Why ChatGPT Can’t Read Your P&IDs?
A Perspective from 35 Years of Engineering Logic
From embedded software to enterprise systems, the golden rule of engineering has always been the same: Context is King, and Ambiguity is the Enemy. I’ve built process control systems and diagnostic imaging platforms. Here is why “Generic AI” fails in high stakes engineering environments.
In my 35+ years of industry experience, I have worn many hats. I spent a decade in Process Control and Automation, where a single missed signal meant a plant shutdown. I spent another 20 years in Digital Health and Medical Technologies, building large diagnostic imaging systems where a data artifact could mean a misdiagnosis.
Today, as the founder of an Industrial AI startup, I see a dangerous trend. The EPC sector is rushing to adopt “Generative AI” tools that are brilliant at language but terrible at logic.
You can upload a specification to a standard Large Language Model (LLM), and it will write a beautiful summary. But try uploading a P&ID and asking a fundamental engineering question:
“What happens if we isolate Valve V-101?”
Standard AI fails. It sees the text “V-101,” but it doesn’t understand the physics. It doesn’t know that closing V-101 stops flow to the heat exchanger downstream.
The “Geometry Gap”: Why Text Models Fail Engineering Having built diagnostic imaging systems, I know that interpreting an image requires understanding spatial relationships, not just pixel values. Similarly, in Process Control, a tag isn’t just a label; it’s a node in a logic loop.
Standard Enterprise AI (like Copilot or Glean) treats your P&IDs like “flat text” or “pictures.” It is blind to the geometry and the connectivity. It cannot see that a line on a drawing represents a pipe carrying superheated steam, nor can it trace the topology from a valve to a pump.
The Cost of “Dumb” Search This results in what I call the “Search Tax”. Your lead engineers—people whose time is incredibly expensive—are wasting 30% of their day verifying data. They search for “Heat Exchanger,” but they have to manually open five different PDFs to check the tag numbers because the search tool doesn’t understand that “Tag 20-E-101” is the heat exchanger.
My Approach: Operationalizing Engineering Memory at Evomaton, we didn’t build a chatbot. We built an Engineering Memory System rooted in the rigor of my background in critical systems.
We (Evomaton) developed ProjectIQ, a solution driven by Structural RAG. We don’t just parse text; we parse the logic of the document. From “Search” to “Certainty”, I believe AI in engineering shouldn’t just be “conversational.” It must be deterministic.
The future of EPC isn’t about writing emails faster. It’s about preserving the logic of your best engineers and making it queryable. It’s about asking, “Find the isolation valve upstream of Pump A,” and getting an answer based on flow paths, not keywords.
We are Operationalizing Engineering Memory—because in our business, context isn’t just nice to have. It’s structural.