Predictive Maintenance With Machine Learning: Promised Savings, Few Field-Validated Results

The use case

In a 2022 peer-reviewed case study, researchers at a university in Jakarta fitted a tube-filling machine with sensors and a model that predicts which components will fail. After three months, they reported that overall equipment effectiveness (OEE, a standard measure of how much planned production time is productive) rose by 13.10% and unplanned failures in the observed part fell by 62.38% (measured without a control group, abstract)³. The result is typical of the field: encouraging, small and hard to generalise.

Predictive maintenance means using machine data to forecast failures so that repairs happen just before they are needed. Only part of this work is AI. In this paper, AI means machine-learning models trained on sensor data to predict failures or remaining life. Older methods are kept apart: fixed alarm thresholds, vibration rules, statistical process control and calendar-based preventive maintenance. Several consultancy sources do not say which method sits behind their results, so those results are not credited to AI. Predictive quality control, maintenance scheduling optimisation and assistants for technicians are out of scope.

Where it is running

No source opened gives a count of plants running machine-learning predictive maintenance. The nearest official figure covers all AI in all functions: 19.95% of EU enterprises with 10 or more employees used at least one AI technology in 2025 (official survey, all AI types)¹. Eurostat lists production processes as a purpose for AI in 33.46% of large and 19.02% of small enterprises using AI in 2025, but the text read does not show maintenance separately¹. Among manufacturers using AI, the most common purposes were marketing and sales (30.41%) and business administration (27.05%) in 2025¹. A U.S. Census-based working paper using 2018 data found advanced-technology adoption rare and skewed towards larger and older firms (abstract; all functions, not maintenance)⁹.

Named or described deployments are few. Beyond the Jakarta case³, the sources describe unnamed companies: a company that predicted about one-quarter of its breakdowns with 85 percent accuracy (2021, case, method not confirmed as machine learning)⁴, an oil producer whose downtime fell from 14 days to six (2017, case, method not confirmed)⁵, and a consumer goods company reporting USD 5 million in annual maintenance savings (2017, case, method not confirmed)⁶. Earlier deployments of this kind date back to at least 2017.

What the evidence shows

Strongest design available. A 2026 systematic review of 89 studies (64 peer-reviewed, 25 preprints, 2007 to 2026) found that 65.6% used weak or unclear validation, and that no study reached operational field validation beyond a temporal or cross-domain split of the data (systematic review, abstract)². This explains why published accuracy figures travel poorly: a model tested on held-out data from the same source is not tested against a running plant.

Single-site measured result. The Jakarta study reports random forest regression at 88% of actual data against 59% for linear regression, and the OEE and failure changes above (2022, abstract, no control group, three months)³. With one machine and a before-and-after comparison, other changes during the period cannot be excluded, and three months says nothing about whether gains last.

Consultancy cases. McKinsey reported a company that saved more than 10 percent on each predicted breakdown, with management estimating more than USD 1 million a year in savings (2021, single company, management estimate)⁴. The same case shows why accuracy alone misleads: a false-positive rate of about 10 percent produced roughly 1,000 extra cases a year that wiped out the overall savings, and shutdown events rose by 10 percent⁴.

Headline claims. McKinsey states that predictive maintenance typically reduces machine downtime by 30 to 50 percent and extends machine life by 20 to 40 percent (2017, no basis stated, method not specified)⁵. Deloitte estimates that poor maintenance strategies cut a plant's productive capacity by 5 to 20 percent and that unplanned downtime costs industrial manufacturers about USD 50 billion a year (2017, estimates covering all maintenance, not AI)⁶. These describe the size of the problem, not the effect of machine learning.

What it takes: cost, integration, governance

No official price or cost-per-unit figure was found for machine-learning predictive maintenance. The Jakarta study used low-cost components, which suggests that sensing hardware need not be the main cost³. The reviews point to other costs: data scarcity, shifting operating conditions, computing limits and distrust of black-box models (2026)², and data heterogeneity and imbalance, labelling costs and limited scalability (2025)⁷. Only 15.6% of the 89 reviewed studies adopted explainability methods (2026)², so crews often receive an alert without a reason.

On rules, Regulation (EU) 2023/1230 on machinery states that software with self-evolving behaviour using machine learning to ensure safety functions should undergo third-party conformity assessment, and that this should not apply to software incapable of learning (recitals 54 and 55, 2023)⁸. Predictive maintenance usually does not ensure safety functions, so these provisions are likely to matter mainly where a model is tied to a safety function; this is an inference, not a finding. The date from which the regulation applies was not confirmed.

Where it falls short

The central test of the usual story is whether the gains survive contact with a working plant. The evidence says they are unproven. Validation is weak in most studies², the one measured trial has no control and lasts three months³, and the headline percentages have no stated basis⁵. False alarms are the clearest failure mechanism: each one sends a crew to a machine that is not failing, and at a 10 percent rate one case lost its entire saving⁴. Gains also depend on people and process, because a prediction creates value only if planners and technicians act on it in time, and none of the sources measures that part.

No evidence was found on whether benefits reach smaller plants or the plants with the oldest machines, and Eurostat shows large enterprises using AI far more often than small ones (2025, all functions)¹. No withdrawn, paused or litigated predictive maintenance programmes were found in the sources opened; this reflects a narrow source base and is not proof that none exist. The deployments are drawn from the EU, one Indonesian case and unnamed companies; no named plant outside these could be confirmed.

Phase: In early use — why

The hypothesis was "In wide use". The evidence does not support it. The share of enterprises using AI in any function was 19.95% in the EU in 2025¹, maintenance is not separately reported, and the method behind most reported savings is unconfirmed. Where machine learning is confirmed, results come from single sites³ and no study in the 2026 review reached operational field validation². The lower of the plausible phases is therefore used.

What to watch next

• Eurostat lists an update to its AI-in-enterprises article for December 2026; a separate maintenance figure would show real adoption¹.
• Multi-site field studies with temporal validation, the main recommendation of the 2026 review, would show whether results hold outside one plant².
• Published false-alarm rates and net savings per predicted breakdown, which the 2021 McKinsey case showed can decide the outcome⁴.

Questions to ask before adopting

• What is the model's false-positive rate on our machines, and what does one false alarm cost in crew time and lost production?
• Was the quoted result measured against a control group or a longer period, or only on held-out historical data?
• Which failures are rare enough that there is too little labelled data to train on?
• Who acts on an alert, within what time, and how will their response be tracked?
• Does the model touch a safety function, and if so which conformity assessment applies?