Abstract
The usual explanation for why industrial automation stalls is cost of capital: robots are expensive, integration is expensive, compute is expensive. That explanation survives because it is easy to put in a budget. It is also mostly wrong. What actually defeats automation on a plant floor is the recurring engineering cost of changing it. Every new part, new fixture, new product, new line speed and new layout triggers another round of specialist work, and that round has no line item, no owner, and no end. This paper names that recurring cost the reprogramming tax, argues that the current wave of industrial AI investment is aimed squarely at it rather than at raw capability, and makes the case that a maintenance organization is already paying the same tax in its parts room, its spares strategy and its standards, whether or not it owns a single robot.
The claim
Automation is bought once and changed forever.
The purchase is visible. It goes through capital approval, it has a number, somebody signs it and somebody else defends it a year later. The changing is invisible. It arrives as a series of small engineering engagements, each one individually defensible, none of them ever added up. A fixture gets modified. A program gets rewritten because the product changed. A vision system gets retrained because the lighting in that corner of the building changed when they replaced the roof.
None of those are failures. They are the normal metabolism of a plant that still makes things people want. The problem is that classical automation charges a specialist's fee for each one, and the fee is charged in the scarcest currency the plant has, which is not money. It is the availability of the one person who understands that cell.
That is the reprogramming tax. It is a recurring cost that behaves like a capital cost, and it is paid in expertise rather than dollars, which is exactly why it never shows up where anyone is looking for it.
What the vendors are actually selling in 2026
Two things landed in September 2026 that are worth reading carefully, because the marketing is about capability and the actual product is about this tax.
Skild AI reported crossing 100 million dollars in annual recurring revenue within ten months of first commercial deployment, with more than 60 paying customers, naming NVIDIA and Foxconn, Sumitomo Wiring Systems and Mitsui among them.[1] The revenue number is the headline and it is the least interesting part. The interesting part is the mechanism the company puts underneath it. Its S1 model uses what the company calls in-context learning, and the description is blunt about what problem that solves: "An operator can record a new demonstration and give it to the robot as a prompt, without updating the model's weights."[1] The company frames this explicitly as adapting to process changes without new datasets or retraining cycles.
Read that as a millwright rather than as an investor. The pitch is not that the robot is smarter. The pitch is that when the task changes, the person who changes it is an operator with a camera rather than an engineer with a schedule three weeks out.
Vention debuted MachineAgent at IMTS 2026 the week before, an agentic layer that runs on plain language across design, programming and operation: generating cell layouts, producing industrial automation programs, and analyzing operating data from deployed machines for troubleshooting.[2] The release is direct about the target, describing physical AI as helping manufacturers tackle "complexity, variability, and the gap in specialized engineering expertise."[2]
Complexity, variability and the gap in specialized engineering expertise. That is a three part description of the reprogramming tax, written by a vendor, in a press release.
Both companies are selling the same thing to the same wound. Neither is promising a machine that does more. Both are promising a machine that costs less to change.
Treat the performance claims as vendor claims. Skild's revenue figure is company reported and its success rates are company reported. Vention's lifecycle coverage is a product description, not a field result. What is evidentially solid here is not that either product works. It is that two companies with real customers and real money have independently concluded that the bottleneck worth attacking is reconfiguration cost. That is a market reading the same problem a plant reads, and it is worth taking seriously even before a single independent deployment result exists.
Twenty kinds of gearbox for the same five-horse conveyor
You do not need a robot to pay this tax. Here is a plant paying it in steel.
Water bottling plant. They broke down one time and had no spare gearbox to put in. I asked why we could not just pull one off another conveyor. Because it is a different gearbox. I looked at the two conveyors and they were basically identical. Why do we have different gearboxes on identical conveyors? Nobody knew. This is what they were sent with. So we waited for a gearbox to come in, with the line down.
Afterward I went through the whole plant. Everything was five-horse motors and gearboxes, and there were something like twenty different kinds of them. I put in work orders to consolidate the lot down to one type, or at least a limited set of types. Instead of keeping hundreds of different spare parts on hand we kept several. It worked like gangbusters. Less inventory sitting there, and we always had the spare, because everything was the same thing.
The thing to notice is that the variety was never a decision. Nobody sat in a room and chose twenty gearbox types. It accumulated one shipment at a time, over years, because nobody owned the standard. Every individual purchase was defensible. The aggregate was indefensible, and it stayed invisible until the day a machine was down and the identical machine beside it could not help.
That is the same shape as the reprogramming tax exactly. A recurring cost with no owner, accumulated one reasonable decision at a time, invisible until the moment it is expensive.
There is a second order to it that is worth stating because it sounds contradictory: standardizing cut the inventory value and raised parts availability at the same time. Coverage stops being per machine and becomes per type, so a handful of spares covers the plant better than hundreds covered it before. The caution that goes with it is that each individual spare now sits on the shelf longer, and a spare sitting still is a spare that can die standing still. Consolidation without storage discipline just changes which failure you get.
Any AI vendor promising to absorb variability is making the same trade. They are proposing to hold the variation in one place instead of twenty. Ask them where the new failure goes.
The unscrambler, and where the capability to change things actually lives
Late in my ten years at that bottling plant I moved into R and D, the group that dictated the machines. I was tasked with making the unscrambler faster, the machine that takes bulk empty bottles and gets them upright on the conveyor. It was doing around forty to fifty bottles a minute, nowhere near what production needed.
I took the machine to the shop and did a lot of modifications. There was a ramp inside, and I built a whole second ramp so there were two of them, and I added airlines to blow off bottles that were oriented the wrong way. Got it to around 100 to 110 bottles a minute. Still not quite enough, so I phoned the manufacturer. Italian company. Phoned Italy, finally got hold of someone who spoke English.
Their man told me flat out: that machine will never do more than eighty bottles a minute.
Holy shit. Okay. Well, we are doing like a hundred and twenty now.
Silence. Then: how did you do that?
You know what? I think I am just going to keep that to myself. And I hung up.
I want to be careful about what that story proves, because the obvious reading is wrong. I did not speed the machine up. I made it more efficient. The manufacturer's eighty was a real ceiling on the variable they were tuning, which was cycle speed. I attacked a different variable, which was yield per cycle. The second ramp doubled the feed path and the air jets rejected wrong-way bottles early instead of letting them clog or waste a slot. The cycle rate could stay exactly where it was and throughput still more than doubled, because the machine stopped squandering its chances. Both ceilings were real. They belonged to different variables.
Here is why that belongs in a paper about reprogramming cost. The OEM knows the machine as designed. It does not know the machine as deployed. The people in Italy who drew that unscrambler could not conceive of the number a millwright in Revelstoke was getting out of it, and that is not a criticism of them. It is a structural fact about where reconfiguration knowledge lives. It lives on the floor, with the person who watches the machine run every day and knows which of its chances it is wasting.
Now put the vendor pitch next to that. Skild's proposition is that an operator records a demonstration and the system adapts.[1] Vention's is that plain language replaces the specialist across design, programming and troubleshooting.[2] Both are attempts to move reconfiguration authority from the engineering queue to the floor.
That is the right direction. It is also the direction that fails quietly if the floor does not actually have the knowledge, or is not permitted to use it. Moving the capability to the operator only pays if the operator is the one who knows which variable the ceiling belongs to. An operator who has been trained to run the machine as designed will demonstrate the machine as designed, and the system will faithfully learn the wrong ceiling.
Forty sheaves at twenty bucks, and the objection that was never about money
There is a version of this tax that gets refused rather than paid, and it is worth seeing because the refusal is always dressed up as economics.
In the OSB plant, after the flakes are cut and the resin is on them, they go into the orienters, rollers carrying blades set at angles that turn each strand so the fibres lay in a criss-cross pattern and the panel comes out stronger. There are a lot of orienters and they are not on individual drives. They are all tied together off one another through a series of sheaves and idlers with long belts.
One time the plant went down because the orienters simply were not turning. The adjustment slack had been taken right out, nothing left to take up, and the sheaves had worn to the point where they would not drive some of the orienters at all. What the contractors had been doing was spraying the belts with a sticky belt dressing to make them grip. It would hold for a while and then fail again.
There were no sheaves in stock. I put in a work order to change all of them. Roughly twenty dollars apiece, roughly forty of them, a few hundred dollars for the whole set. My line was that the plant had been down for this, they are twenty bucks each, change them all and it is gone for years.
The guys did not want to do it because it was a lot of work, and the argument they kept making was the price.
The price was never the objection. The labour was. And that matters more than it looks, because when the stated objection is price and the price is trivial, price is not the objection, and if you keep arguing the price you never get to the real one.
The diagnostic tell in that story is the adjustment being at its limit. When there is no slack left to take, the wear has already gone past what adjustment can cover, and nothing but replacement works from there. A consumable was being used to mask a worn drive component, and the plant was buying repeated downtime to avoid a few hundred dollars in parts.
Translate that into automation. The belt dressing is the workaround program, the manual step, the operator who knows to nudge the fixture before the cycle starts. The worn sheave is the configuration that no longer fits what the plant now makes. The adjustment at its limit is the point where no further parameter tweak will help. And the reason the real fix does not happen is almost never the cost of the fix. It is that somebody would have to do a lot of work, and nobody wants to own it.
Any honest evaluation of an AI automation product has to ask which of those two costs it removes. If it removes the parts cost but not the labour cost of reconfiguration, it has solved the objection nobody actually had.
The conveyor that came the wrong length
One more, because it shows the tax arriving as pure downtime.
The replacement for the vibrating conveyor arrived the wrong length. I do not know if anyone measured it, or measured it wrong. I had to retrofit both the infeed and the discharge to make it fit, and it took the whole line down for an entire week. A sixteen hour scheduled window turned into an unplanned week because the part was wrong.
That machine would rip itself apart in a day if anything was off, so after this it got examined every day, full bolt-torque checks, the works. And it is what got us talking about the reliability matrix. Until then we had undervalued the machine. Nobody had clocked its importance or its potential for failure, the inspection work orders captured neither, and it was on a once a month frequency for a machine that self-destructs in a day. The conversation widened into redoing inspection frequencies across other machines, and then they brought in a team to build the matrix. I did not build it. I participated in it. It turned out to be a fancy name for what I was already doing in my head, put on paper. Effective tool. We overpaid the outside consultants badly for it.
A week of unplanned downtime, from a dimension. That is the reprogramming tax with the invoice attached, and the invoice went to production, not to engineering, which is why nobody filed it under automation cost.
What this means for a maintenance organization
Four things follow, and none of them require buying anything.
First, find out what you are already paying. The tax is measurable if you decide to measure it. Count the engineering hours spent in a year changing equipment that already worked, as opposed to fixing equipment that broke or installing equipment that is new. Most plants have never separated those three, and the separation is where the number lives.
Second, ask who owns the standard. Twenty gearbox types is a reprogramming tax paid in steel. If nobody owns the standard, variation accumulates for free and is paid for all at once, on a day nobody chose. The same is true of programs, fixtures, recipes and PLC logic. Ownership of a standard is cheap. Its absence is not.
Third, when a vendor offers to absorb variability, ask where the variability goes. It does not disappear. It moves. Consolidating spares moved the risk from availability to shelf life. Consolidating reconfiguration into a learned model moves the risk from the engineer's queue to the quality of what the model was shown. Both are probably good trades. Neither is free, and a vendor who says it is free has not run a plant.
Fourth, and this is the one that decides whether any of it works: check whether the floor is permitted to change anything. Every one of these products is a bet that the person nearest the machine can be trusted to reconfigure it. In a plant where that person cannot currently change a PM frequency without three signatures, the technology will land in the engineering queue anyway, and the tax will be paid exactly as before with a new logo on it.
What to ask a vendor
Five questions, in order, and none of them are about accuracy.
- When the product changes, who performs the change, and what is their job title?
- What does the system need to be shown, and who decides what it gets shown?
- When the demonstration is wrong, how does anybody find out?
- What happens to the configuration when the person who set it up leaves?
- What did the last customer's engineering hours look like before and after?
The fifth one is the whole audit. If the vendor cannot answer it, the product has not yet been deployed long enough for anybody to know whether it works, which is a fine thing to be, as long as everyone says so out loud.
Conclusion
The cost that kills automation is not on the quote. It is the standing obligation to keep paying specialists to make the thing fit reality again, forever, and it is invisible precisely because every individual payment is small and reasonable.
The plants that will get value out of the 2026 wave of industrial AI are not the ones with the best data or the newest robots. They are the ones that already know what their reconfiguration cost is, already own their standards, and already let the person closest to the machine change it. Those plants were paying a lower tax before any of this existed. That is not a coincidence, and it is not an argument against the technology. It is an argument about what has to be true before the technology has anything to bite on.
Twenty gearboxes for the same five-horse conveyor was never a technology problem. Nobody owned the standard. An AI that absorbs variability, dropped into a plant where nobody owns the standard, will absorb it beautifully and hide the fact that it is still there.
References
- Skild AI, "Skild Crosses 100M ARR," 10 September 2026. https://www.skild.ai/blogs/skild-crosses-100m-arr
- Vention, "Vention Facilitates Manufacturing at IMTS 2026 with Physical AI and Agentic AI in One Platform," PR Newswire, 10 September 2026. https://www.prnewswire.com/news-releases/vention-facilitates-manufacturing-at-imts-2026-with-physical-ai-and-agentic-ai-in-one-platform-302874435.html