The integration of Artificial Intelligence (AI) into the industrial landscape has moved beyond the experimental phase. As manufacturing facilities and energy plants grapple with the complexities of digital transformation, the conversation has shifted from simple predictive maintenance to the more sophisticated, actionable domain of Prescriptive AI.
Unlike predictive systems that merely flag an anomaly, Prescriptive AI acts as a digital foreman, explicitly telling operators which component to replace, which specific maintenance action to perform, and the optimal timing for execution. During the "Unlocking Plant Reliability with Verticalized Prescriptive AI" session at the Industrial AI Summit 2026, industry experts Boris Scharinger of Siemens and Karthikeyan Natarajan of Infinite Uptime tackled the most pressing questions facing modern plant managers, challenging the status quo of how we define "data maturity" and "operational trust."
The Core Debate: Does Data Maturity Gating Hinder Innovation?
A central tension in the industrial world is the "chicken and egg" problem of data. Many organizations hesitate to deploy AI because their historical data is incomplete, messy, or siloed. The question posed to the panel was whether AI should be deferred until a company achieves a high level of "data maturity" or if it should be an engine for that maturity.
The Strategic View vs. The Physics-Aware Approach
Boris Scharinger, author of Industrial AI: From Pilot to Profit, suggests that the debate is often ideological. "CFOs look for ROI levers, CTOs look for strategic data capabilities, and CIOs often hope that existing ERP systems like SAP will solve everything implicitly," Scharinger notes. He emphasizes that a robust data layer is a foundational capital investment that cannot always be tied directly to a single use-case ROI.
Conversely, Karthikeyan Natarajan of Infinite Uptime argues that waiting for "clean" data is a multi-year trap. "Most plants will never have perfect historical data due to legacy equipment and sparse instrumentation," Natarajan explains. His solution is to leverage physics-aware models rather than purely data-dependent ones. By starting with known asset physics—such as vibration, thermal patterns, and vertical fault signatures—a system can provide actionable prescriptions within weeks, not years. In this model, the intelligence layer and data maturity grow in parallel, with the system generating structured, validated data as it operates.
Bridging the Gap: Retrofitting Legacy Plants
For the majority of industrial facilities, the challenge is not designing a "greenfield" site but upgrading "brownfield" legacy plants from Industry 3.0 to the standards of Industry 4.0 and 5.0.
Multi-Modal Sensing: A Path to Modernization
Attendees highlighted the use of cameras for computer vision and thermal monitoring on legacy milling machines to track both performance and material quality. The panel agreed that this is a gold-standard approach. Scharinger notes that sensor retrofitting enhances machine vendor independence, allowing operators to gain insights regardless of the age or origin of their hardware.
Natarajan adds that combining modalities—vibration sensors for mechanical degradation alongside vision and thermal sensors for process-induced quality issues—creates a comprehensive "asset reality." By capturing different physics-based failure modes simultaneously, operators can achieve a level of visibility that legacy sensors simply could not provide.
The Diagnosis Hierarchy: From Anomaly to Action
One of the most technical challenges in prescriptive maintenance is determining the right diagnosis approach. The panel addressed three prevailing methods: human review, supervised machine learning (requiring labeled data), and the increasingly popular RAG+LLM (Retrieval-Augmented Generation + Large Language Models) approach for documentation lookup.
The Panel’s Verdict: A Layered Strategy
Scharinger advises that these tools should not be conflated. Anomaly detection is merely the "beginner’s start." Classification, which requires labeled data, is necessary to map symptoms to specific problems. He advocates for decoupling the AI that detects the anomaly from the "look-up" AI (RAG+LLM) that searches manuals and FAQs.
Natarajan warns against using LLMs as the primary diagnostic engine for physical failure. "Treating an LLM as the detector of physical failure modes skips the physics and creates real risk," he explains. Instead, he proposes a sequence:
- Physics-informed detection: The first layer of classification.
- Human expert verification: An essential gatekeeper to translate asset reality into frontline action.
- RAG+LLM layer: An explanation and search tool that sits on top to provide the "why" and "how" to the maintenance team.
Economic Realities: When to Predict and When to Replace
A practical concern for every plant manager is the cost-benefit analysis of predictive maintenance. If a part is inexpensive, is it worth the investment in complex AI monitoring?
Scharinger acknowledges that there is no universal "rule of thumb." Instead, he recommends building a quantitative decision model for every use case. "Consider the cost of unplanned downtime, the residual useful life of the part, and the labor costs of replacement," he says. He suggests using AI-driven simulations, such as Monte Carlo analysis, to factor in uncertainty.
The Psychology of Outages
Interestingly, Scharinger points out that human psychology plays a massive role in maintenance investment. In mid-tier risk environments, teams often normalize outages, and there is a sense of pride in "crisis management." In high-risk environments, firms have built in so much redundancy that they often struggle to calculate the cost-saving potential of replacing that physical redundancy with "perfect" predictive maintenance.
Ensuring Trust: Data Quality and Security
The final frontier of Industrial AI is the issue of trust and data sovereignty. Attendees questioned whether AI data is held to the same quality rigor as physical products.
The MLOps Standard
Scharinger is clear: "If your AI is making decisions on data that hasn’t passed the same quality rigor you’d apply to a physical product, it is not trustworthy—it is just fast." He argues that industrial-grade AI requires the "cumbersome" but necessary world of MLOps, including strict version control for training sets, test sets, and model integration.
For those using tools like Copilot or Claude for internal knowledge management, the risk of data leakage is paramount. Scharinger’s advice is categorical: "You must operate within your own corporate environment. Negotiate contracts that prevent providers from using your data for model training, or look into hosting your own open-weight models."
The Feedback Loop of Trust
Natarajan concludes that trust is not a precondition but an outcome. It is built through a closed-loop system where:
- Every AI-generated prescription is verified by a human expert before implementation.
- Every real-world outcome is tracked.
- That outcome data is fed back into the system to sharpen the intelligence for future predictions.
Implications for the Future of Manufacturing
The takeaway from the Industrial AI Summit 2026 is clear: the transition to prescriptive maintenance is not merely a technological upgrade; it is a structural shift in how industrial organizations value their data and their experts.
Key Implications:
- Physics is the Foundation: Relying on domain-specific physics rather than waiting for historical datasets allows for faster, more accurate deployments.
- Human-in-the-loop is Non-negotiable: For high-stakes industrial environments, the "black box" is unacceptable. AI must serve as a decision-support tool that validates its prescriptions against human expertise.
- Data Quality as Product Quality: Organizations that treat their data pipelines with the same rigorous engineering standards as their physical assembly lines will be the ones that achieve true, long-term competitive advantage.
As the industry moves toward 2030, the divide will not be between those who have the most data and those who don’t, but between those who can successfully integrate their physical asset reality with the prescriptive power of AI, all while maintaining the security and auditability required for high-performance operations.
