Prescriptive AI tells plant operators which component to replace, which maintenance action to take, and when. Boris Scharinger, author of Industrial AI: From Pilot to Profit and R&D Excellence, DataOps and Master Data Management at Siemens, and Karthikeyan Natarajan of Infinite Uptime answer eight audience questions about deploying prescriptive AI in manufacturing from the “Unlocking Plant Reliability with Verticalized Prescriptive AI” session at the Industrial AI Summit 2026.
Is Industrial AI Really the Answer When Most Manufacturers and Energy Companies Still Can’t Get Their Data Right?
Many manufacturers and energy companies still struggle to get data available and right. The original audience question challenges whether AI deployment should be staged by maturity level, starting with data management: “Based on this discussion I am wondering if Industrial AI is really the solution of today. A lot of manufacturers and energy companies have problems getting the data available and right. Shouldn’t we discuss how to integrate AI based on the level of the maturity? Isn’t it more a question of Industrial Intelligence, starting with Data Management?”
Boris Scharinger, Book Author, R&D Excellence, DataOps and Master Data Management at Siemens:
The question of what comes first, the use case that pulls efforts to get the data challenges sorted or the data strategy put in place first, is almost an ideological one. CFOs tend to look at use cases as the lever to pull ROI, CTOs look at a data strategy (and related investments) as something that establishes a strategic capability. CIOs ask why we are not doing everything with SAP as the data issue is solved implicitly (haha). My book elaborates on RoI calculation including the topic of a data layer as a foundational capability that cannot be attributed to the Use Case RoI layer.
Karthikeyan Natarajan, CEO of Infinite Uptime:
Agree with the premise, but waiting for a clean data foundation before deploying AI is a multi-year tax before any outcome shows up, most plants will never have complete historical data given decades-old equipment and sparse instrumentation. This is exactly why physics-aware beats data-dependent: sensor coverage from day one, first prescription inside two weeks, plant-wide in 90 days, because the intelligence starts from known asset physics and vertical fault patterns, not from a customer’s historical dataset. The system generates clean, structured, validated-outcome data going forward rather than requiring it upfront. Data maturity and the intelligence layer build each other in parallel, not one gating the other. Where the questioner is right: prioritize by production consequence first, start with the highest-consequence, best-instrumented assets.
How Do You Get Old Plants to the Industrial AI Level?
One attendee described transforming a track manufacturing plant from Industry 3.0 to 4.0/5.0 using cameras for vision and temperature, monitoring both machine performance and material quality on milling machines. The original audience question: “We are transforming our Track Manufacturing Plant from Industry 3.0 to 4.0/5.0 using cameras (vision and temperature) and AI to get additional data and information for the process. For example, for the milling machine we look at the performance of the machine and the quality of the material. What do you think, is this the way to get old plants to the Industrial AI level?”
Boris Scharinger, Book Author, R&D Excellence, DataOps and Master Data Management at Siemens:
Sensor retrofit can be an excellent way to move a brownfield environment to Industry 4.0 and intelligent ways to leverage machine vision for it, sounds great! It can also contribute to strengthening your machine vendor independence if this is of concern.
Karthikeyan Natarajan, CEO of Infinite Uptime:
Yes, and it’s a clean example of connecting asset reality to production consequence through more than one sensing modality. Vision/thermal and vibration-based sensing catch different failure physics: vibration catches mechanical degradation, vision and thermal catch process-induced and quality-correlated issues vibration can’t see. Combining both covers more of the asset-reality picture than either alone.
Which Diagnosis Approach Works for Prescriptive Maintenance?
The original audience question: “For predictive maintenance diagnosis, this is more challenging than anomaly detection and I have seen a variety of approaches. Some companies use humans to review the data to make a judgement. I have seen some that wait until there is labelled training data to train supervised classification type models. I have seen some consider the RAG+LLM approach that uses reference documents. Which approach have you used, and which approach has been the most successful for prescriptive maintenance?”
Boris Scharinger, Book Author, R&D Excellence, DataOps and Master Data Management at Siemens:
First of all: Anomaly detection is the beginner’s start. Humans will have to review the anomaly to make sense of it and define measures. Classification requires labeling (labeling can be AI supported and accelerated though) and provides a mapping of anomaly patterns (=symptoms) to diagnosis (actual problems), so speeding up the way to “therapy”. Error types (=classification / labels) are instrumental to hop into documentation (troubleshooting, manuals, FAQs, “known issues”) so here’s the link between labels and leveraging RAG+LLMs for “documentation integration”. At the current state I would decouple the anomaly/classification AI from the “look-up AI”. In any case, linking diagnosis to automated therapy suggestions, accelerates problem resolution further and increases efficiency. Furthermore: Solution providers who have built and deployed that type of logic will typically help you achieve the desired outcome faster.
Karthikeyan Natarajan, CEO of Infinite Uptime:
Not one of the three alone, a sequence, and it starts with physics, not data. Physics-informed fault signatures, built from domain expertise across mechanical and process-induced fault modes, tuned by vertical, do the initial classification, so accuracy doesn’t depend on a customer’s own failure history from day one. That’s the asset-reality layer. A human reliability expert then verifies every AI-generated prescription before it reaches the plant floor, connecting that asset reality to production consequence and translating it into frontline action, not just a flagged anomaly. RAG+LLM sits on top as an explanation/search layer, not the diagnostic engine; treating an LLM as the thing detecting physical failure modes skips the physics and is a real risk.
When Is Replacing a Part Cheaper Than Predicting Its Failure?
The original audience question: “Boris commented on not spending too much money on predictive maintenance if it’s cheaper to replace the part. What would you recommend a rough threshold for part value, or which metrics and parameters should you consider to come to a decision?”
Boris Scharinger, Book Author, R&D Excellence, DataOps and Master Data Management at Siemens:
Unfortunately, I cannot provide a rule of thumb here but simply the recommendation to build a quantitative decision model for the individual use case. Beside the factor of the cost of an unplanned downtime, there is the factor of value lost by cutting the remaining useful lifetime of a part (e.g., an expensive CNC cutting tool) short. There is the cost of replacement (cost of planned downtime plus labor costs of actual replacement work) that must be factored in in many scenarios. My recommendation: Use AI to create a mock-up of a decision model, maybe even factoring uncertainty in by e.g. a bit of Monte-Carlo-Simulation!
How Much Does It Cost to Audit Predictive Maintenance Algorithms?
The original audience question: “Was the cost of the internal audit on the algorithms calculated? What did the internal audit on the algorithms cost?”
Boris Scharinger, Book Author, R&D Excellence, DataOps and Master Data Management at Siemens:
Whatever an audit costs. Siemens chose not to allocate any audit costs to the business units “affected”. However, auditing algorithms, the underlying business assumptions (and their likelihood), defining a risk assessment approach, running some monte-carlos, is a 2-3 people x 6 weeks effort.
Why Don’t Manufacturers Invest More in Predictive Maintenance?
The original audience question: “Boris, why a company only pay 1000 euros if a startup will deliver a 0% downtime? Why companies don’t value this more?”
Boris Scharinger, Book Author, R&D Excellence, DataOps and Master Data Management at Siemens:
My explanation for this is: In low- to mid-tier risk areas, people and management got used to outages. A high-performance-team of “marine-type” men can also prove their value by solving a crisis. So I guess there is some psychology involved. In high-risk-areas, everything has been built in a redundant setup. And it is yet not foreseeable (nor a transparent tough-numbers calculation) what costly elements of redundancy can be eliminated once predictive maintenance becomes “perfect”.
How Do You Keep Proprietary Data Safe in Tools Like Copilot and Claude?
The original audience question: “I’m using Copilot + Claude for Lessons Learned searches in my company. We have 20 years of data, and the agent is providing great and complex answers in less than 1 minute. When it comes to data safety, how can I make sure that this is properly covered?”
Boris Scharinger, Book Author, R&D Excellence, DataOps and Master Data Management at Siemens:
You must move into your own corporate environment of an LLM and negotiate a contract that does not allow the LLM provider to leverage your data for training. Examples for these types of environments: MS Azure OpenAI Services. Alternatively, you work on hosting your own open weight / open source model with the help of a hosting provider.
Would Your AI Data Pass the Same Quality Rigor as a Physical Product?
The original audience question: “Most Industrial AI readiness conversations focus on data infrastructure and governance but how many organizations are actually building AI trust on a foundation of validated, audit-ready quality data? If your AI is making decisions on data that hasn’t passed the same quality rigor you’d apply to a physical product, is that AI actually trustworthy or just fast?”
The question implies that few organizations hold their AI data to the same quality standard as their physical products.
Boris Scharinger, Book Author, R&D Excellence, DataOps and Master Data Management at Siemens:
Correct. This includes tight version control of everything: training data sets, test data sets, models, and model integration (aka “Harness”). This could imply even LLM specific testing. Welcome to the cumbersome world of industrial MLOps. Once you master this, you have built a critical differentiation, up to the capability to use AI even in regulated environments.
Karthikeyan Natarajan, CEO of Infinite Uptime:
Fair challenge, very few organizations have audit-ready data quality when they deploy industrial AI, and pretending otherwise is what fuels AI skepticism. Trust doesn’t come from flawless input data upfront; it comes from a closed loop that connects every step through to a validated outcome: physics-aware detection, human expert verification before anything reaches the plant floor, real-world outcome tracking on every prescription, and that outcome data feeding back to sharpen the system.
These questions were submitted by attendees of the “Unlocking Plant Reliability with Verticalized Prescriptive AI” session at the Industrial AI Summit 2026. Answers provided by Boris Scharinger of Siemens and Karthikeyan Natarajan of Infinite Uptime. The session also featured Jan Pingel of Ingersoll Rand, and the session was moderated by Rob Tiffany of IDC. The panel was sponsored by Infinite Uptime.
Related from IIoT World
