The retreat has started
The retreat has started. You can feel it in the room before it shows up in a survey.
PwC's CEO survey: 56% of chief executives saw no significant financial return from AI. One in eight got both cost and revenue improvement. KPMG: nearly half of executives have pulled back on AI agents, mostly over cost. Boards spent two years asking "what's our AI strategy." Now they're asking what we got for it.
I'm not going to defend the people who can't answer. Most of what got funded deserved to be cut. A chat window on top of a document repository was never going to move a production number. Everyone in the room knew it. Some of us said so.
But there are two ways to read a retreat. The technology was oversold. Or the money went to the wrong half of the company.
Watch where the smart ones go
The executives worth watching aren't exiting. They're relocating.
Budget is moving out of the office and into the plant. Off the desktop, onto the asset. Away from summarizing what's already written, toward seeing what's happening right now in a physical place where the consequences are physical too.
The industry calls this physical AI. Perception and action in the real world, instead of text about the real world.
That's the right read of the same data pushing everyone else out. Here's why.
Physical AI has a denominator
When AI drafts memos, nobody can tell you what a good memo is worth. So the return conversation turns into a debate about time saved. Time saved never lands on a budget line. Eventually a CFO asks where the money went and there's no honest answer.
When AI counts things, inspects things, or watches for a condition that used to need a person standing there, the value is arithmetic. The count exists or it doesn't. Somebody drove out there or they didn't. The leak got caught Tuesday or Friday.
That doesn't make the work easy. It makes it measurable. Measurable survives a board meeting. Vibes don't.
Second reason, and the one I care about more: in the field, the hard part is finite.
What kills industrial AI projects isn't the model. The model has been the easy part for a while. What kills them is integration labour. The undefined weeks between "we have a model" and "it runs unattended somewhere with no network and nobody to babysit it." Getting a device on a network. Getting it to survive a reboot. Getting data off it when the connection drops. Knowing which of your forty devices went quiet.
That labour is real. It's also not novel. It's the same work every time. Which means you can solve it once and stop paying for it. A model deployment needs constant minding to keep producing something nobody can price.
You find the end state by chipping at it
Nobody knows what the end state of industrial AI looks like. I don't. Be suspicious of anyone selling you a picture of it.
That's exactly why the retreat is happening. The last two years asked executives to fund a destination nobody could describe, on a timeline nobody could defend, through a transformation program that had to be right the first time. When the invoice came due, the honest ones admitted they couldn't say what they'd bought.
The alternative isn't a better roadmap. It's a smaller unit of work. Put something in the field. Learn what it tells you. Let the next thing be shaped by what the last thing taught you.
If you just cut an AI budget and are wondering what to fund instead, test the next proposal against this:
Can you name the person on site who wants it? By name. If not, it's an executive's idea, and it'll be deployed to people who were never asked.
Does it produce a number that didn't exist before? Not a nicer summary of numbers you already had. A new number.
Would a human have had to drive out to get it? That trip is your baseline. It has a cost, and everyone already agrees on what it is.
Can your own people deploy the second one without the vendor? If not, you didn't buy a capability. You bought a pilot. Pilots are where this money went to die the first time.
Every site has a list of things everybody watches and nobody records. The gate that gets left open. The line that runs hot on nights and nobody can prove it. The tank someone eyeballs out of habit. None of it is a safety incident yet. None of it is a production number. It's just what the people there know, in their heads, unwritten because nobody was going to fund a six-week integration project to write it down.
Your site is full of those. The people who work there can name a dozen before lunch, and most have never been asked. The companies that come out of this holding something real won't be the ones with the best AI strategy. They'll be the ones that put the most small things in the field and paid attention to what came back.
