Has AI actually changed manufacturing? Follow the money.

Three or four years ago, when ChatGPT landed and the whole world briefly lost its mind, I took a position that did not win me many friends: there was not one good, monetisable use case for AI in manufacturing. Not one. It is 2026 now, and the honest truth is that I have changed my mind a lot less than I expected to.

So let me push back on myself for a second, because I want to be fair to the technology. Has AI moved the dial in manufacturing, digital, distributed, decentralised, whatever label you prefer? A little. There is a handful of genuine use cases. But are they clearly monetised? Not really. The vendor selling you real-time vision inspection is making money. The manufacturer? AI is just sitting inside the automation bill. It is not its own line item. It is not its own return.

The dark factory is not an AI story

Everyone points at the same thing: the lights-out factory, nobody on the floor, the future made real. So look at what is actually holding it up. Robotics, sensors, PLCs, orchestration, years of patient automation engineering. AI sits on top as one layer, and a thin one at that. I am not being cynical. I think that is simply what it is.

The real 2026 story is not AI being the smartest thing in the building. It is a whole wave of technologies that are enabled by AI. Perception. Data. Sorting signal from noise. That is the part that is genuinely working, and it is worth being precise about, because the difference matters.

What AI actually does well

Here is a good example, and almost everyone reads it backwards. An AI early-warning system at a Toronto hospital cut unexpected deaths by 26%. A real number, properly tested, peer-reviewed and published. But look closely at what the AI did, and did not, do. The researchers were explicit: it does not prescribe, it does not treat, it does not decide. What it does is watch a hundred data points across thirty patients that no human can hold in their head at once, and it prompts a clinician to go and look.

That is the whole thing. The win is not the decision. The win is surfacing the right signal and putting it in front of a human in time to act on it. And that is exactly what AI should be doing on a factory floor.

Follow the venture money

If you want to know what an industry truly believes about itself, follow the money. In the first quarter of 2026, four companies, OpenAI, Anthropic, xAI and Waymo, raised $188 billion between them. That is 65% of all global venture funding for the quarter. AI as a whole took 80%. Eighty.

Now, the manufacturing and robotics rounds are real. Mind Robotics raised $615 million. Hadrian raised $260 million to build AI-powered factories. Bezos put $10 billion into an AI manufacturing fund. But set against the model labs, it is a rounding error. That is the honest word for it. A rounding error.

Why manufacturing keeps getting skipped

So the real question is why. Are we actively avoiding manufacturing because it is genuinely hard to apply? Legacy systems, cultural inertia, plants that have run the same way for thirty years and do not want a model anywhere near them? Or is it simpler than that? Money flows where the hype is, and the hype is foundational models. And honestly, that is understandable. Software scales. A factory does not.

Is there a shift toward physical AI? Yes. Is it happening fast enough? That is the real question. The answer people give you is that the foundation models are the critical layer, and the physical stuff comes later. Maybe. But here is my point. Demand does not wait for later.

What does the future actually hold?

So let me bring it home. The UK. Europe. What does the future actually hold? Dark factories? Highly automated ones? The one-person factory, one person running the whole floor? Or do we keep importing the story while the money quietly goes into building models, not machines?

This was never really about AI

Here is the thing I keep coming back to. This is not really a piece about AI. It is a piece about engineering value. Software scales, models improve, but somewhere, someone still has to make something. I went deeper on all of it on the Automate UK podcast, if you want the long version.

So tell me where I am wrong.

Engineering has always been judged by measurable value. AI should be judged the same way.