Why Predictive Maintenance Lost Its Top Spot in Industrial AI

On October 19, 2026, IIoT World published an article on why predictive maintenance (PdM) no longer tops the lists of industrial AI use cases. It summarizes a panel at the Industrial AI Summit 2026 featuring Boris Scharinger (Siemens), Karthikeyan Natarajan (CEO of Infinite Uptime) and Jan Pingel (Ingersoll Rand), moderated by Rob Tiffany (IDC).

An important caveat: the article is sponsored by Infinite Uptime, although it carries the label "Editorially Independent", and it states that AI tools were also used to summarize it. All performance data cited here are vendor claims and have not been independently verified.

Why it lost ground

PdM was the first "safe" AI use case on the factory floor because it does not interfere with a running line. The problem is the financial return, which has a ceiling. According to the rule of thumb cited, eliminating all unplanned downtime on a $1 million asset saves roughly $1,000 per month, or about $12,000 per year. With a ceiling like that, a vendor needs hundreds of plants and thousands of machines to build a sustainable business.

Anomaly detection versus true prediction

According to the panel, most deployments stopped at anomaly detection, which only says "something is wrong, someone should check." True PdM is a prescriptive classification, for example "bearing problem, replace within 10 days." This requires analysis of vibration and sound, and models trained on a large, diverse fleet of machines.

The operator trust threshold

With 30-40% false positives, operators abandon the system within a month. The panel puts the threshold for lasting adoption at more than 95% true positives and less than 5% false positives. Infinite Uptime claims:

• 99.97% accuracy across 55,000 prescriptions;
• 95% of prescriptions acted on by operators;
• 175,000 hours of downtime avoided.

The article also cites roughly $100 billion in annual unplanned downtime costs in manufacturing.

The vendor's approach

The models were built over 8 years across more than 1,000 plants in 27 countries. The "vertical AI Shields" cover 18 asset categories and add to vibration data the process data, inspection records (route-based, ultrasound, thermography), maintenance history, OEM procedures, and technician input in text, audio and video. The product was launched in the last three months; three shields are nearly complete, with steel, chemicals, food and beverage, and cement planned. Value must be demonstrable within 3-6 months: more than 600 plants are scaling up, and over 300 have reached 70-80% coverage of critical equipment.

The next pool of value: redundancy

Plants carry 30-40% in extra costs for standby lines, backup equipment held in stock for 12-18 months, and on-call technicians. Accurate remaining useful life prediction could make it possible to eliminate this redundancy at the design stage. Scharinger expects PdM to return to the top three use cases within five years.

Key takeaways

For those running or evaluating PdM projects, the article suggests a few checks:

• a pilot that only produces alarms is an anomaly detector, not predictive maintenance;
• operator trust, with false positives under 5%, matters more than model novelty;
• prescriptions need context: PLC and process data, maintenance records and asset classes;
• the business case should shift from maintenance planning to plant design.

Since this is sponsored content, it is wise to ask for verifiable evidence before accepting the metrics at face value. The underlying message, however, matches the experience of many projects: without concrete action and the trust of the people working in the field, predictive AI remains just a dashboard.