Predictive Maintenance: Why It Stalled and Where the Value Lies Next

On 19 October 2026, IIoT World published a summary of a panel at the Industrial AI Summit 2026 built around an uncomfortable question: why has predictive maintenance lost its position as the number one industrial AI use case? The discussion featured Boris Scharinger (Siemens), Karthikeyan Natarajan (CEO of Infinite Uptime) and Jan Pingel (Ingersoll Rand), moderated by Rob Tiffany (IDC).

A necessary caveat: the article is labeled as sponsored by Infinite Uptime, although it states that it is editorially independent. The company's performance figures are therefore vendor claims and should be read as such.

A safe use case, but with an economic ceiling

According to the panel, predictive maintenance was the first AI use case precisely because it does not interfere with a running line. But most installations stopped at anomaly detection, which boils down to a generic "something is wrong, go check."

There is also an economic limit. As a rule of thumb, the speakers said that eliminating all unplanned downtime on a $1 million asset is worth roughly $1,000 a month, or $12,000 a year. A ceiling that low has made it hard for vendors to build a sustainable business model.

True prediction, they explained, requires classification: not "anomaly" but, for example, "bearing fault, replace within 10 days." This calls for vibration and sound analysis and models trained on a large and varied machine fleet.

Operator trust has a threshold

The most concrete point concerns accuracy. Operators who receive 30–40% false positives abandon the system within a month. For the panel, lasting adoption requires more than 95% true positives and fewer than 5% false positives.

Infinite Uptime reports the following among its results:

• 99.97% accuracy across 55,000 prescriptions, with 95% acted on by operators;
• 175,000 hours of downtime avoided;
• 8 years of model development across more than 1,000 plants in 27 countries, of which over 600 are in the scale-up phase and over 300 have 70–80% coverage of critical equipment.

The article also cites a cost of unplanned downtime for the manufacturing industry of about $100 billion, a figure reported as stated in the text. The panel added that value must be demonstrable within 3–6 months.

The next pool of value: redundancy

According to the speakers, plants bear an additional cost of 30–40% for standby lines, backup equipment held in stock for 12–18 months, and on-call technicians. Accurate remaining useful life predictions could reduce this redundancy. Scharinger expects predictive maintenance to return to the top three AI use cases in about five years.

On the product side, over the past three months Infinite Uptime has introduced "Verticalized AI Shields," which combine process data, inspection records, maintenance history, manufacturer procedures and technician input across 18 asset categories. Shields are planned for steel, chemicals, food and beverage, and cement.

What to do with this on the plant floor

The article does not specifically address Italy or MES systems, but its takeaways transfer to anyone evaluating solutions of this kind. Some useful questions to put to vendors:

• What does the system output: a generic alarm or a prescription with diagnosis and timing?
• How many false positives does it produce, and how were they measured?
• What return can it demonstrate within 6 months?
• What asset, process and maintenance context data does it need, and is it already available?

An alert without a prescription costs you operator trust.

This last point is perhaps the most underestimated: without well-structured data on assets, process and interventions, even the best model remains of little use. Before choosing the algorithm, it is worth checking the quality of the context in which it will have to operate.