finds out when the larger customer's supplier audit asks which systems have handled its drawings and records in the past year and nobody can answer, and then finds a first-article report with a measurement the technician never took. The fix takes two weeks and one uncomfortable call. The estimating manager is named as owner, with a half-page scope. The owner writes the data rule: no customer drawings, no controlled or confidential files, no routings or quote history into an outside tool, and no tool without written vendor terms. The never-build list goes on the page. The free tool is replaced by the company plan, the drawings are deleted from the vendor's account with the deletion confirmed in writing, the inspector redoes the report, and the owner calls the customer's quality manager with the page in hand. This is a hypothesis, a composite and not a client.
The one page
What is in use, by department and task, including the free tools and the assistants vendors switched on. Who owns the question, with the written scope. The data rule in one sentence, with the vendor terms listed and a line on where the data is kept and who owns what each tool produces. The never-build list, the three items above. The review date and who signs. The owner signs it, and the quality manager keeps it for the next supplier questionnaire. The board-briefing template has the same page in a form a fabricator can fill in.
Mistakes that repeat
Letting the software vendor switch it on. An assistant arrives in an update, defaults to on, and a quote carries a lead time nobody checked. The owner decides, feature by feature.
Customer drawings in a free tool. The one mistake that can end a customer relationship in a single email, and the one a supplier audit finds. The data rule exists for this.
A quote the tool priced. The tool finds the last similar job and adjusts it. It does not know the material price moved or the customer changed the tolerance. A person prices; a tool prepares.
An inspection record nobody measured. A tool can make a report look complete. A named inspector measures and signs, every time.
Terms nobody read. The vendor's page lets it keep what it is given, and the shop finds out when a customer asks for proof of deletion.
No review date. Customers change their audits and tools change their terms. A page without a date is a page nobody opens.
Frequently asked questions
Should a metal fabrication shop use AI?
It already does. The decision worth making is who owns the question, what may never go into an outside tool, and the three things the shop never lets a machine do alone: a quote or price sent by a tool unread, an inspection or first-article record signed without a named inspector, and a customer drawing uploaded to a tool without a written rule.
How can a fabrication shop start with AI safely?
Start with the page, not the tool. Name the person who owns the question, write the data rule with customer drawings first and written vendor terms as a condition, and put the never-build list on paper. Then pick one repeatable job and run it inside the rule for a month.
Who should own AI in a job shop?
The person the floor already asks when a new tool shows up, usually the estimating manager, the operations manager or the quality manager, with a written scope of half a page and a review date. Not the IT provider and not the software vendor.
Can AI quote jobs and write programs for our machines?
It can draft a quote letter from a price a person has built and help with work instructions. It should not set the price or release a program a programmer has not checked and proven. The estimator's name is on the quote.
What about customer drawings and export-controlled files in these tools?
They are first in the data rule. No customer's drawings or specifications, and no file that is export-controlled or marked confidential, go into a tool the shop has not approved, and no tool is approved to hold customer information without written vendor terms. Export control is a matter for counsel; in general terms, keep such files out of anything unapproved. If drawings have already gone into a free tool, delete them from the vendor's account and get the deletion confirmed in writing.
Why does my bank, my insurer or my biggest customer ask about AI?
Because each carries the risk when a price, a part or a record a tool touched is wrong, and wants to know whether a named person stands behind it. One dated page with names on it answers all of them.
Not sure where you stand? Take the 3-minute AI readiness scorecard.