As an engineer, it is easy to get caught up in the wave of new AI capabilities. But once the initial excitement around having an AI agent answer your e-mails, and plan your week, most engineers have to get back to their day job in the project office where the real engineering work takes place. And most engineers dont just work with normal word and excel documents, they have to work with drawings, design software, CAD, simulations and more.
Many industrial organisations are now asking a sensible question: can artificial intelligence help us get more value from the thousands of drawings, diagrams, specifications and reports sitting in our engineering systems? The answer is yes, but with an important qualification. Standard AI tools are not, on their own, enough to reliably interpret complex engineering documents such as piping and instrumentation diagrams, or P&IDs.
A P&ID is not just a picture. It is a dense engineering representation of equipment, piping, valves, instruments, control loops, line numbers, tag numbers and relationships. Much of its value lies not in the symbols themselves, but in how those symbols connect to each other.
Too many AI discussions among engineers are far too optimistic about the capabilities of the current tools to accurately interpret their drawings. Over time the tools may improve, but being a specialised area, this is unlikely to be the focus of the AI vendors right now.
From Static Drawings to Structured Data
A large language model, or LLM, is an AI system trained to work with language. Tools such as ChatGPT are examples of LLM-based systems. They are good at summarising text, answering questions, drafting explanations and identifying patterns in written information.
But a P&ID is not ordinary text.
Even if a drawing is available as a PDF or image, the important engineering meaning is often embedded in geometry, symbols, line connections, labels and conventions. A generic AI tool may recognise some text or visual features, but that does not mean it has correctly understood the process structure.
For AI to be useful in this context, the document first needs to be converted into structured data. In plain terms, this means representing the drawing as objects and relationships that software can read. For example:
- Pump P-101 is connected to line L-204.
- Line L-204 includes valve XV-210.
- Instrument FT-301 measures flow on that line.
- Controller FIC-301 receives the signal and acts on a final control element.
This kind of information can be represented in formats such as JSON, where entities, properties and relationships are explicitly described. Once the data is in that form, AI systems can reason over it far more reliably.
Why Extraction Is the Real Bottleneck
The difficult step is extraction.
It is one thing to ask an AI assistant, “What instruments are on this drawing?” It is another thing to prove that the answer is complete, accurate and traceable back to the source document.
Engineering offices know this problem well. A small error in a drawing interpretation can become a procurement error, a commissioning delay, a control system mismatch or a safety concern. In industrial environments, “mostly right” is rarely good enough.
This is why purpose-built platforms are emerging around engineering document intelligence. Their value is that they add the specialised layer needed to identify engineering entities, extract properties, understand connections and preserve the underlying relationships.
Replicating that capability inside an average engineering office is not trivial. It may require computer vision, symbol libraries, drawing standards, domain rules, validation workflows, data models, integration tools and experienced engineering oversight. Some organisations may have the resources and appetite to build this. Many will not.
The Single Source of Engineering Truth
The phrase “single source of truth” is often overused, but in this case it matters.
If engineering data is scattered across PDFs, CAD files, spreadsheets, control system databases and maintenance systems, each team may be working from a slightly different version of reality. AI does not automatically fix that. In fact, AI can make the problem worse if it confidently surfaces information from weak or inconsistent sources.
A more useful goal is to build a trusted engineering data layer. That means the organisation has a reliable way to move from documents to structured information, and from structured information to operational decisions.
Standards and initiatives such as DEXPI, which focuses on data exchange for process industry information including P&IDs, point in this direction. The broader lesson is that industrial AI depends heavily on data quality, interoperability and governance — not just model capability.
Practical Questions for Technical Leaders
Before investing in AI-driven engineering document tools, technical leaders should ask a few practical questions:
- What engineering documents are most valuable to digitise first?
- Are we trying to extract text, symbols, relationships or full engineering meaning?
- How will extracted data be validated by competent engineers?
- Can the output integrate with our DCS, engineering tools, asset systems or data platforms?
- What level of accuracy is required before the information can be trusted?
- Are we buying a mature capability, building one, or experimenting to understand the gap?
The key is to avoid treating AI as a shortcut around engineering discipline. The better approach is to use AI to strengthen that discipline: reduce manual bottlenecks, improve consistency and make engineering knowledge easier to access.
Conclusion
AI will change how engineering information is used, but the foundation is still engineering quality.
For P&IDs and similar documents, the real opportunity is not simply to “chat with drawings”. It is to convert static documents into structured, validated, machine-readable engineering intelligence. Once that layer exists, AI becomes far more useful — not as a magic interpreter, but as a practical assistant working on trusted data.
Industrial organisations should approach this area with interest, but also with healthy scepticism. The winners will not be those who adopt AI vocabulary fastest. They will be those who understand their engineering information well enough to make AI useful.
Call to Action
If your organisation is exploring AI for engineering documents, start by examining the quality and structure of the underlying data. Follow my blog for more practical reflections on AI, automation and industrial technology strategy.
References used in compiling this article
- DEXPI Initiative, “Data Exchange in the Process Industry,” DEXPI, https://dexpi.org/
- National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework (AI RMF 1.0),” NIST, January 2023, https://www.nist.gov/itl/ai-risk-management-framework
