The project sounds simple: you install sensors, you train a model, and maintenance costs fall. Six months later, the solution runs on a handful of machines, technicians barely use it, and costs have not moved.
This scenario is common, and the cause is rarely technical. A large share of industrial AI projects fall short of the expected outcome, most often for reasons of organisation and method. The sites that succeed stand out first through field adoption, not through the sophistication of the algorithm [McKinsey, 2023]. Mimorian is an industrial intelligence platform that models equipment, structures failure diagnosis and captures the know-how of maintenance teams through a multi-agent AI architecture. Here are the five mistakes that kill a project, and how to avoid them.
Mistake 1: deploying without the technicians
The classic scenario: the IT department selects a diagnostic AI, works for six months with a supplier, then presents the tool to the technicians. They look at a system that ignores the local context. Adoption never comes.
Technicians are the real experts. Without them in the design, the AI misses the reality of the field. Good practice: start from them. Involve them from the outset, ask them which diagnoses take up the most of their time, and co-build the guided diagnoses and alert thresholds with them.
Mistake 2: neglecting data quality
You have a CMMS with ten years of history, but with poorly categorised entries, inconsistent descriptions and empty fields. An AI trained on this data reproduces the disorder.
Good practice comes in four stages:
- Audit: sample the intervention reports to measure the share of degraded data.
- Standardise: build a dictionary of clear categories, validated by the technicians.
- Enrich: clean and tag the raw data before any training.
- Measure continuously: monitor the quality of new data through regular audits.
Data cleaning conditions the performance of everything else.
Mistake 3: aiming for prediction before knowledge capture
The stated ambition: to predict failures in advance. The problem: model-based prediction requires many documented failures of the same component. If a machine has broken down only a few times in ten years, the model produces mostly false alerts, and the technicians stop listening to it.
The order that works starts with knowledge, not with prediction:
- Knowledge capture first: document symptoms, causes and remedies so that the knowledge becomes reusable.
- Then guided diagnosis: ranked hypotheses that save the technician time across the whole fleet.
- Prediction last, on only the critical equipment that has enough history and sensors.
Mistake 4: ignoring change management
Experienced technicians who see AI as a threat quietly work around it. The tool is there, but no one uses it. Change management weighs at least as much as technology in the success of a project.
What helps:
- Train the managers so that they understand what the tool does, and does not, do.
- Repeat the right message: AI amplifies expertise, it does not replace it.
- Rely on champions: two or three respected technicians test it first and demonstrate it to their peers.
- Track adoption: usage rate, frequency, feedback.
- Make the gains visible: every failure resolved faster, every breakdown avoided.
This logic answers a fundamental challenge: 89% of industrial leaders acknowledge a talent shortage [Deloitte/Manufacturing Institute, 2018]. Keeping the experts' knowledge inside the plant becomes a matter of operational survival.
Mistake 5: failing to measure the result
Without measurement, the tool seems to be used, but no one knows whether it is useful. After a year, the finance department asks for the assessment, there is no answer, and the budget is cut.
Define the indicators before you start, and measure the baseline so that you have a point of comparison:
- MTTR (mean time to repair), the most telling indicator for maintenance.
- OEE (overall equipment effectiveness), for the impact on production.
- The cost of interventions and the share of breakdowns avoided.
A significant share of unplanned downtime comes from human error [Vanson Bourne/ServiceMax, 2017], precisely what guided diagnosis and captured knowledge help to reduce. It still has to be measured to be demonstrated.
Conclusion
The success of a maintenance AI project rests on five reflexes: involve the technicians from the outset, clean the data before training anything, start by capturing knowledge before aiming for prediction, take care of change management, and measure the result from day one.
For the complete framework of trustworthy AI in industrial maintenance, see our complete guide to trustworthy AI for industrial maintenance.