Out of the catalogue
Ask a vendor what your problem is and the answer is whatever the vendor sells. Ask an AI vendor and the answer is AI. I am one.
Gartner has a sentence for this. It is in the June 2025 release that predicted over 40% of agentic AI projects will be cancelled by the end of 2027, and I would have put it first: "Many use cases positioned as agentic today don't require agentic implementations." The same release estimates that "only about 130 of the thousands of agentic AI vendors are real".
Gartner gives no count for that sentence. I have one, for one list. The list is older than agents. The choice on it was between a model and arithmetic, and I think it is the same step being skipped.
This year we sorted a client's list. It was not a list of AI projects. It was their assurance programme: about two thousand periodic checks, each written as something a technician confirms by hand. Verify that the generators share load correctly. Pressure-test the pumps to their rated pressure. Every line was a candidate for automation, and we are the ones who sell the automation.
We sorted it into three piles. A calculation is a number derived from data already being recorded, with no alert attached. A report is evidence that a test took place: when, for how long, how close to the limit. A model is something that has to watch continuously.
The sorting was not AI either. A script read the words.
"Function test" and "pressure test" went to report. "Verify setpoint" and "calculate" went to calculation. "Monitor", "deviation" and "imbalance" went to model.
A verb describes how the check is done by hand today, not what the machine needs. "Verify that the generators share load" starts with verify, and load sharing is something you watch all the time. So the rules read more than the first verb, an engineer's comment on a line overrode the rule, and every line kept the word that had decided it, so an engineer could correct the rule and not the row.
About half of the list was a report. Nobody needs a model to know whether a pressure test took place. The rest split between calculations and lines that do need a model. So at least half of that list needed no model of any kind.
For the lines that did need one, we considered deep learning and rejected it. It needs labelled failures that nobody has, and our rule was that a maintenance superintendent must be able to repeat the explanation back after hearing it once. What shipped compares each machine with its siblings.
I still arrived with a catalogue. Twice the client's specialists corrected us. One model began as anomaly detection, and they told us a physics-based approach was the right one. We abandoned ours and rebuilt it. On another they told us that one of our data sources would add noise and not signal, and we dropped it.
They knew the equipment. We had brought the method.
What went back to the client was the list: each line in its pile, with the word that put it there, and the decision column empty. About half of it was a report, and we built those too. Nobody means a report when they say AI.
What I'd bet on
A lot of industrial AI now starts by putting a language model on top of the problem. A good part of what industry needs solved does not need one. I'd bet most lists like this one sort the same way, with at least half the lines needing no model of any kind, let alone a language model. Anyone who has such a list can check. Our own platform has a language model at the front of it, so the rule applies to us first: the model reads the question and writes the query, and the number comes from the data.
Diego Mercadal started as a commissioning engineer on high-voltage motors and generators, joined an offshore drilling contractor as a rig hand, worked several positions in the drilling crew, and ended up running its AI/ML function. He is now co-founder and CEO of Wonder DataLabs.