Prescriptive AI for Maintenance: Layering Diagnosis and Decision-Making

On October 8, 2026, IIoT World published a Q&A covering eight audience questions raised during the session "Unlocking Plant Reliability with Verticalized Prescriptive AI" at the Industrial AI Summit 2026. The respondents are Boris Scharinger (Siemens, DataOps and Master Data Management, author of "Industrial AI: From Pilot to Profit") and Karthikeyan Natarajan, CEO of the vendor Infinite Uptime. The content is labeled as sponsored by Infinite Uptime, while stating that it is editorially independent, so it is worth separating statements of principle from commercial claims.

Prescriptive AI means a system that does more than flag an anomaly: it indicates which component to replace, which maintenance action to perform, and when.

Data: use case first or strategy first?

Scharinger describes the choice between a use-case-driven and a data-strategy-driven approach as an almost ideological divide. His point is that the data layer is a foundational capability, whose cost cannot be charged to the ROI of a single use case.

Natarajan, speaking from the vendor's perspective, argues that waiting for clean data is a "multi-year tax". He claims that a physics-aware approach can deliver a first prescription in two weeks and cover the whole plant in 90 days, starting with the most critical and best-instrumented assets. These timelines are vendor-stated and have not been independently verified.

Existing plants: sensor retrofit

For older plants, Scharinger points to sensor retrofits, including machine vision, as a valid path; it can also strengthen independence from machine builders. According to Natarajan, vibration detects mechanical degradation, while vision and thermography catch process-induced failures and quality problems: combining them broadens coverage.

A layered diagnostic architecture

The most useful part for system designers is how the techniques are combined. The two views are close but not identical:

Key points that emerged:

• Anomaly detection is only the starting point, and anomalies still need to be reviewed by people.
• Classification requires labels, which AI can help produce; the labels are linked to documentation via RAG+LLM.
• For now, Scharinger would keep anomaly/classification AI separate from look-up AI.
• Natarajan proposes a sequence: physics-informed failure signatures, expert verification of every prescription, and RAG+LLM only as an explanation and search layer.
• In his view, using an LLM to detect physical failure modes is a real risk.

Replace or predict? A decision model

Scharinger offers no empirical thresholds. He recommends a quantitative model for each use case, including the cost of unplanned downtime, the remaining useful life lost through early replacement, and the cost of planned replacement (downtime plus labor). He suggests adding a Monte Carlo simulation to handle uncertainty.

For auditing the algorithms and business assumptions, including defining the risk approach and Monte Carlo, he estimates roughly 2–3 people for 6 weeks; Siemens did not charge the cost to the business units involved.

Why companies underinvest

Scharinger invokes psychology: in medium-to-low-risk areas, organizations get used to downtime; in high-risk areas, redundancy is already built in. The available text cuts off during this answer, so we do not know how it continues.

Practical takeaways

The picture suggests layering: rules or physics first, then ML classification, and finally an LLM for documentation, always with a person validating. The cost-based decision model avoids treating prediction as an end in itself. Sensor retrofits show that you need not wait for perfect historical data to start, but the promised timelines should be validated with a pilot on a few critical assets.