Why Construction AI Keeps Missing the Mark (And How Domain-Specific Tools Fix It)
From pizza glue to RAMS reviews: Building AI that actually understands construction
The Translation Problem
I distinctly remember my confusion and disgust the first time I encoutered the term "gobbo", to my ears it just sounded wrong. I was a couple of years out of university and recently started as a site engineer for Henry Boot Construction working on the Mansfield Police Headquarters project. The bricklaying foreman, with a strong Barnsley accent, stormed over to me, and demanded to know where his "gobbo" was. I had already encoutered the "let's wind up the new lad" terms, such as "sky hooks" and "a long stand" and assumed that was a Barnsley equivalent. After my Senior Engineer stepped in as peacemaker and translator, I eventually learned that "gobbo" is a common Barnsley term for a mortar mix.
"Have you buttered the bricks?"
Despite all my education and training, I had no idea what the bricklaying foreman was talking about. In construction, we have our own language - and that's exactly why generic AI tools keep getting it wrong.
The Generic AI Trap
When Google's AI Told People to Put Glue on Pizza
When Google's AI Overview feature told users to add Elmer's glue to pizza sauce to keep cheese from sliding off, it wasn't just embarrassing - it was dangerous. The AI had used a decade-old Reddit joke and presented it as fact.
The Comedy of Construction Terms
Construction language is a minefield for outsiders - and for AI. Consider these genuine site phrases that would send HR into meltdown:
The Gland Incident: 'Have you tightened your glands?' We were talking about cable glands - metal fittings that secure cables to panels.
The Stuffing Box Situation: 'Lube up the stuffing gland and push the flex through' is a perfectly normal instruction for installing flexible cables.
When AI Meets Site Slang
This isn't just amusing - it's a serious problem. I've seen:
- Safety documents blocked because they mention 'dead loads' (permanent structural weight)
- RAMS flagged for discussing 'striking formwork' (removing temporary supports)
- Technical specifications flagged for inappropriate content for mentioning 'male and female couplers' (standard pipe connectors)
Now Imagine That Same AI Reading Your Method Statement
In construction, the stakes are higher. When AI doesn't understand that 'striking' means removing formwork, or that a 'wacker' is referrring to a compacting plate, not a violent act, we're not just getting bad advice - we're risking safety, compliance, and increased project costs.
Why Prompt Engineering Isn't Enough
The all leads into how we communicate with AI. There is a lot of talk in the AI world about prompt engineering and context engineering. But what does that actually mean?
The New Starter Analogy
Imagine sending a brilliant graduate to site for the first time. You can give them perfect instructions (prompt engineering), but without understanding how a particular project has been designed and how it interfaces with temporary works and other trades, they're still going to struggle with complex tasks.
- Prompt engineering = Asking the question clearly: “Could you summarise how to erect a steel frame?”
- Context engineering = Giving the new starter the right document first: “Here’s the steel frame erection sequence for the project.”
Even the smartest new starter (or LLM) can’t give a perfect answer if they don’t have the right context, and even with the right context, an ambiguous question leads to messy results.
Context Engineering: Giving AI the Right Textbook
"This is where context engineering comes in. Instead of just asking better questions, we're giving AI the construction method statements first. But even that's not enough..."
The Missing Piece: Domain-Specific Tools
From Chatbots to Workflows
Chatbots are brilliant for exploration. Ask ChatGPT about construction law or materials science, and you'll get something useful. But construction isn't about unique conversations - it's about consistent, repeatable processes that need to be right every time.
Why This Matters More Than You Think
This is exactly why we can't just throw generic AI at construction problems. We need AI that knows a 'kicker' is a concrete upstand, not a footballer. That understands 'making good' means repair work, not moral improvement. That recognises 'muck' as mortar, not dirt.
Building Intelligence Into Workflows
We've built tools that understand construction.
The RAMS Revolution
Our RAMS processor doesn't just extract text - it scores documents against Method Statement Assessment standards across 11 criteria. It knows the difference between a generic risk and a critical safety failure. It understands construction.
The Drawing Comparison Tool
Our drawing comparison tool doesn't just extract text - it compares drawing revisions against each other. It knows the difference between a generic revision reference change and change that could impact the construction process. It understands construction.
Knowing When Humans Must Decide
The Human-in-the-Loop Imperative
Construction AI also needs guardrails - clear boundaries where human expertise takes over. When our RAMS tool spots a critical safety issue, it doesn't try to fix it. It escalates it. That's not a limitation - it's intelligence.
Examples of essential guardrails:
- Commercial decisions over threshold amounts
- Safety-critical actions
- Contractual interpretations
- Design coordination conflicts
The Compound Effect
From Tools to Transformation
Once you have domain-specific tools that understand construction, something magical happens. You can chain them together. Drawing analysis feeds into quantity takeoffs. RAMS reviews trigger permit applications. Each tool does one thing brilliantly, and together they transform how we work.
Every tool we build at MyDrawIt is designed to be used in a workflow. We don't just build tools, we build workflows.
Trust Before Hype
I'm not chasing AI hype. I'm transforming organisations that want to change and building trust through tools that deliver measurable wins - time saved on RAMS reviews, errors caught in drawing revisions, compliance improved automatically. Once teams trust these specific tools, we can build the workflows that finally break our miniscule 50-year productivity gains.
Generic AI may never understand that 'making good' isn't about morality. That's why construction needs its own AI tools - built by people worked on site, reviewed and interpreted the drawings, and know the difference between striking and a strike.
If any of this resonates—if you’ve been the person who completely misunderstood a term, or have encountered AI not doing what you expected—I’m building for you. Try MyDrawIt. You’ll get 10 free credits to start. Use the simple drag and drop interface or the APIs if you want to go deeper. And follow along as I keep sharing what I’m learning about practical AI in construction.
Ian Yeo is the former CEO of Operance and founder of MyDrawIt. He specialises in helping construction organisations navigate digital transformation through practical AI implementation at Yeo Innovation. Learn more at ianyeo.com or try MyDrawIt at mydrawit.app.