The Floor Report #8: Your AI Is Just Expensive Autocomplete
Why the algorithm is never the bottleneck, and what actually is. A practical take on AI in maintenance. Not anti-technology. Anti-theater.
I watched a plant spend $340,000 on a predictive maintenance pilot last year. One vibration sensor on one compressor. It caught a bearing degradation pattern two weeks before failure. Everyone celebrated. The VP got a slide. The vendor got a case study. The reliability engineer got a pat on the back.
Then somebody asked the question that nobody wanted to hear: what about the other 400 assets?
Silence.
Not because the AI was bad. The sensor worked exactly as advertised. The algorithm did its job. The problem was everything underneath it. The CMMS had six years of failure history coded as "other." The timestamps on half the work orders were wrong because the crew was closing them in batches at the end of the shift. The PM records for the assets adjacent to that compressor were so thin that the data science team couldn't even build a training set.
The plant had spent $340,000 proving that AI works when you hand it clean data and one asset to watch. It had also proven, without meaning to, that it couldn't scale the thing to asset number two.
This is Pilot Purgatory. And most plants I talk to are stuck in it.
The problem is not the algorithm
Google Trends tells a clear story right now. "Industrial AI" has a search interest score of 49 over the past twelve months. "Predictive maintenance" is at 15. "CMMS" is at 12. The conversation has moved. Everyone is talking about AI. Vendors are selling AI. Conference panels are about AI. Plant managers are being asked by their VPs when the AI strategy is coming.
But here is what the data actually says about the plants doing the asking.
Limble's 2026 benchmark report found that only 20% of maintenance organizations genuinely trust their asset data across teams. Not "have data." Trust it. Believe the numbers enough to make decisions from them. Among organizations with strong data discipline practices, that number rises to 51%. Among the rest, it falls to single digits.
MaintainX's 2026 report found that many maintenance teams still spend less than 40% of their time on planned work. The rest is reactive. Firefighting. Chasing the breakdown that nobody predicted because nobody was doing the inspections that would have caught it.
So the picture is this: an industry racing to adopt AI while most of its members can't trust the data the AI would need to learn from, and can't execute the planned work that would generate that data in the first place.
If your CMMS history is noise, your AI layer is just expensive autocomplete. It will pattern-match on garbage and give you confident-sounding predictions built on a foundation of "other," "miscellaneous," and "repaired. RTO."
The skills gap is the AI gap
The second problem nobody in the vendor demos wants to talk about is who runs this thing after the pilot team leaves.
AI copilots for maintenance are being positioned as the answer to the skilled-trades shortage. And the pitch sounds reasonable. Your senior tech is retiring. He's the only one who can diagnose the Paper Machine drive system by sound. What if we captured that knowledge in an AI that could guide the junior techs?
Good pitch. One problem.
Who trains the AI? Who validates its recommendations? Who catches it when it confidently tells a second-year mechanic to check the wrong bearing on the wrong end of the drive? The same senior tech who is already burned out from being the only person who can troubleshoot three systems after hours.
Plant Services' April 2026 panel coverage keeps surfacing the same theme: manpower and skill constraints are the everyday bottleneck. Not technology constraints. Not budget constraints. The constraint is that you have two people who actually understand the equipment, and you're asking them to also become AI trainers, data validators, and integration consultants on top of their existing jobs.
AI copilots don't replace the skills gap. They require the same people the gap is already crushing. The bottleneck hasn't changed. You've just added a more expensive layer on top of it.
What actually has to happen first
I am not anti-AI. I've seen what good predictive analytics can do when the foundation is there. But the foundation is not optional, and it is not exciting, and nobody is going to put it on a conference slide.
Here is what "AI-ready" actually looks like at the plant level:
Your failure codes mean something. If your top three failure codes are "other," "general repair," and "see notes," you don't have failure data. You have a diary nobody reads. An algorithm trained on that will learn one thing: that your plant breaks in ways nobody bothered to describe.
Your work orders close with useful information. Not "repaired, RTO." What failed. Why it failed. What was done. How long it took. What parts were used. If the close-out field is a formality that gets filled in during the last ten minutes of the shift, every work order is a lie that the AI will memorize.
Your PM completion is honest. A PM that was "completed" but skipped the inspection steps is worse than a PM that wasn't done at all. Because the skipped PM creates a false record of health. The AI sees a clean inspection history and concludes the asset is fine. The asset is not fine. Nobody looked.
Your planned work ratio is real. If less than 40% of your labor is going to planned work, you are reactive. Calling yourself preventive on a PowerPoint does not change the ratio. Fix the ratio before you layer intelligence on top of it.
Your people have time. If your maintenance team is running from breakdown to breakdown, they don't have the bandwidth to generate good data, let alone review what an AI recommends. You can't automate your way out of a staffing problem by giving the same overworked people a new tool.
None of that is glamorous. None of it fits on a vendor slide. All of it is the actual work.
The vendor won't tell you this
The vendor selling you predictive maintenance has a product that works. That is true. The sensor is accurate. The algorithm is real. The cloud dashboard is pretty.
What the vendor will not tell you is that their product only works in environments that have already done the boring, expensive, years-long work of building data discipline, work-order quality, and inspection consistency. The demo environment they showed you had all of that. Your plant does not.
They also won't tell you that the pilot was designed to succeed. One asset. Their best sensor. Their data science team cleaning the inputs. Their dashboard tuned to that specific failure mode. The pilot was a controlled experiment. Your plant is not a controlled experiment. Your plant is a place where somebody pulled a bearing from a kit last Friday for a hot job and nobody restocked it.
The gap between "the pilot worked" and "this scales across 400 assets" is not a technology gap. It is an execution gap. Planning discipline. Close-out quality. Parts availability. Inspection consistency. Communication between shifts. The boring stuff.
AI doesn't solve the boring stuff. The boring stuff is what makes AI possible.
What to do Monday
If your plant is in the early stages of an AI conversation, or if you've been through a pilot and are stuck wondering why it won't scale, here is what I'd do this week:
Audit your failure codes. Pull a report of the top ten failure codes by frequency across the last twelve months. If "other" or any equivalent catch-all is in the top three, your data is not AI-ready. Fix the coding structure first.
Read five random closed work orders. Not the ones your planner flagged as examples. Five random ones. Read the close-out notes. If three of the five say something like "repaired" or "replaced part" with no useful detail, that is what the AI will learn from. That is what it will become.
Ask your reliability engineer one question. "If I gave you twelve months of our CMMS data and asked you to predict which assets will fail next quarter, could you?" If the answer is no, the AI can't either. It just won't tell you it can't.
Pick five assets, not 500. If you're going to pilot, build the data discipline on five critical assets first. Clean failure history. Honest PM records. Consistent inspection data. Prove you can sustain that discipline for six months. Then talk about sensors.
Protect your experts' time. If your AI strategy requires your most experienced technicians to also train the model, validate its outputs, and troubleshoot its failures, you don't have an AI strategy. You have a plan to burn out your last remaining experts faster.
Save this before your next digital transformation meeting.
If this landed, forward it to the person at your plant who keeps getting asked "when is our AI strategy coming." They need ammunition.
