Predictive maintenance is not a technology

Every few years the industry launches a new maintenance trend. Condition-based, predictive, prescriptive, predictive-with-context, AI-powered. The labels change. The fear behind them does not.

The fear is simple: an unplanned production stop. In a heavy industrial plant a single critical failure during one shift can wipe out seven figures of output. The more surprising the stop, the more it costs — because surprise is what removes the ability to plan around it. That fear has been the engine of every maintenance trend I've seen in this industry, and it is the reason a new trend keeps appearing. None of them have closed the gap. If any of them had, we wouldn't need the next one.

I want to argue something more uncomfortable than "the next trend will finally do it." I want to argue that we've been looking in the wrong place the whole time. Predictive maintenance is not a technology. It's a human workflow. The market keeps selling slices of the technology stack and calling them predictive maintenance. That marketing has increased ambiguity and made it difficult to compare solutions in the market. Everybody is selling AI — a label broad enough to cover machine learning, deep learning, and AI agents under one word.

Predictive maintenance vs. predictive maintenance technology

These are two different things, and conflating them is the source of most of the confusion in this market.

Predictive maintenance is the end-to-end process. Acquire data. Decide which machines need attention. Figure out what is actually wrong. Decide when to fix it. Plan the work. Do the work. Confirm the fix worked.

Predictive maintenance technology is the stack of tools that support that process — sensors, analysis software, machine learning models, planning systems, EAM. The stack is the toolbox. The process is the job.

Most of what is sold under the "predictive maintenance" label is a piece of the toolbox. A vibration sensor system. A machine learning model that flags anomalies. A dashboard. Buyers expect the process and receive a component. That gap is where ROI quietly disappears, and it's also where the disruption-fear sneaks back in.

The workflow, and where the tech actually fails it

Walk the workflow end to end with me. You don't need to be a reliability engineer to notice the pattern — the steps closest to a sensor are well-tooled, and the steps closest to a decision are not.

Filtering. The first useful question on any plant floor is: which machines are healthy and which aren't? Today the answer is built on thresholds and alarms, and engineers live between two bad outcomes. False negatives are the scary ones — a missed developing failure is exactly the surprise stop everyone is afraid of. False positives are the expensive ones — they don't break the plant, they just eat the engineer's day. Every filtering tool on the market is, in the end, a tuning exercise between those two errors.

Troubleshooting. Once a machine is flagged, an engineer has to figure out what is actually wrong with it. This is where the technology gets thinnest. The engineer sifts through signals, spectra, sideband patterns, and trend graphs across two or three disconnected systems. The work is intensely manual and the limiting resource isn't data — it's concentration. The tooling is built as if expert attention were infinite. It isn't.

Assessment. Here is the question the entire workflow exists to answer: when should we fix this machine? Not "is something wrong" — when do we act. I have looked, and there is no commercial software product that genuinely answers this. It runs on experience and tacit knowledge. The engineer remembers a similar gearbox or bearing three years ago and how long it lasted. They benchmark, by hand, against historical cases. The most important decision in the workflow is the one with the least tooling behind it.

Planning and execution. Past assessment, the workflow leaves the condition monitoring software entirely. Information moves by phone call, email, SAP ticket, and sometimes printed paper. Each handoff is a chance for the workflow to break, and every break is a chance for the disruption you were trying to prevent in the first place.

You can see the pattern. We invested heavily in the parts of the workflow that are easy to instrument — measurement and detection — and largely left the rest to humans, meetings, and inboxes. Then we labeled the result "predictive maintenance" and wondered why the fear came back.

The paradigm shift is not better technology

This is the part where I disagree with most of the industry.

The dominant story right now is that better technology will finally crack the problem. Better sensors. Better ML. More data. More AI. I don't think that's the shift that matters. Better sensors don't fix the troubleshooting step. Better ML doesn't fix the assessment step. Putting more data into a workflow whose bottleneck is expert attention will, predictably, make the bottleneck worse.

The shift that matters is software that makes the human better at every step — not software that tries to replace the human at one step. The contest in industry isn't humans versus AI. It's humans with AI versus humans without. The plants that win the next decade will be the ones whose reliability engineers can scale their judgment to every machine in the plant, not just the handful they have time to look at today.

That is the actual paradigm shift. It is unglamorous, and it does not fit on a vendor slide as cleanly as "AI-powered." But it's the one that closes the gap.

What this looks like in practice

Design the stack around the workflow and the day job changes — not because the engineer is replaced, but because the slow, manual parts stop being manual.

A world without alarms. Stop filtering with thresholds. Apply expert-level reasoning to all the data, continuously. Every machine has a current health assessment, kept fresh 24/7. The engineer doesn't get pinged by a buzzer. They open the system and see, at a glance, which machines need their attention today.

A maintenance plan that stays current on its own. Health assessments roll forward into recommended actions automatically. The system writes a draft plan; the expert verifies it. Planning stops being a translation step between two disconnected worlds — condition monitoring and maintenance execution — and starts being one continuous flow.

Verification that closes the loop. After the fix, the same expert-level baseline checks whether the machine is back to nominal, automatically. Assembly errors get caught before they become the next surprise. The workflow finally has an end, not just a handoff.

None of this is anomaly detection. None of it is a smarter alarm. All of it is the workflow finally being done as a workflow.

Back to the fear

Unplanned disruptions cost the most when they are the most surprising. Surprise is what removes your ability to plan, to schedule, to ensure spare parts exist, to keep the production targets you sold to your customers. Every maintenance trend of the last thirty years has been an attempt to take surprise out of the system. None has done it, because each tried to bolt technology onto a workflow nobody was running end to end.

There is no next maintenance trend coming. Or rather, there shouldn't be. What's coming is predictive maintenance done properly — the workflow factories have always been trying to run, finally supported end to end, with software that lets a small team of experts deliver their judgment across an entire plant.

That is what reduces surprise. Not better sensors. Not bigger models. The workflow.