Somebody at the trade show told you robots are coming
You've heard it twice this year. A vendor booth, a podcast, a nephew. The AI can see, it can reason, it can plan the job — surely the part where it does the job is nearly here.
So here's the honest version, from the wiring outward.
A model can already write you a genuinely good diagnostic sequence for a rooftop unit. Check the filter, read the superheat, look at the contactor. It's correct, it's ordered, and a competent tech could follow it.
It cannot turn the valve. Not "isn't allowed to" — cannot. And the reason is more interesting than "the technology isn't ready," because it tells you exactly which parts are close and which aren't.
Three clocks that don't match
The single most useful thing to understand about AI running hardware is that the stack has at least three layers, and each one runs at a wildly different speed.
- The language model — the part that reasons about the job — takes something like a second or several per decision. It is thinking in sentences.
- The policy — the part that turns "open the damper" into joint movements — runs somewhere between a handful and a few dozen times a second.
- The controller — the part that actually holds a motor at a position against a load — runs hundreds to thousands of times a second, and never stops.
Those aren't three settings on one dial. They're three different kinds of software, and the gap between the top and the bottom is roughly a thousand to one.
Ask why you can't just skip the middle and let the model command the motor directly, and the clocks answer it. In the second the model spends deciding, a motor under load has already moved — possibly into something. A control loop can't wait for a sentence.
The layer that closes the gap has a name
The middle layer — the part that takes a camera image and an instruction like "pick up the wrench" and emits actual movement — is a policy. The current generation of these are called vision-language-action models, VLAs, and they're the genuinely new thing of the last few years.
They're worth knowing about because they're the piece that was missing. Google DeepMind's RT-2 was the first widely noticed one; open models like OpenVLA, and work from labs like Physical Intelligence, have followed. The idea in one line: instead of programming a robot's motion, you train it on demonstrations the way you'd train a language model on text, and it outputs movements instead of words.
They work. They also work far less reliably than a language model does at writing, and on a much narrower range of tasks, which is exactly what you'd expect from a field that's several years behind.
The physical world doesn't do retries
Here's the asymmetry that makes this hard in a way text never was.
When a model writes a bad paragraph, you delete it. The cost of a wrong answer is that you read it. That's why chat tools got useful so fast — the failure mode was free.
When a model sends a bad torque command, something bends. Or a panel comes off, or a line gets over-pressurized, or a person who was standing in the wrong place has a bad day. The failure mode has a repair bill and sometimes a lawyer.
Everything strange-looking about robotics — the caution, the slowness, the fact that impressive demos take years to become products — comes out of that one difference. It isn't timidity. It's that you can't iterate your way through failures that cost money to undo.
Every real machine is a little bit wrong
The other thing that separates a demo from a deployment: on paper, a machine has exact dimensions and its sensors read true. In your shop, none of that holds.
The temperature probe reads two degrees high. The pressure transducer has wandered since it was installed. The actuator overshoots slightly when it's cold and less when it's been running an hour. The nameplate says one thing and the machine does another, and your senior tech knows all of this and compensates without thinking about it.
A model does not know any of it. The number it receives is the number it believes.
This is why "it worked in the lab" and "it works in your building" are two different projects, and why the boring, unglamorous work of measuring what your equipment actually does is worth more right now than any model is. More on that in what to do now if your business runs on equipment.
What this changes about the question
"Are robots coming for the trades" is the wrong question, because it treats one stack as one thing.
The reasoning layer is here now and it's good. The policy layer is real, improving quickly, and nowhere near reliable enough to leave alone. The control and safety layers underneath have existed for decades, are well understood, and are not going to be replaced by a model — nor should they be.
So the honest answer is: the top of the stack is already in your business if you want it, the bottom of the stack was always there, and the middle is the part being fought over. Nobody's shipping a machine that watches a job and figures out the rest.
And there's a second problem, which is that even where the pieces work, none of them agree on how to connect to each other. That's the missing standard, and it's the reason this stays a research story longer than the demos suggest.
This week
Take the one physical task you'd most want handled — the after-hours filter change, the tank top-off, whatever it is — and write its three clocks:
[DECIDE] What has to be judged? How often?
[MOVE] What physically moves? How fast does it need correcting?
[HOLD] What keeps it safe if everything above fails?
If [HOLD] is empty, that task isn't a candidate for anything automated yet, and that's a useful thing to know in ten minutes rather than after a purchase order.
If the task turns out to have no physical movement in it at all — it's judgment, paperwork and a phone call — congratulations, you found the one that's actually ready. That one is a harness, and you can build it this quarter.
Most of what businesses want from "AI robots" turns out to be a scheduling, dispatch or paperwork problem wearing a hard hat. If you want a straight answer on which half of your problem is which, book a 30-minute call and bring the task.