Cobot Welding for High-Mix Shops: Who Programs the Next Part?

A welding cobot quote is an itemized document. The arm, the power source, the torch and consumables, the cart, the operator panel, the software seat, the install. Vectis publishes $95,000 to $140,000 for most of its systems and tells buyers that most of its customers report payback in two years or less. Hirebotics starts at $105,000 and says its welder runs right out of the box, no special training required.

Now look down the quote for the programmer.

He isn’t a line item, because he isn’t something the vendor sells. He’s a person your shop employs or doesn’t, and which one is true decides whether that payback figure describes your shop or somebody else’s. Two shops buy the same cell at the same price. One runs it. One parks it. The variable that separated them was on the payroll before the crate arrived.

Christian Lefebvre, a welding engineer at Pemamek, told The Tube & Pipe Journal in June 2026 that in many, many cases the cobots he saw were gathering dust. He sells welding automation for a living. The reason he gives starts with a number most shops can’t produce.

The number your mix turns on

Automotive built the robot welding playbook, and it works there because of one division. A body shop writes a program once and runs it across 100,000 identical parts. Divide the programming cost by 100,000 and it vanishes. Every robotic welding economic argument you have ever heard descends from that division.

Run it again on 12 brackets. Then on 30 of something else. Then on a one-off nobody reorders. Programming cost per part stops being a rounding error and becomes the dominant term, and nothing else on the quote moves it.

This is why job shops sat out robotic welding for a generation. A 2014 trade-press account of teach-pendant and offline programming put the break-even at roughly 100 units, below which programming time made the robot uneconomic. That figure is twelve years old and predates the cobot market almost entirely, so read it as history. It is the history that explains the present, because every product in this article exists to beat that threshold.

Which gives you one number to carry into every demo. Setup time per new part, on a part that resembles your work, timed end to end, from the moment the job lands to the moment it welds.

Vendors will offer cycle time instead. Cycle time measures the speed of the weld after someone has already done the programming, which is the part you are trying to evaluate.

You also need your own baseline, or the vendor’s number has nothing to sit against. Lefebvre’s method for getting one is a stopwatch. He has advised customers to stand next to their welders on the line and record arc-on time and cycle times by hand before automating anything. Manual welding in a normal shop, he says, runs 10 to 15 percent arc-on time. It sounds bad and it’s realistic.

That number is one half of an arithmetic you can finish yourself, and it’s the half nobody hands you. Take the setup time on a part like yours, multiply it by how many new parts you take on in a year, then multiply that by what an hour of programming costs you. That is what the cell spends on programming every year, and it appears nowhere on the quote. Set it beside the cell’s own annual cost and the two-year payback either holds or it doesn’t.

Four numbers run it. Two are yours and you already know them: how many new parts a year, and what an hour of that work costs. One you make the vendor prove on your part instead of the demo part. The last is on the quote. If the two of you can’t put all four on the table before you sign, the demo is the only evidence in the room, and the demo part was programmed by somebody else before it arrived.

Five answers, and where each hands it back

The industry has five answers to “who programs the next part.” Each one removes more of the programming than the one before it, which is the order they come in below. Each one also hands something back, and the part it hands back is never on the quote.

Seam tracking

Start with the answer that removes no programming whatsoever. A sensor reads the joint in real time and nudges the torch to follow it, so a slightly imperfect program still lays a good weld. Valk Welding’s ARC-EYE is a capable example, though it’s bound to Panasonic controllers and Valk has no North American branch. Servo-Robot sells laser seam tracking out of Wauwatosa, Wisconsin, which is the one you can buy here.

Laurent Gravel of Groupe Gravel, the job shop in the last section of this piece, put the limit better in 2016 than any vendor has. Offline programming, he said, “doesn’t change the physics of welding. If a job required seam tracking before offline programming, because of inconsistencies in the weld joint, the job will continue to require such tracking.”

Tracking corrects a path. It does not write one. Everything it fixes sits downstream of a program someone already had to produce, which is why it hands the job straight back to whoever wrote it. The correction has a ceiling too, and Lefebvre draws it himself. Adaptive welding is “really good” now, he says, “but there’s always a limit.” On a joint whose gap wanders from zero to a quarter inch, “something that’s a bit more challenging,” he would “likely just keep a human there.”

Cobot plus hand-guided teaching

The welder drags the arm through the weld path by hand, sets a few parameters, and the cobot repeats the motion. No robot language, no teach-pendant coordinates. This is the approach that made cobot welding a category, and the US channel for it is dense. Hirebotics, Miller, Lincoln Electric, ESAB, and Vectis all sell hand-guided cobot welders domestically, and Fronius (Portage, Indiana) and Cloos (Schaumburg, Illinois) sell their cells here directly.

The promise is that programming disappears. What disappears is the robot language, which is a real gain and a smaller one than it sounds. Every new part still needs a person to walk the arm through it, decide the sequence, and judge the result. Hand-guiding trades a skill for a duration. That person needs no robot training now, and he needs the same hour he always needed, once per part, forever. It works while paths are short and repeat enough to earn the teaching time back, and it runs out of room exactly where a high-mix shop lives.

CAD-based automatic programming

Here the promise gets specific, and for once both numbers are public. Upload the 3-D model, and software writes a collision-free robot program from the geometry. Novarc’s NovPlan, which is Verbotics underneath, advertises cutting programming time by up to 90%, and Verbotics markets a runnable program “within minutes.”

Novarc’s own worked example, on the same page as that claim, is a part with 175 weld paths that would normally take over 35 hours to program, ready in 1.5 hours. Both numbers are vendor-published, both are true, and both describe the same product. The headline is what the software does to a weld. The worked example is what a person does to a part, and it still costs most of a shift.

There’s a boundary underneath it that matters more than the arithmetic. CAD-based programming works where the parts are already standard CAD objects. AGT Robotics says its Cortex software batch-processes an entire building from a CAD file in one click, and it runs on structural steel through Tekla and SDS/2, where beams and connections are parametric library shapes. A 2022 case study of Petrosmith, an oil-and-gas vessel fabricator running Cloos robots with German “Moses” software, worked for the same reason: flanges, manways, couplings, and vessel heads have industry-standard designs, so the software builds weldments from a library.

The author of that case study wrote his own caveat, and it’s the line every buyer here should read. For industries with very low lot sizes and very high mix that aren’t based on industry-standard CAD components, the software “may or may not be the best solution.” A 40-person shop welding a one-off bracket from a PDF is precisely that case. The technology that eliminates programming has landed in the sub-segments that were already the most standardized, and not in the high-mix chaos that needs it most.

And where it does land, it relocates the programmer rather than deleting him. The Petrosmith author is explicit that the knowledge sits with a programmer who “certainly needs to be a skilled welder/robotic programmer,” and that one such person with one software seat can write programs for several robots while a less-skilled operator loads parts. For a shop that already employs him, this is the best deal in the article. For a shop with zero of them, it’s a job posting.

AI and vision

The largest removal anyone offers without leaving the category. The robot scans the real part with vision or laser sensing, builds its own weld plan, and adapts as it goes. Path Robotics and Novarc are the names most associated with it, and Inrotech’s vision system now sits inside Lincoln Electric, which bought the company in 2024.

The promise is that the program writes itself, and the defeat sits upstream of the software entirely. Dan Colvin of ESAB put it in June 2026: a manual welder adjusts for every part, and an automation system expects the same parts in the same place in the same profile every time. Colvin’s summary is blunt. The number one problem people have when they say let’s automate is poor fit-up.

Which is where the largest removal hands back the largest job. Vision narrows the fit-up problem and does not repeal it, so the work moves upstream into fixturing, tacking discipline, and holding tolerances you may not hold today. Somebody has to own that, and it’s the same engineer the software was supposed to make unnecessary, now standing further up the line.

Ask this category for a setup time per new part and notice what comes back. There isn’t a published figure, from any of them. The answer with the largest claim on your programmer is the one answer that has never said what it does to your number.

Handheld laser

The only honest answer in the set, and it wins by not entering the contest.

A handheld laser welder trains a competent welder to acceptable welds in a few hours, runs 4 to 10 times faster than TIG on thin-gauge joints, and costs roughly $7,000 to $15,000 for a mid-range setup. There is no program, so there is nobody to hand the programming back to. The full case, with thickness tables, real costs, and the Class-4 laser safety burden that comes with it, is in the handheld laser guide.

Put the money side by side. A cobot cell is $75,000 to $140,000 in vendor-published pricing and needs a programming capability the shop may not have. A handheld laser is roughly a tenth of that and amplifies a welder already on the payroll before lunch.

It also gives up more than any other answer here. The laser tops out around 8mm, it carries a Class-4 safety load with a designated laser safety officer and eyewear for everyone in the room, and it welds one joint at a time with a human holding it. It automates nothing. For a shop running thin-gauge work that says it can’t hire welders, that trade is worth making, and the cobot has to beat it on the numbers rather than on the brochure.

The objection: even MEC is still evaluating

The obvious answer to all of this is to go get the programmer. Hire him, train him, and the payback figure comes true.

Mayville Engineering is the test of that answer, and it complicates it.

MEC is a contract fabricator north of $500M in revenue that has topped The Fabricator’s FAB 40 for over a decade. It has robots. It has programmers. It sorts work by quantity with a discipline most shops never reach: low-quantity jobs go to manual stations, high-quantity jobs go to traditional robots, and the cobot’s job is everything in the middle. David Higgs, MEC’s VP of Business Excellence, describes an operating model where one person handles the fixturing, tacking, and programming on a cobot, against a traditional robot cell that might need one person programming while another runs it. The Fabricator’s account is candid that the cobot can’t touch the speed of a traditional cell.

And in 2024, with all of that in hand, MEC was still “deep in the middle of the evaluation process.”

Read that as a failure of the thesis and it looks fatal. If having the capability doesn’t settle the question, what is the capability for?

Read what MEC is actually doing and the answer is narrower. MEC can sit in evaluation because MEC can run the arithmetic. It knows its arc-on time, its quantity bands, its fixturing cost, and roughly what a cobot does to each. Its “not yet” is what those four numbers look like when somebody actually produces them, on a specific job, and finds they don’t clear.

So the programmer’s first job happens at the quote, months before the cell exists. He’s the person who can produce your two numbers and check the vendor’s two, and MEC is what it looks like when a shop can do that and the answer comes back no.

What to ask before you sign

The technology is available. Every category above has a US channel, on the hardware side and the software side, and any vendor who implies scarcity is selling urgency. What a shop cannot buy off the shelf is the ability to tell a real fit from a slick demo.

Every gate below came out of a section above. The setup-time gate is the arithmetic from the top of this piece. The programmer gate is what the CAD section found when it went looking for him and turned him up in an office. The standard-CAD-objects gate is Petrosmith’s own caveat. The fit-up gate is Colvin’s number-one problem. None of these are hard questions. They are only unanswerable if nobody on your side can produce a number.

A realistic first cell looks like this.

The labor case for automation, that it augments welders instead of replacing them, is made in detail in Why the US Shipbuilding Revival Is a Welding Capacity Problem.

The shop that beat it

Groupe Gravel is a 25-person shop in Marieville, Quebec, and it is the one job shop the trade press has found genuinely running one-off robotic welding. The trade press presents it as atypical, which is the useful part.

It worked because the shop had spent years offline-programming its CNC milling and turning centers in Mastercam. Employees programmed those machines to run one-off parts all the time, so when robotic arc welding arrived, they asked why the same thing couldn’t be done for welds. When jobs showed up without 3-D models, and many did, Gravel bought a 3-D scanner and made its own.

Faced with the exact boundary that stops CAD-based programming, the shop that beat cobot welding responded by acquiring another capability. That is the shape of the one winning story the trade press has found, and there is no version of it where the machine supplies the answer.

Gravel brought the capability in with it, and the robot was the last thing it bought.

Which leaves one question before you sign anything. The cell can probably weld your part. Can you name the person who programs part number 13, and can that person produce your two numbers by Friday?