This is the eighth part in a series on how AI is crossing out of software and into the physical world. This time, Jan Bosch looks at the layer that does the verifying and what it actually costs to install.
In 2016, a mechanical lidar cost something like 80,000-100,000 dollars; a 64-beam Velodyne was around 80,000 dollars. Today, solid-state units are being targeted below 200 dollars, with a stated long-term goal of 100 dollars and several competitors aiming under 500 dollars. That’s a reduction of two to three orders of magnitude in a decade. The same curve, less dramatically, has run through cameras, MEMS accelerometers, gas sensors and thermal imagers. If your plan requires putting an eye or an ear somewhere, the sensor is no longer the reason you can’t.
So, the comfortable story is that the world is becoming machine-readable by default. Instruments proliferate, data exhaust accumulates and AI arrives later to mine it. I want to argue that this story has the right economics and the wrong object. The sensors are, in most industrial settings, already there. But they’re simply the wrong sensors – and not because anyone made a mistake.
Here’s the thing I think is genuinely underappreciated. Almost all installed industrial sensing exists to close a control loop. A thermocouple holding a setpoint, a pressure transducer tripping an interlock, an encoder feeding a servo. For that job, what you need is the present value, repeatability and low latency.
Notice what you don’t need. You don’t need absolute accuracy, because the loop only cares that the reading is consistent, not that it’s true. You don’t need retained history, because the loop only cares about now, so the old values get overwritten or thrown away. You don’t need timestamps that align with other subsystems, because each loop runs independently. And you emphatically don’t need context – what product was running, which operator was on shift, which batch of feedstock went in – because the controller doesn’t care and nobody was going to read it.
Every one of those four is mandatory for learning, and none of them is mandatory for control. Which means decades of capital spent on instrumentation produced plants that are thoroughly instrumented and barely observable. Those are different properties, and the industry has been conflating them for about ten years.
The downstream numbers are what you’d predict. Industry analyses put the share of industrial asset data that goes unused at something like 95 percent, and the reasons given aren’t about volume or compute; they’re missing metadata, inconsistent labelling, no standardization, time-series locked in legacy systems and engineers’ spreadsheets – data described, memorably, as “some of the most ungoverned and ungovernable.” Making it usable is reported to take months or years of human effort per site.
This also deepens something from the fifth part in this series. I argued there that a model trained in one plant doesn’t transfer to the next because the cameras hang differently and the tags are named differently. That’s true, and it’s the shallow version. The deeper version is that the measurements were never constructed to be comparable in the first place, because comparability wasn’t a requirement of the job they were installed to do.
Which brings me to the cost structure, and to the sentence I’d put on the wall of every industrial AI program: The capex collapsed and the opex didn’t.
A sensor isn’t an asset; it’s a maintenance obligation with a price tag attached. It drifts. It fouls. It needs recalibration on a schedule somebody has to own. Nodes die quietly and are discovered months later. Install ten thousand cheap sensors and you haven’t acquired ten thousand data sources; you’ve acquired a permanent staffing commitment, and that commitment is the line item that kills more business cases than any model ever has.
Drift deserves particular attention, because it fails in the worst possible way. An uncalibrated sensor doesn’t stop producing readings; it produces confident, plausible, slightly wrong ones, and a model trained on them learns the drift as if it were physics. You get a system that’s precisely tuned to an instrument’s decay. There’s no error message for this.
Underneath all of it sits a point I keep coming back to. Sensing doesn’t produce data; it produces a claim – and a claim needs a context to mean anything. “Vibration amplitude 4.2” isn’t information. “Vibration amplitude 4.2, on bearing three, on the line running product B, six hours after a bearing change, with feedstock from the new supplier” is information. The signal is nearly free now; the context is expensive, largely human and doesn’t get cheaper when the sensor does.
For a startup, that’s the whole opportunity, and it isn’t the one most people pitch. Don’t sell the sensor; sell the claim. The defensible products in this layer are the ones that attach meaning to signal: context capture that records what was happening and not merely what was measured, semantic and metadata layers that make one site’s readings commensurable with another’s, calibration and drift detection as an ongoing service rather than a scheduled visit and tooling that establishes ground truth cheaply in domains where labelling currently requires an expert standing next to the machine.
Here’s a test I’d apply to any company in this space. Can you tell your customer what a given reading means without asking them? If the answer is no, you’re selling instrumentation and the customer is still doing the expensive part.
For established manufacturers and asset owners, the decision has been misframed, in my experience, almost universally. The question is presented as whether to instrument; it’s actually whether to re-instrument – to go back to equipment that already has sensors and add the observability those sensors were never specified to provide. That’s a capital allocation question competing against replacement, not a technology question, and it should be argued in those terms.
There’s also a quieter governance issue. Whoever owns the semantic layer – the translation from raw signal to meaning – owns the leverage over everything built on top. At the moment, that layer is usually supplied by a vendor, which means the asset owner has outsourced the definition of what its own plant is doing. I’d think hard about that before signing.
For society, this is the part of the series where the technology is least visible and the stakes are most personal, and the honest framing is about who gets asked. Inside a fence, instrumenting is a procurement decision. Point a camera at a machine, buy the license, done. Outside the fence – or inside it but pointed at people rather than equipment – the same stack becomes a political decision, and the industry has a long habit of shipping the first and discovering the second.
Europe has already drawn part of that line, more sharply than many engineers realize. Since 2 February 2025, the AI Act’s prohibitions have been in force, and two of them land directly on the sensing layer: Systems that infer emotions of people in the workplace or in educational settings are banned outright, and so is biometric categorization that infers race, political opinions, trade union membership, religious belief, sexual orientation or health status. These are prohibitions, not high-risk obligations – there’s no compliance path, and the penalty ceiling is 35 million euros or 7 percent of worldwide turnover. A factory camera program that includes “operator attentiveness” or “sentiment” in its feature list is no longer an engineering decision with an HR footnote.
The deeper difficulty is structural rather than legal. Ambient sensing works precisely by not asking. Its value proposition is that it observes without requiring anyone to do anything, which is the same sentence as “without requiring anyone to agree.” Consent is the one input this layer can’t measure, and the cheaper sensing becomes, the more the default drifts toward installing first.
One last connection, because it ties the series together more tightly than I expected when I started. I’ve argued throughout that the durable artifact in AI systems is becoming the contract a system must honor plus the continuous evidence that it does. Every evidence chain in this series – the factory’s yield statistics, the vehicle’s safety case, the laboratory’s measurement – terminates in a sensor. If that bottom link is drifting, uncontextualized and unowned, then every contract stacked above it is unverifiable, however rigorous the layers in between. The evaluation problem doesn’t begin with the model; it begins with the instrument.
Which is why I want to close with a line that’s almost always attributed to Einstein but was in fact written by a sociologist. William Bruce Cameron, in Informal Sociology in 1963: “Not everything that can be counted counts and not everything that counts can be counted.” The Ultimate Quotable Einstein files the Einstein version under “probably not by Einstein,” which is itself a small lesson about unverified claims propagating because they sound right.
Cameron’s two clauses are the two failure modes of this entire layer. The first is optimizing what the sensor happens to measure because it’s measurable. The second is losing what matters because no instrument was pointed at it. A sensing program is, whether anyone frames it this way or not, a decision about which of those two categories your organization will be able to see for the next decade. Most organizations are making that decision by accident – in a procurement process, on price, by buying whichever sensor had a datasheet. It deserves to be made on purpose, by someone who can be named.
