AI for MEP Engineers: A Practical Guide to Using GPT

8/12/2025

Here’s the thing about artificial intelligence: the hype is both completely justified & entirely overblown, all at the same time. We see these demos of AI creating entire videos from a single prompt or writing flawless code, & it’s easy to think it’s a magic wand that will solve every problem. And then you try to use it for a very specific, technical task, & it falls flat on its face.

This is EXACTLY the situation many of us in the MEP (Mechanical, Electrical, & Plumbing) engineering world are facing. We hear about the coming wave of GPT-5 & the promise of AI-driven design, & we get excited. But when we kick the tires on the current models, we run into some serious, almost comical, roadblocks. A classic example? Something that seems simple on the surface: equipment identification.

You’d think you could just feed an AI a project manual or a schedule & ask it to list the equipment. But what you often get back is a mess. It might list every single VAV box terminal as a major piece of equipment, confuse a unit designation like "AHU-1" for a manufacturer's model number, or miss the main chiller entirely. It confidently gives you information that is just… wrong.

The problem isn't that the AI is "dumb." The problem is that it's a brilliant, well-read, incredibly articulate intern who has never set foot on a construction site or opened a set of drawings. It has read the entire internet, but it hasn't read your project's spec book, your manufacturer's submittals, or your company's standard details.

So, how do we bridge that gap? How do we use these powerful new tools in a field that demands precision, accuracy, & deep domain knowledge? Turns out, it's not about waiting for a perfect AI. It's about learning how to work with an imperfect one. It’s about becoming a better collaborator, a better data wrangler, & a better prompt artist. Let's dive into what's really going on & how to make this tech work for us, right now.

The Real Issue: AI's "Context Blindness" in MEP

The core of the issue isn't really about identifying a specific pump or fan. It's about context. An LLM's greatest strength—its massive, general knowledge—is also its biggest weakness in a specialized field like MEP engineering. It’s trained on petabytes of text & data from the public internet, which is great for writing an email or a poem. But your project’s One-Line Diagram, mechanical schedules, & Basis of Design narrative? That’s not on the public internet.

This leads to a few critical failures.

1. It Doesn't Speak Our Language

MEP engineering, like any specialized field, has its own shorthand. We use cryptic abbreviations & acronyms that are second nature to us but completely alien to a generalist AI. One study looking at enterprise data found that LLMs that were masters at understanding public data struggled immensely when faced with internal, coded language like "KUNNR" for "customer number."

We have the exact same problem. We see "CHW" & "HW" & instantly know it's chilled water & hot water. We see "MCC" & know it's a motor control center. An AI, without specific training, might misinterpret these or just ignore them. It sees a schedule filled with designations like

EF-G-01

VAV-3-12

, &

P-HW-02

& it lacks the fundamental context to know what those mean. It doesn't understand the hierarchy—that the

P-HW-02

is a part of the larger hot water system, which is a key part of the building's mechanical design.

2. The "Hallucination" Problem is a Liability

One of the most well-known limitations of LLMs is their tendency to "hallucinate"—a polite way of saying they make stuff up. Because their fundamental job is to predict the next logical word, they can sometimes generate very plausible-sounding but factually incorrect information. In creative writing, this is a feature. In engineering, it's a catastrophic failure waiting to happen.

You can't have an AI "hallucinating" a fire pump's flow rate or the voltage of a switchboard. An engineer's entire career is built on a foundation of trust & accuracy. A real-world example from Pelles.ai, a company building AI for MEP, tested GPT by asking it to list major equipment from a bid package. The AI confidently provided a list, but it completely missed the main Makeup Air Unit (MAU) & Packaged Condensing Unit (PCU)—arguably two of the most significant pieces of equipment. It saw the letters but didn't grasp their importance. That's not just an error; it's a fundamental misunderstanding of the task.

3. Knowledge Cutoffs & The Data Silo

LLMs have a knowledge cutoff date. A model trained up to April 2023 knows nothing about events, products, or codes released after that date. This is a major issue in an industry where codes are constantly updated & new products are released all the time.

But the bigger issue is the data silo. Your firm's knowledge isn't stored in a public database. It's locked away in PDF spec books, proprietary manufacturer software, Revit & AutoCAD files, and spreadsheets. The AI simply can’t access it. It can't tell you the efficiency of the specific Daikin unit you specified because it doesn't have the submittal data. It can't tell you how many diffusers are on the third floor because it can't count them from a drawing that it can't properly interpret. The same Pelles.ai test found that GPT got diffuser types right, but the quantities completely wrong because that required counting them on a drawing—a multimodal task that was beyond its reach.

So we're left with a tool that's brilliant on paper but lacks the specific, contextual, & up-to-date knowledge required for professional-grade work. The answer isn't to throw the tool away. The answer is to get smarter about how we use it.

Making It Work: Practical Strategies for MEP Professionals

The path forward isn't waiting for some mythical, all-knowing GPT-5. It's about adopting new workflows that leverage the AI for what it's good at while mitigating its weaknesses. It's about treating the AI less like an autonomous engineer & more like a super-powered apprentice.

Strategy 1: The "Copilot," Not the "Pilot"

The first & most important mental shift is to reframe the AI's role. It's not the Professional Engineer of Record. It's a copilot. It's an assistant that can accelerate your workflow, but you are ALWAYS in command, making the final checks & decisions.

What does this look like in practice?

Drafting, Not Finalizing: Use it to generate a first draft of an RFI response, a scope of work narrative, or an email to a client. It's incredibly good at this kind of wordsmithing. But you MUST review & edit it for technical accuracy.
Code "Search," Not Interpretation: You can ask it, "What does the NEC say about receptacle spacing in a commercial kitchen?" & it will likely give you a pretty good, sourced answer to start from. But it is NOT a substitute for you, the engineer, reading & interpreting the code yourself in the context of your specific project. It can point you to the right chapter, but you need to do the reading.
Brainstorming & Idea Generation: Stuck on a design problem? You can describe the constraints to an AI & ask it to suggest three potential HVAC system types for a small office building. It might suggest a system you hadn't considered. This is a great way to break through creative blocks, but the actual design, calculation, & selection is still on you.

Strategy 2: Grounding AI in Reality with Your Data (RAG)

This is probably the single most important strategy for making AI useful in a technical field. The problem is the AI doesn't have your data. The solution? Give it your data.

This is done through a technique called Retrieval-Augmented Generation (RAG). In simple terms, RAG connects a general-purpose LLM to a private, curated knowledge base. When you ask a question, the system first retrieves relevant information from your private data & then feeds that information to the LLM as context to generate its answer. The AI isn't just relying on its hazy memory of the public internet; it's using your specific project documents as its source of truth.

This changes EVERYTHING.

Suddenly, you can overcome the knowledge limitations. You can build a system that contains:

All your project-specific documents: Spec books, drawings, schedules, meeting minutes.
Manufacturer data: Cut sheets, installation manuals, & performance curves for the equipment you use most.
Internal Standards: Your company's standard details, calculation methodologies, & checklists.
Building Codes & Standards: The specific, up-to-date versions of ASHRAE, NEC, & local codes that apply to your projects.

This is where the industry is heading, & it’s becoming more accessible every day. For example, businesses are already using platforms like Arsturn to tackle this exact problem. Arsturn helps businesses create no-code AI chatbots that are trained on their own specific data. An MEP firm could use this to create an internal "Project Assistant" chatbot. Imagine being able to ask:

"What is the specified VFD model for pump P-HW-02 on the Downtown Tower project?"
"Pull up the cooling capacity & EER for the Trane rooftop unit specified in the basis of design."
"Summarize all RFIs related to the electrical service entrance for the last 3 weeks."

The chatbot, grounded in your private data, could provide instant, accurate answers with sources. It's not doing the engineering, but it's DRASTICALLY reducing the time spent searching for information, allowing engineers to focus on, well, engineering. This approach turns the AI from a lying intern into a hyper-efficient librarian who knows your projects inside & out.

Strategy 3: The Fine Art of the Technical Prompt

Interacting with an AI is a skill, & just like any skill, you get better with practice. This is often called "prompt engineering." Vague questions get vague answers. Detailed, specific questions get detailed, specific answers.

Let's look at an example for our equipment ID problem:

Bad Prompt: "List the equipment in this document."
- Why it's bad: It's lazy. It doesn't define "equipment." It doesn't specify the source. The AI is forced to guess your intent, & it will probably guess wrong.
Good Prompt: "Act as an MEP engineering assistant. I am providing you with a 15-page PDF of a mechanical equipment schedule for the 'Metropolis Health Clinic' project. Your task is to parse ONLY this document & create a table with the following columns: 'Equipment Tag', 'Equipment Type', 'Manufacturer', 'Model Number', 'Specified Airflow (CFM)', & 'Associated Drawing Sheet'. Please ignore all plumbing fixtures & focus only on air-handling units (AHUs), fans (EF, SF, RF), & variable air volume (VAV) boxes. If any information for a column is not present in the schedule for a given piece of equipment, leave the cell blank."
Why it's good:
- It sets a role: "Act as an MEP engineering assistant."
- It defines the source: "a 15-page PDF of a mechanical equipment schedule."
- It gives a clear, specific task: "create a table."
- It defines the output format: "with the following columns..."
- It provides constraints & negative constraints: "ignore all plumbing fixtures," "focus only on..."
- It tells the AI how to handle missing data: "leave the cell blank."

This level of detail removes the guesswork. You are guiding the AI step-by-step, dramatically increasing the odds of getting a useful, accurate output.

Strategy 4: Integration, Not Isolation

The true power of AI will be unlocked when it's not a standalone chatbot window but is deeply integrated into the tools we already use. We're starting to see this with the concept of "AI agents" – AIs that can perform multi-step tasks & interact with other software via APIs.

Think about the workflow of laying out receptacles in a room. An engineer might look at an architectural PDF in Bluebeam, mark it up with NEC requirements in mind, then a designer has to manually place those receptacles in a Revit model.

An integrated AI workflow might look different. An AI agent could be given the goal: "Lay out receptacles in this room according to NEC 406.5 & our firm's standard of one receptacle per 10 feet of wall space." The agent could then:

Access the architectural background via a Revit API.
Interpret the geometry of the room.
Place the receptacles according to the coded rules it was given.
Flag any ambiguities or conflicts for the engineer to review.

We're not quite there yet, but it's where things are headed. The most forward-thinking firms are already exploring how to connect different software platforms, using AI as the "glue." The key takeaway is to think of AI not as a replacement for Revit or Bluebeam, but as a powerful layer that can automate the tedious data transfer between them.

The Future: From Chatbots to True "Agentic AI"

Where does this all lead? The buzzword you'll start hearing more is "agentic AI." This is the evolution from reactive chatbots to proactive, goal-driven assistants. Instead of just answering a question, an AI agent can be given a complex goal & can then autonomously break that goal down into sub-tasks, execute them, & adapt as it goes.

For our industry, this could mean moving from asking "What's the airflow for AHU-1?" to telling it, "Generate a preliminary HVAC equipment schedule for Project Phoenix based on the attached architectural floor plans, our standard Basis of Design document, & local energy codes."

This is the holy grail, but it will require even more sophisticated data grounding & integration. It also mirrors a shift we'll see in all aspects of business. As companies build these complex internal systems, they'll also need to improve how they communicate with the outside world. This is another area where a platform like Arsturn becomes incredibly relevant. Arsturn helps businesses build conversational AI to forge meaningful connections with their audience. The same principles of training an AI on specific data to answer an engineer's questions can be applied to a company's website. An AI chatbot trained on a firm's portfolio, services, & case studies can engage potential clients 24/7, answer their initial questions, qualify them as leads, & schedule a consultation with a human engineer. This frees up the human team to focus on the high-value work of design & client relationships, which is exactly the goal of using AI internally.

Tying It All Together

Look, GPT-5, or whatever comes next, is going to be incredibly powerful. It will be faster, have a larger context window, & be better at reasoning. But it will NOT be a magical replacement for an experienced MEP engineer. It will still lack the deep, nuanced, project-specific context that is the hallmark of our profession.

The hype is real, but it's not about a robot taking your job. It's about a new class of tools that will fundamentally change your workflow. The engineers who will thrive in the coming decade are the ones who see AI not as a threat, but as a powerful, if flawed, collaborator. They will be the ones who master the art of the technical prompt, who champion the use of RAG to ground AI in their firm's knowledge, & who see the potential for AI to automate the 80% of their day that is drudgery so they can focus on the 20% that is true engineering.

It's a pretty exciting time to be in this field. We're on the cusp of a major shift, & by understanding both the incredible potential & the very real limitations of these tools, we can be the ones to lead the way.

Hope this was helpful. Let me know what you think.