Prescriptive AI on the Plant Floor: What It Really Takes, per Siemens and Infinite Uptime

On October 8, 2026, IIoT World published a Q&A in which Boris Scharinger (Siemens, author of "Industrial AI: From Pilot to Profit") and Karthikeyan Natarajan (CEO of Infinite Uptime) answer eight questions raised during the "Unlocking Plant Reliability with Verticalized Prescriptive AI" session at the Industrial AI Summit 2026. Prescriptive AI goes beyond predicting a failure: it indicates which component to replace, which maintenance action to perform and when. This analysis is based on the first six answers; the last two were not legible in the text we reviewed.

A note of caution: the content is marked "Sponsored by Infinite Uptime | Editorially Independent". The performance figures attributed to Natarajan are therefore vendor claims and have not been independently verified.

Data first or use case first?

Scharinger describes the debate between "use case first" and "data strategy first" as almost ideological: CFOs tend to prefer the return from a single use case, CTOs the data strategy. In his view the data layer is a foundational capability, and its cost cannot be charged to a single use case.

Natarajan, by contrast, argues that waiting for clean data is a "multi-year tax", because most plants lack a complete history. He claims a physics-aware approach could deliver a first prescription in two weeks and full-plant coverage in 90 days (vendor claim). His broader advice is to start with the most critical, best-instrumented assets.

Retrofitting existing plants

Both support adding sensors and computer vision to older plants. Scharinger notes that this can also reduce dependence on machine OEMs. According to Natarajan, vibration catches mechanical degradation, while vision and thermal sensors detect process and quality problems that vibration cannot see.

Diagnosis: keeping the human in the loop

For Scharinger, anomaly detection is only a starting point, and anomalies should be reviewed by people. Classification requires labels, which AI can help produce; the labels are then linked to documentation through RAG and LLMs. He suggests keeping anomaly and classification AI separate from the AI used for retrieval.

Natarajan proposes a similar sequence:

• physics-based failure signatures for initial classification;
• verification of every prescription by a reliability expert before it reaches the shop floor;
• RAG+LLM used only as an explanation and search layer.

Using an LLM to identify physical failure modes is a real risk (Natarajan).

Replace or predict? A decision model

There is no empirical threshold based on component value. Scharinger recommends a quantitative model for each use case, covering the cost of unplanned downtime, the remaining useful life lost through early replacement (for example an expensive CNC tool), and the cost of planned replacement, including scheduled downtime and labor. He suggests using AI to draft the model and adding a Monte Carlo simulation to handle uncertainty.

On auditing, Siemens used about 2-3 people for 6 weeks to review algorithms, business assumptions and risk assessment, including the Monte Carlo simulations, without charging the cost to the business units involved.

Key takeaways

Practical guidance for technical and IT leaders:

• treat data as shared infrastructure, but start with critical, well-instrumented assets;
• evaluate retrofits with different sensor types, choosing the technology by failure type;
• separate failure classification from document retrieval, with mandatory human review;
• build an economic model for each case with explicit uncertainty, and plan for audit work.

The value of the Q&A lies mainly in its framing: prescriptive maintenance is as much a problem of decision-making processes and governance as of algorithms. Claims about time to launch, however, should be verified with a pilot on your own assets.