Everyone in industrial machinery is "doing AI." Not one can prove it's worth a cent.

The big consulting firms have been hinting at this for a year. I went and checked it on eight real companies. Guess how many could show a number. Zero.

Lucas Barrios · Applied AI & Operational Transformation · Berlin

Let me save you the suspense.

Everyone is adopting AI. Almost nobody can tell you what they got for it.

That's not me being cynical. It's what the biggest research shops have been quietly reporting for a year. MIT's Project NANDA looked at 300 enterprise AI deployments and found that 95% of them delivered no measurable impact on the P&L. McKinsey surveyed nearly 2,000 companies and found only about 39% could tie AI to any profit impact at all, and for most of them it was under 5%.

Huge adoption. Tiny proof. That's the whole story.

But those are surveys. Big, anonymous, averaged. I wanted to know if it held up when you put real names on the table. So I did.

I benchmarked eight machinery makers. The measurement column was empty for all of them.

GEA. Krones. TRUMPF. Siemens. KION. Sandvik. Heidelberger. Dürr. Eight of the biggest industrial machinery companies in Europe.

I scored each one across six dimensions, using a rubric I locked before I looked at a single company, so I couldn't move the goalposts later. Everything is public. Every score has a quote, a link, and a date behind it.

One finding survived every test I threw at it:

Not one of them publishes a single measured result from AI in their own operations. Not one real number with a starting point and a scope attached. On the "measured outcomes" score, the average was 0.75 out of 5. Nobody cracked a 1.

And here's the kicker. It's not that they can't count.

They count obsessively. Just never for themselves. Siemens brags about a 25% maintenance improvement, for product pilots. GEA publishes an 8% energy gain, for customers running GEA machines. I threw out roughly forty pieces of evidence full of precise numbers, because every one of them measured somebody else's result, not their own AI adoption.

The whole sector is a measurement machine pointed in every direction except the mirror.

The companies actually winning at AI aren't the ones with the best tech. They're the ones who decided to measure.

This is the part nobody wants to hear.

The winners in McKinsey's data and Accenture's research have one boring thing in common: they decide who owns value measurement before they deploy anything. Not a dashboard they build afterward to justify the spend. A discipline they install up front.

So flip the empty column around. For a machinery company, that gap isn't embarrassing. It's the easiest lead in the market. Every single competitor left the box blank. One honest internal number, with a baseline and a scope, and you're ahead of the entire field overnight. Not because the others have no results. Because none of them bothered to measure and say so.

That's not a technology advantage. That's a courage advantage.

"Are we mature at AI?" is a garbage question.

Everybody asks it. It leads nowhere.

Here's why. The two things people lump into "AI maturity" don't even move together. Look at what the benchmark turned up:

KION runs a generative AI tool in production and discloses zero governance. Its main sustainability report doesn't mention a single AI-governance term. Krones is the exact opposite: best governance in the group, an AI advisory council, EU AI Act prep, and their flagship internal use case was still a pilot. Sandvik has a Chief AI Officer and a whole AI center, and not one named internal deployment to point to.

Nobody scored well on both.

So stop asking "are we mature." Ask "which half are we missing, the doing or the governing." That question actually goes somewhere, and the answer is different for every company. Accenture found the same split at scale: 36% have scaled generative AI, but only 13% are getting real enterprise-level value out of it. The gap between those two numbers is the whole game.

Everyone's obsessed with the model. The model is 10% of the job.

Your data is a mess. That's the real bottleneck, and almost nobody wants to admit it because fixing data isn't sexy and doesn't demo well.

In the benchmark, data foundation was the second-weakest thing across all eight companies. Most of them say data is a priority and show nothing built behind it. One announced a big data platform back in early 2024. Two and a half years later, no source confirms it actually exists.

BCG put a number on this that every founder should tattoo somewhere. Their 10-20-70 rule: the algorithm is 10% of an AI transformation. The tech and data plumbing is 20%. People and process is the other 70%.

Read that again. The thing everyone argues about on LinkedIn, which model, which vendor, is 10% of the work. The unglamorous stuff is where the value actually lives. Fragmented data quietly kills agentic and generative systems before they ever pay off. Data first. Shiny layer second. In that order, or don't bother.

The field isn't a spectrum. It's two tribes.

Six companies landed between 23% and 45%. Two crashed at 11%, less than half the next lowest.

And the bottom two got there in completely opposite ways. One is all governance talk with no real deployment. The other has a real deployment with no governance at all. Same low score, opposite diseases.

That's the actual takeaway, more than the ranking. There's a group that discloses something about their own AI, and a group that discloses basically nothing. Two tribes.

Rank	Company	Score (% of max)
1	GEA Group	45%
2	Krones	41%
3	TRUMPF	40%
4	Siemens	39%
5	KION	30%
6	Sandvik	23%
7	Heidelberger	11%
8	Dürr	11%

That percentage is out of a theoretical max nobody could realistically hit. It's a reading aid, not a target. The six dimensions: data foundation, process automation, agentic and generative deployment, governance, workforce enablement, and measured outcomes.

I'm showing my work, including the parts I got wrong.

Most benchmarks hand you a ranking and hide the machinery. I did the opposite.

The rubric and weights were locked before company one. The git history proves the order. Every score has a quote, a URL, and a date. And there's a live explorer where you can throw out my weights, plug in your own, and watch the ranking rearrange itself in real time. If you think I weighted governance too high, prove me wrong in thirty seconds.

I changed the framework five times while working, out in the open. The process caught two of my own mistakes on re-run. Siemens jumped from 31% to 39% once I read a deployment in full instead of skimming a summary. GEA went from 39% to 45% when I found a governance policy I'd missed the first time. Both fixes are on the record, with the wrong versions kept next to the right ones.

A benchmark isn't credible because it nailed everything on attempt one. It's credible because you can see it correcting itself.

One thing worth stealing for how you read any AI coverage: loud isn't the same as mature. On my first pass, a private company that publishes way less than its listed rivals beat a much bigger one, because a single honest interview about its own operations was worth more than a giant pile of reports aimed at selling to customers.

What this does not prove

Let me kill the obvious objection before you type it.

This is about what companies disclose, not what they actually run inside. A company scoring low here might be quietly doing serious AI it just hasn't published. Honestly, given the gap between what these firms sell and what they admit to running, that's probably true. Real internal adoption is almost certainly further along than any outsider can see.

This is one person's desk study, in English and German sources, at one moment in a field that moves weekly. Treat the scores as a starting point for your own thinking. Don't quote them as gospel.

The one line to walk away with

The consultants and my benchmark meet in the middle from opposite ends. MIT names the 95% who can't show a return. McKinsey names the 39% who can show anything. Accenture names the 13% getting real value. BCG says 70% of the work is people and process, not tech. And eight machinery giants, checked one by one, show what all of that looks like up close: everyone's adopting, nobody's measuring.

This sector's AI problem was never capability. It's measurement, sequencing, and actually connecting governance to deployment instead of building them in separate rooms at different speeds.

That's a business problem before it's a tech problem. Which, funny enough, is exactly where the real money in AI has been hiding the whole time.

Want the full thing? The complete analysis, the evidence behind every score, and the explorer where you can reweight everything yourself are all public:

Lucas Barrios helps DACH and EU companies turn AI from a slide deck into governed workflows with numbers attached. Let's talk.