Most companies putting AI to work in their operations spend all their energy on the model. They compare vendors, tune prompts, and argue about which platform to standardize on. Far fewer stop to ask a simpler question: where are the data actually coming from?
For a lot of AI-driven monitoring and automation, the answer is a physical sensor sitting on a wall, a pipe, or a factory floor. If that sensor's wrong, the AI is wrong right along with it—confidently and at scale. As these systems shift from drawing charts to making the call themselves, that quiet dependency turns into a real exposure.
The ceiling under every model
An AI system can only reason about the data it's given. A model watching a building's air quality, a plant's emissions, or a cold chain's temperature has no independent way of knowing whether a reading is real. It takes the number and acts on it, throttling ventilation or signing off on a shipment. That makes the humble sensor the real foundation of the whole thing.
Security teams are waking up to this, and SecureWorld has looked at how AI can make or break IoT security, with the outcome coming down almost entirely to the quality of what you feed it. There are also more of these foundations every year. IoT Analytics expects the number of connected IoT devices to reach 21.1 billion by the end of 2025, with a growing share of them piping data straight into analytics and automated systems. The idea here is old and unglamorous: garbage in, garbage out. What's changed is the blast radius. A bad reading used to mean a misleading chart, but now, with a model driving the response, it means a wrong action taken instantly, with no human in the loop.
When the data are wrong on purpose
Some bad input is deliberate. A sensor is a way in, and an attacker who can shape what it reports can shape everything downstream. Feed a monitoring model a slow drip of doctored readings and you can train it to see an abnormal state as normal, so the alert never fires when it really matters. The U.S. National Institute of Standards and Technology (NIST) maps out these techniques in its work on adversarial machine learning, including the data poisoning attacks that corrupt a model through the very information it learns from and leans on.
The tricky part is that an attack like this hardly looks like an attack. There's no breach notification, no ransom note, just a system confidently making the wrong call because someone quietly rearranged its picture of the world through the devices it trusts. These make appealing targets precisely because nobody's really watching them, and because the payoff is quiet influence over a decision instead of a noisy smash-and-grab. If you're leaning on AI to run operations, that's the kind of failure mode worth losing sleep over.
When the data is wrong by accident
Most bad sensor data isn't an attack at all, just cheap or aging hardware doing what cheap or aging hardware does. Take gas detection, one of the most common jobs that sensors get handed in commercial and industrial buildings. The devices that track oxygen, carbon monoxide, or air quality lose sensitivity as they age. They get slower to respond, and they start reporting numbers that look plausible but aren't right. This is the most dangerous kind of wrong, because nothing obvious flags it. A model can't tell a confident accurate reading from a confident inaccurate one, and it's got no gut instinct telling it a number feels off.
Here comes the point where the choice of hardware stops being a line on a purchase order and turns into a data-integrity decision. The better makers of electrochemical gas sensors put every unit through defined accuracy and response testing before it ships, because their customers are building life-safety and decision-critical instruments on top of them. A sensor that holds its calibration for years is one a model can keep trusting. One that quietly drifts out of spec becomes a slow leak of bad data into every decision it touches, and the AI will defend that decision as confidently as it defends a good one.
The sensors nobody owns
There's one more problem that has nothing to do with the data and everything to do with who's watching it. Sensors may be installed by facilities teams, safety managers, and outside contractors, but regularly go unlogged in the IT or security asset inventory. They then sit on the network, feed data into systems that make real decisions, and answer to nobody in particular.
You can't secure, patch, or sanity-check a device you don't know exists, and you definitely can't judge whether its readings have earned the trust your AI is putting in them. Every unmanaged sensor is a small act of blind faith wired into the foundation, and most operations are running more of them than anyone's ever counted.
Closing the gap
The fix isn't complicated, though it does mean treating sensor data as something you earn rather than assume. Start with an inventory, so you know what's deployed, where it sits, and what shape it's in. Fold the physical layer into your risk thinking like any other part of the stack, with owners, review dates, and a replacement plan. Lean toward devices built to recognized security and performance standards, something regulators are pushing through programs like the U.S. Cyber Trust Mark.
Choose your sensing hardware with the same care you'd give the model itself, because the smartest AI in your operation is still taking a physical device at its word. Choose the right sensor and everything above it has a shot at being right too. The alternative is likely to mean all you've really done is automate a mistake.
