AI for Maintenance Operations: Which Project to Run First

In 2021, McKinsey described a predictive maintenance system that did what its vendor promised and still lost money. The model caught about a quarter of breakdowns at 85% accuracy, and each one it caught saved over 10% of the repair cost. Its false-positive rate was about 10%, and on a product serviced four times a year at roughly $1,500 a visit, that meant about 1,000 extra service calls a year. Those calls erased projected savings of over $1 million, and service events rose 10%, which added to total shutdown time.

That example explains a number from this year. MaintainX surveyed 2,234 maintenance and operations leaders across the US and Canada for its 2026 State of Industrial Maintenance report and found 58% already using AI, with 75% reporting measurable return inside six months. The same survey found 79% of teams with unplanned downtime flat or rising over the past year. Adoption climbed through the year while downtime stayed where it was.

This post is for directors of maintenance and reliability, vice presidents of operations, and the plant, facilities and fleet managers who report to them, in any industry that runs equipment. You get a four-signal map that sorts every maintenance AI use case by the data it consumes, and four ordered checks that tell you which project to run first. A table then shows where that first project usually lands in six industries.

For a team that spends under 40% of its hours on planned work, and MaintainX’s 2026 survey puts half of all teams there, learned failure prediction is the fourth AI project rather than the first. The three projects before it run on data you already hold, and they are what move a team from firefighting to planned work in the first place.

What does AI for maintenance operations cover?

AI for maintenance operations covers four families of software, and each family reads a different kind of data. Knowledge AI reads the words in work orders and manuals. Planning AI works on the schedules, criticality rankings and parts records that sit in the same system, and vision AI reads inspection photos and camera feeds. Condition AI reads vibration, current and temperature streams to flag assets drifting toward failure. Predictive maintenance is one use case inside that fourth family, and vendor marketing has made it stand for the whole field.

The families differ in what they need from you, which is why the distinction matters. A computerized maintenance management system (CMMS) assistant such as Fiix MAX, released in July 2026, answers questions like which assets had the most downtime last month from records the team already keeps. IBM Maximo Application Suite 9.0 added generative AI assistants for work-order intelligence in June 2024, and MaintainX ships a report builder that answers questions about labour, inventory and asset performance. None of those needs a new sensor.

A condition-monitoring platform such as Siemens Senseye or Augury sits at the other end. It needs sensors on the asset and months of readings before it can say anything useful, and a learned failure model needs examples of failure on top of the readings. Our AI consulting and strategy work with maintenance teams starts by establishing which of the four kinds of data a site holds in usable form, because that answer decides the order of everything else. Each family needs its own signal, so the useful question for your site is which signal is ready today

Why do maintenance AI pilots stall when adoption is this high?

Maintenance AI pilots stall for two reasons that have little to do with model quality. The first is that the team cannot act on what the model says, because its hours are already consumed by breakdowns. The second is that the model was trained on too few failures to separate a real warning from noise, so its alerts cost more than the failures they prevent. Both reasons show up in the 2026 survey data and in McKinsey’s published case.

Capacity is the first of the two reasons. MaintainX’s 2026 survey found half of all teams spending less than 40% of their time on planned work, and it ranked labour shortages and poor knowledge transfer among the top causes of unplanned downtime. Downtime is also getting dearer, with 39% of leaders saying so in 2026 against 31% a year earlier. Siemens’ True Cost of Downtime 2024 report was built on 181 interviews across automotive, consumer goods, heavy industry and oil and gas. It estimates that the world’s 500 largest companies lose about $1.4 trillion a year to unplanned downtime, 11% of their revenue. The average large plant in that study lost 27 hours a month across 25 incidents, and average recovery time had grown from 49 minutes to 81 over five years.

A team in that state receives a prediction the way it receives a breakdown. Someone has to find the hours, the parts and the permit, and the job joins a backlog that is already late. The prediction was correct and the asset still failed on schedule, which is the outcome the pilot gets judged on.

Data is the second of the two reasons. A learned failure model needs examples of failure, and failures on well-maintained equipment are rare by design. Practitioners writing for the Institution of Engineering and Technology put failure records at under 1% of the telemetry a programme collects. A 2025 study of automotive fleet data in Springer’s Discover Applied Sciences describes failure instances as scarce even when sensors record continuously, partly because the failures that do happen often go undocumented. A classifier that calls every reading normal scores 99% accuracy on data like that and prevents nothing. The sensors themselves are often in place, since Siemens found 87% of the manufacturers it surveyed collecting data that could support predictive maintenance, with about half collecting at least one of current, vibration or temperature. What those sites lack is the labelled failure history to train on.

If the choice in front of you is whether to build the model, buy a platform or hold off, our build vs buy AI decision framework scores that call on five factors. For maintenance specifically, the two reasons above decide the order of projects before any sourcing question comes up.

A prediction only has value when a planner can turn it into scheduled work before the asset fails, and half of all teams have no room on the schedule for it yet.

What is the Four-Signal Maintenance AI Map?

The Four-Signal Maintenance AI Map sorts every maintenance AI use case by the data it consumes, then attaches a readiness test to each signal that you can run on your own records inside a week. Words, schedules, pictures and sensor streams are the four signals. A signal that passes its test is ready for an AI project now, and a signal that fails tells you which data work comes first.

The Four-Signal Maintenance AI Map

The counts in the fourth column are the thresholds we use when scoping, and your own history can move them in either direction. A site where every work order already carries a failure code skips the tagging work in the first row. A site with a decade of vibration routes, on the other hand, may clear the fourth row on its first critical asset class without any new instrumentation.

What can AI do with the words in your work orders?

Knowledge AI turns years of work-order text into something a technician can search and a planner can count. It tags each past job with the asset, the symptom, the cause and the fix, drafts procedures from the jobs that went well, and answers questions from original equipment manufacturer (OEM) manuals at the machine. It needs no new sensors, which is why it is usually the first project that passes its readiness test.

The US National Institute of Standards and Technology (NIST) has studied this problem for years. Its researchers Thurston Sexton and Michael Brundage built Nestor, an open-source tool for tagging the free text in maintenance work orders. The tool exists because technicians write lines like “hyd leak at saw attachment” while the structured fields stay empty. A 2021 analysis hosted by NIST of 405 work orders from one building found about 13% carrying an asset number in the structured field. Extracting asset numbers from the text lifted that to 22%. That is the usual state of the words signal, rich in the text and thin in the fields. It is also why the readiness test counts asset numbers and descriptions rather than the number of records.

A large language model (LLM) pipeline does at scale what Nestor does with a human tagger. The first pass tags the backlog of past work orders with asset, symptom, cause and fix, with a senior technician correcting a sample until the tags hold. The second pass puts retrieval-augmented generation (RAG) over the tagged history and the OEM manuals. A technician at a failed asset can then ask what fixed this fault last time and get the work order, the parts used and the procedure. We run a document AI pipeline for one client that processes 50,000+ documents a month across 47 formats and cut manual review by 80%. Work orders are shorter documents with the same shape of problem.

The RAG development and AI agent and copilot work that follows is where the assistant learns your site’s vocabulary rather than the vendor’s.

The CMMS vendors’ own assistants cover the question-answering layer over their own records, and for a single-system site with clean fields they are the right first stop. Custom work earns its place when history sits across several systems, on paper, or in a retired technician’s memory. Either way, keep AI-drafted work orders behind a human approval until the share approved without edits passes 90% for a full month, which is the rule we use before any drafting assistant goes live.

What can AI do with your preventive maintenance schedules and parts records?

Planning AI reviews each preventive maintenance (PM) task against the failures it was meant to stop. It also orders the backlog by asset criticality and risk, and forecasts the parts a job will need before the technician opens the cabinet. It runs on the schedule and parts data a CMMS already holds, so it passes its test wherever the asset hierarchy and PM compliance are maintained.

The case for reviewing intervals is old and well documented. Nowlan and Heap’s 1978 study of United Airlines failure data, the report that reliability-centred maintenance rests on, found only 11% of the items it studied followed an age-related failure pattern. Time-based replacement on the other 89% swaps parts that were fine and sometimes introduces the fault it meant to prevent. Planning AI finds those tasks in your own history. A PM completed 40 times without a recorded defect is a candidate for a longer interval. A PM whose asset keeps failing between visits is a candidate for condition monitoring rather than a shorter one. Deloitte’s 2017 estimate from its client work puts maintenance planning time 20% to 50% lower with predictive techniques in place, with maintenance costs 5% to 10% lower.

Two industries cap what this family may change. In US hospitals, 42 CFR 482.41(c)(2) requires equipment to be maintained to manufacturer recommendations unless the hospital runs a documented Alternate Equipment Management (AEM) programme, and imaging equipment stays on the manufacturer’s schedule regardless. Pharmaceutical plants answer to 21 CFR 211.67 which requires written procedures for equipment maintenance, so an AI-suggested interval change goes through change control like any other. For both, planning AI works as a review queue that a qualified person signs off, which is how we build it for pharma and healthcare clients.

What can AI do with inspection photos and camera feeds?

Vision AI replaces the scheduled walk-round for defects a camera can see, which includes belt wear and tracking, leaks, corrosion, blocked filters, gauge readings and missing guards. It flags frames for a person to review rather than closing work orders on its own. It passes its test when you can gather about 200 labelled images of each defect class and mount a camera where the defect shows.

The stack is the one we use for production inspection. Our tire inspection system runs YOLO-based detection on an NVIDIA Jetson Orin with TensorRT FP16. It reaches 99.2% defect detection at 200+ tires an hour, with decisions under 50 milliseconds, on a floor that reaches 42°C ambient. Pointed at a conveyor return roller or a filling-line guard, the same computer vision development stack watches for the defect a technician would otherwise walk out to find. Running inference at the asset on edge hardware matters here, because a stop decision cannot wait for a round trip to the cloud and many plant networks cannot carry continuous video anyway.

Two capabilities separate this work from a generic video analytics package. Some defects are dimensions, such as belt sag, a gap at a seal or the volume left in a hopper. For those, a Stereolabs ZED 2i stereo camera gives metric depth in real metres where a standard camera can only guess. And after a failure, the question is usually about the minutes before it. Our Intelligent NVR answers prompts such as “Show me the Dock 4 conveyor in the ten minutes before the jam” across every camera on site. It handles “Find every forklift that entered the chemical store this month” the same way. That turns the existing recorder into a maintenance record. If you are not sure what your recorder can hold, our guide to what an NVR does covers the five numbers that cap it. In warehouses, the Occupational Safety and Health Administration’s rule at 29 CFR 1910.178(q)(7) requires forklifts to be examined before being placed in service. A camera at the charging bay can confirm that the examination happened, and that is one of the

logistics

use cases that clears its readiness test fastest.

What can AI do with vibration, current and temperature streams?

Condition AI reads sensor streams in three steps of rising difficulty. Thresholds raise an alert when a reading leaves its normal band. Anomaly detection learns what normal looks like for one asset and flags drift without any failure examples. Learned failure prediction, the step vendors call predictive maintenance, needs labelled failures to train on, and it carries the strictest entry ticket of anything on the map.

The published returns describe mature programmes rather than first pilots. McKinsey’s 2017 manufacturing analytics research states that predictive maintenance typically reduces machine downtime by 30 to 50 percent and increases machine life by 20 to 40 percent. The US Department of Energy’s Operations and Maintenance Best Practices Guide is more conservative and names its baseline. It estimates 8% to 12% savings for a predictive programme over a preventive one, and savings opportunities exceeding 30% to 40% for a facility that relies on reactive maintenance. Those two baselines explain the order of projects in this post. A reactive shop gains more from reaching planned work than from any model, and the first two families are what get it there.

McKinsey’s 2021 article is also where the middle step earns its place. It reports a technology manufacturer that cut labour, downtime, parts and related costs by 30% with condition-based maintenance. It reserves learned prediction for assets where a breakdown is costly or dangerous enough to justify the false alarms, or where failures are predictable and false positives are cheap. Our rule in scoping is the one in the map. Fewer than five labelled failures per failure mode on an asset class in 24 months means thresholds and anomaly detection for that class, with the failure labels collected from here on. Where the plant network cannot carry raw vibration data, inference runs on an NVIDIA Jetson at the asset and only the alert travels, which is how we deploy for manufacturing clients with air-gapped lines.

In what order should you run the four maintenance AI projects?

Run four checks in a fixed order and start the first project the checks point to. Planned-work share comes first because it decides whether anyone can act on an alert, and the work-order export comes second because it is the cheapest data to test. Third is the failure count on your critical assets. A vision check runs alongside whichever project wins, because a camera needs no history.

1. Measure your planned-work share.
   Use the Society for Maintenance and Reliability Professionals (SMRP) definition, where reactive work is work that interrupts the weekly schedule, and count labour hours. If planned work is under 40% of hours, Knowledge AI and Planning AI are your first two projects, in that order, because they raise that share. Hold learned prediction until planned work passes half of hours.
2. Export 12 months of work orders.
   Count the share with an asset number and a description of ten words or more. At 60% or above, start Knowledge AI now. Below it, the first project is a six-to-eight-week tagging pass in the Nestor pattern, with an LLM drafting the tags and a senior technician correcting a sample until the error rate holds under 5%.
3. Count labelled failures on your 20 most critical assets over 24 months.
   Five or more per failure mode on an asset class, with sensor data covering the days before each one, clears learned prediction on that class. Fewer clears thresholds and anomaly detection only, and the count restarts from today with failures labelled at close-out.
4. Pick vision where inspection is a scheduled human walk.
   List the rounds a technician does with eyes and a camera phone, then pick the one with the highest hours or the highest safety exposure. Gather 200 labelled images of the target defect before you buy a camera or a model.

Two sites show how the order changes with the data [ILLUSTRATIVE]. A 400-person food plant measures planned work at 35% of hours and finds 71% of its work orders carrying an asset number and a description. It counts two labelled failures per mode on its critical compressors, and it runs a weekly thermography round on the palletiser. Its order is Knowledge AI in the first eight weeks and Planning AI on the PM schedule next, with a camera on the palletiser conveyor in parallel. Anomaly detection on the compressors follows once the first two projects have lifted planned work past half. Learned prediction waits for 24 months of labelled failures.

A 600-bed hospital’s facilities team measures planned work at 55%, because its AEM programme already forces scheduling. It finds only 40% of its work orders carrying an asset number, since requests arrive as free text from the wards. Its order starts with the tagging pass and moves to Knowledge AI on request triage and work-order drafting. Planning AI then runs as a review queue inside the AEM boundaries, and anomaly detection on the air handlers comes last, fed by the building management system’s temperature and pressure readings.

Where does maintenance AI land first in each industry

The first project lands in a different place in each industry because the four signals mature at different rates. Manufacturing is where the sensor data already exists, with Siemens finding about half of large manufacturers collecting current, vibration or temperature readings. Hospitals and pharma plants carry regulated maintenance intervals. Data centres lose more major outages to power and procedure failures than to any learned pattern, by Uptime Institute’s count, which points them at words first.

Where the first maintenance AI project usually lands, by industry

Uptime Institute’s 2025 outage analysis found nearly 40% of organisations had suffered a major outage caused by human error in the past three years. The share of those caused by failure to follow procedures rose ten points on the year before. Network World’s coverage put power behind 54% of major outages. Data Center Knowledge, reporting on the same analysis, had human error accounting for two-thirds to three-quarters of outages. That second number is the one a procedure assistant exists to move. For a financial services data centre, an assistant that puts the right procedure in front of the technician at the switchgear addresses the cause the data names. The same logic sends a construction fleet manager to service records before sensors, because a 40-machine fleet produces too few failures per model to train on. It sends a warehouse operator to cameras, where conveyors and forklifts give a defect class hundreds of examples a month. Within the six industries, the signal that is ready first is the one your people already produce without being asked.

Turn the four checks into a brief a vendor can quote against

Get the AI Project Scoping Template and record your planned-work share, your work-order export result and your failure count on one page, so every proposal you receive is priced against the same facts.

How do the ways of buying maintenance AI compare?

You can buy maintenance AI four ways. You can switch on your CMMS vendor’s AI features, subscribe to a condition-monitoring platform, hire a generalist AI agency, or engage a specialist that builds on your data and hands you the result. Each route covers a different part of the map, and the first two are often the right answer for the signal they cover.

Brainy Neurals works on the fourth route, so read that column knowing where we sit, and hold it to the same standard as the other three.

Four ways to buy maintenance AI, compared

Two of those columns deserve a straight recommendation. If your site runs one CMMS with clean fields and your question is which assets cost you the most, switch on the vendor’s assistant before you talk to anyone. It covers the words signal for its own records at subscription cost. If your plant already instruments its critical assets and wants thresholds and anomaly alerts, a condition-monitoring platform covers the fourth signal faster than a custom build. The specialist route earns its place when the history sits across systems, when the work has to run at the edge, or when you want to own what gets built. Our guides on how to choose an AI development company and on AI proof of concept development cover the selection and the pilot. If you expect a model to live on your floor for years, a digital twin is where the condition data ends up.

Mitesh Patel, on where maintenance AI starts

“Across the 40+ AI systems we’ve shipped to production for 12+ years, the maintenance teams that got value fastest started with the words their technicians had already written. One document pipeline we run handles 50,000+ documents a month across 47 formats and cut manual review by 80%, and a work-order backlog is the same problem in shorter documents. Our tire inspection system holds 99.2% defect detection at 200+ tires an hour on an NVIDIA Jetson, and even that project began with 200 labelled images before anyone bought a camera. Prediction comes when the failures have been counted, and at half of the sites we scope that count is still under five.”

Mitesh Patel, NVIDIA Certified AI Architect, Founder & Director, Brainy Neurals

What do the numbers say about AI in maintenance operations?

Five figures frame the decision in front of you. The first two size the problem and show how little adoption has moved it, the third and fourth set expectations for predictive programmes, and the fifth shows what a words-first project returned for one client.

Read each figure with its method in mind. Siemens extrapolates from 181 interviews and calls its cross-sector totals indicative, and its own product-performance claims are left out of this post. MaintainX surveyed 2,234 leaders in two countries. The Department of Energy figures date from the guide’s 2010 release and describe older industrial studies, and McKinsey’s case is a single programme. The Brainy Neurals figure is one client’s pipeline, not a maintenance deployment, and it is cited here for the shape of the work.

When does the Four-Signal Map give the wrong answer?

The map gives the wrong answer when one of its four inputs is measured wrong or when a rule outside the data sets the interval. Five conditions produce that result, and each has a fix you can apply before the first project is scoped.

• Planned work was measured as PM completion. A site that counts completed PM tasks as planned work overstates the share. Count labour hours against the weekly schedule, as SMRP defines it, and rerun the first check.
• The work-order text is too thin to tag. If the median description is under ten words, Knowledge AI has nothing to read. Put four structured prompts on the mobile close-out form, symptom, cause, action and parts, for 90 days before the tagging pass.
• The interval is set by a regulator. Under the AEM rules in US hospitals and under good manufacturing practice in pharma, a model’s suggested interval has no standing on its own. Run Planning AI as a review queue with a qualified signatory, and log the decision with the evidence the model showed.
• False alerts cost more than the failures. McKinsey’s 2021 case is the warning here. Set an alert budget per asset per month in scoping, and count every false alert as a truck roll or a teardown at its real cost. Tune the model to the budget rather than to accuracy.
• The alert has no named owner. MaintainX co-founder Nick Haase’s line in the 2026 report’s release, that reliability gains come from execution maturity rather than system adoption alone, is the whole story here. Every alert becomes a CMMS work order assigned to a named planner, or the integration does not ship.

When does the Four-Signal Map give the wrong answer?

The map gives the wrong answer when one of its four inputs is measured wrong or when a rule outside the data sets the interval. Five conditions produce that result, and each has a fix you can apply before the first project is scoped.

• Planned work was measured as PM completion.
  A site that counts completed PM tasks as planned work overstates the share. Count labour hours against the weekly schedule, as SMRP defines it, and rerun the first check.
• The work-order text is too thin to tag.
  If the median description is under ten words, Knowledge AI has nothing to read. Put four structured prompts on the mobile close-out form, symptom, cause, action and parts, for 90 days before the tagging pass.
• The interval is set by a regulator.
  Under the AEM rules in US hospitals and under good manufacturing practice in pharma, a model’s suggested interval has no standing on its own. Run Planning AI as a review queue with a qualified signatory, and log the decision with the evidence the model showed.
• False alerts cost more than the failures.
  McKinsey’s 2021 case is the warning here. Set an alert budget per asset per month in scoping, and count every false alert as a truck roll or a teardown at its real cost. Tune the model to the budget rather than to accuracy.
• The alert has no named owner.
  MaintainX co-founder Nick Haase’s line in the 2026 report’s release, that reliability gains come from execution maturity rather than system adoption alone, is the whole story here. Every alert becomes a CMMS work order assigned to a named planner, or the integration does not ship.

A confession from our scoping work

We’ve rewritten more than 40 AI specifications mid-project at this point, and the maintenance ones share a pattern I can now spot in the first meeting. The specification asks for failure prediction on assets whose failure history lives in a retired technician’s memory and a dozen spreadsheets. The fix is rarely a better model, however good the vendor demo looked. It is six weeks of tagging the work orders the team already wrote, a project that seldom appears in the original budget because it does not demo well.

What should you check before your next maintenance AI budget meeting?

Bring four numbers to the meeting and let them set the agenda. This week, pull labour hours against the weekly schedule and compute the planned-work share. Export 12 months of work orders and count the ones with an asset number and a real description. Pick your 20 most critical assets and count the labelled failures per mode over two years. Then list the inspection rounds your technicians walk with a camera phone. Those four numbers tell you which signal is ready, and a vendor proposal that ignores them is pricing a project for a different site. If the numbers point at words or schedules first, our AI consulting and strategy team scopes those projects at fixed price, and we will tell you when your CMMS vendor’s own assistant covers the job.