Quick answer: AI Predictive Maintenance vendors promise outcomes in 90 days. The honest math is 12 months minimum from kickoff to a model that actually fires correct work orders. The 12-Month Rule says: any plant that cannot satisfy four foundational prerequisites cannot fast-forward to AI PdM, regardless of how much you spend. This guide explains the prerequisites, the realistic readiness sequence, and the decision matrix for buyers evaluating AI PdM right now.
Related deep-dives: increase MTBF using native OEE · why OEE improvement stalls · closing the OEE-CMMS loop · Computer Vision OEE.
Every AI Predictive Maintenance vendor pitch goes like this: "Plug us in, we train on your data, you save 30% on downtime by Q2." The 90-day pitch is real in slide deck math. In plant-floor reality, no AI model can predict failures before the data exists to learn from. The 12-Month Rule is the honest sequence between "we bought AI PdM" and "AI PdM works in our plant."
The 12-Month Rule says: the time from kickoff to a production-ready AI PdM model is at least 12 months, regardless of vendor promises. The reason is not vendor competence, it is data physics. A predictive model needs a minimum of 12 months of cleanly-coded failure data per asset class to find useful patterns. If you do not have that data today, no shortcut exists.
AI PdM is supervised learning. It needs labeled examples: "here is what an asset looked like before it failed". Most plants have 12 months of work-order data, but the failure codes are generic ("general repair", "misc") instead of specific ("bearing seizure", "belt tension drift"). Clean failure-coding is the first 3-month sub-project of the 12-Month Rule.
AI cannot predict failure from work orders alone. It needs sensor data: vibration, temperature, current draw, acoustic emission, oil quality. Most plants are sensored on 10–20% of their critical assets. The 12-Month Rule mandates 80%+ coverage on the assets you actually want predictions for.
The model output is a predicted failure with a recommended action. That recommendation must flow into a CMMS that creates a work order, routes it to the right maintainer, tracks completion, and feeds the outcome back to the model. If your CMMS is a file cabinet, the model has no place to send its predictions.
This is the prerequisite the vendor will never name. AI PdM works only when maintainers act on model predictions, not when they say "the model doesn't know my plant". Building that trust requires transparent model explanations, conservative initial thresholds, and a culture that values data over heroic individual diagnoses. This takes longer than any technical task.
Audit your last 12 months of work orders. Recode generic entries with specific failure modes. Verify your asset registry, every critical asset has a unique ID, a hierarchy parent, and a criticality rating. This is unsexy and essential. Most plants discover 30–60% of their data is unusable until they do this work.
List the top 10 critical assets where prediction would change the economics. Add sensors where missing, vibration for rotating equipment, temperature for bearings and motors, current for drives, acoustic for compressed-air systems. Budget realistically: €5,000–€15,000 per asset for proper coverage.
Wire sensor data, asset history, and CMMS into one pipeline. Sensors → time-series database → CMMS event bus. Webhook from CMMS into the model when a work order closes (so the model learns from the outcome). The integration is the most underestimated part of the 12 months.
Train the model on your now-clean, now-complete data. Validate predictions against held-out data. Tune thresholds conservatively. Run a shadow phase where the model fires predictions but maintainers act on them only if the supervisor approves. Convert to live operation once trust is established.
Track this number: % of AI-predicted failures that were prevented before causing downtime. Starting baseline at month 12: 25–40%. Mature operation at month 24: 60–75%. A plant above 80% is using AI PdM as designed. Below 25% at month 12 means a prerequisite is not yet satisfied, go back and fix it.
This is a programme, not a purchase. Read the OEE software pricing breakdown, the Intelligence Gap article, and the closing the OEE loop guide for the foundation underneath AI PdM.
Either (1) you already have the four prerequisites and they will deliver, or (2) they are training a generic model on assumptions and you will get false positives that destroy team trust in month two. Ask for three reference customers at your asset count and complexity. If they can't produce them, the 90-day claim is marketing.
Limited. Work-order-only AI can spot frequency patterns but cannot predict acute failures. Combining work-order AI with computer-vision OEE (no PLC needed) is a middle path for plants without sensor budget.
Condition-based maintenance acts on threshold rules (vibration > X, temperature > Y). AI PdM uses pattern recognition across multiple signals to predict failure before any single threshold trips. CBM is a prerequisite for AI PdM, not an alternative.
Generative AI (LLMs) is excellent for work-order summarization, technician assistance, and knowledge-base search. It is not predictive maintenance. Do not confuse the two when evaluating vendors.
AI Predictive Maintenance is real, and it works, once the foundation is in place. The 12-Month Rule exists because the foundation cannot be skipped. Vendors who promise faster are either selling a generic model that will produce false positives or counting on you already having the foundation. Audit your data, sensor coverage, CMMS readiness, and team culture before signing. If three of four are missing, do the 12-month foundation work first; the AI conversation goes much better afterward.
Login
Blog
