Built For Tomorrow - The Construction Technology Leadership Podcast

Title options

1. Nigel Wea on Why MeltPlan Won't Ship AI Below 95% Accuracy
2. Workflow 10/10, AI 2/10: How Nigel Wea Ended Up at MeltPlan
3. Two Weeks of Document Review Nobody Gets Paid For

Show notes

Nigel Wea spent his early career as a civil and environmental engineer in water infrastructure, sizing pipes, lift stations and reservoirs and spending his first day on the job binding plan sets by hand in the print room for six hours. He built Codex Civil to pull together the standards, calculations and background research behind a preliminary design report. A comment on a LinkedIn post about MeltPlan's funding led to a call with its CEO, who rated his work: workflow 10 out of 10, UI/UX eight out of ten, understanding of AI two out of ten. Nigel is now Founding Product & Customer Success at MeltPlan, which builds AI for pre-construction: bid leveling, plan and spec review and code compliance, with a takeoff model in development.

His central argument is about honesty. Frontier models used straight out of the box get construction documents roughly 50 to 70% right, and too many buyers treat that output as finished. MeltPlan keeps a product out of the market until its benchmarks reach 95% or better, and puts its own estimators on the last 10% in its estimating service. For anyone leading technology adoption at a contractor, consultant or design firm, this is a practical way to judge AI tools by where they fail.

What you'll learn

• Why a frontier model with no guardrails or harnesses tends to land around 60% on construction work, and what that means for the 40% your team still has to check.
• How MeltPlan levels subcontractor bids apples to apples, flagging exclusions and lump sums against broken-out pricing, with the output in Excel.
• Why Nigel thinks document review is the easiest return on investment in construction: five or six people, two weeks, and none of it billable.
• How engineering fees, often around 10% of construction cost on a lump sum, get eaten when one municipal review mistake costs a six-week turnaround.
• Why seven- and fourteen-day pilots on a project the client already knows went wrong earn more trust than any demo.

Newsletter issue

Subject: The 60% problem with AI in construction

The most useful number from my conversation with Nigel Wea this week is 60%.

Nigel is Founding Product & Customer Success at MeltPlan, and before that he was a civil engineer in water infrastructure. He says that if you point a frontier model at construction documents straight out of the box, with no guardrails or harnesses, you get something like 50 to 70% accuracy. That isn't the problem on its own. The problem is that most of the GCs and engineers using it don't know the figure, so they skip the 40% of review time the output still needs. In his words:

"They're taking it as viable. They're taking it as this is accurate. And that's where a lot of the mistakes are happening."

He hears a version of the same complaint at least once a week: a firm has spent heavily on software and finds it takes longer to fix the output than to have done the work by hand. His view is that tools doing the low-hanging 60 or 70% are fine, as long as they say so. The harm comes from calling it done.

MeltPlan's answer is a release rule. Nothing goes into the marketplace until its benchmarks hit 95% or better, and the areas where the last 5% might fall short are flagged for the user. Their takeoff vision model is sitting in the 80s, so it stays internal for now. In their estimating service, the AI does about 90% of the work quickly and their lead estimator and his team verify the rest.

Nigel's reason is trust. One bad experience doesn't just lose a customer. It can put someone off AI for everything, and that sets back the whole industry.

We also got into where he thinks the easiest return sits: document review. On a typical project he has seen five or six people spend two weeks reviewing the same plans and specs, work nobody is paid for. A tool that does a reliable first pass in an hour would cut that QAQC time by about 90%.

And there's the story of how he joined MeltPlan. He commented on a LinkedIn post about their raise, the CEO called him, and the verdict on what he'd built was blunt: workflow 10 out of 10, UI/UX eight, understanding of AI two. Nigel said he appreciated the honesty, because that is how he learns.

If you're evaluating AI tools for pre-construction, ask every vendor one question: what's your accuracy, and where does it break?

The full conversation with Nigel is on Built For Tomorrow now.

LinkedIn posts

Post 1

Ask any AI vendor in construction one question: what percentage does it get right, and where does it fail?

Nigel Wea of MeltPlan was refreshingly specific about this on Built For Tomorrow. A frontier model used straight out of the box on construction documents, no guardrails or harnesses, sits at roughly 50 to 70%. A proprietary model built and harnessed for the job can get into the 90 to 95% range.

Neither number is the issue. The issue is buyers not knowing which one they have. If it's 60%, your team needs to spend 40% of the effort reviewing. Treat it as finished and that's where the mistakes come from.

MeltPlan's rule is that nothing goes to market until benchmarks reach 95% or better, with the weak spots flagged. Their takeoff model is in the 80s, so it waits.

Nigel says he has the same conversation at least once a week with a firm that has spent heavily on software and now finds fixing the output slower than doing the work by hand.

Transparency about accuracy is what protects trust in AI across the industry. Full episode on Built For Tomorrow.

Post 2

On his first day as a civil engineer, Nigel Wea spent six hours in the print room printing large plan sets and spec sheets and binding them together by hand.

That day stayed with him. Years later he built Codex Civil, a tool that took a project's type, location and municipal owner and pulled together the standards, calculations and background research for a preliminary design report.

His reasoning was pure economics. An engineering fee is often around 10% of construction cost, and on a lump sum it isn't a replenished pot. Every hour spent compiling standards at the start is an hour you don't have at 60 or 90%, when municipal comments arrive and one small mistake can mean a six-week turnaround. Get preliminary design done in days rather than months and you keep budget and time for the detailed design that matters.

Nigel called it attacking the first domino.

He later commented on a LinkedIn post about MeltPlan's funding, the CEO got in touch, and he's now Founding Product & Customer Success there.

We talked through all of it on this week's Built For Tomorrow.

Post 3

Construction companies are not short of data. Nigel Wea of MeltPlan put it plainly on Built For Tomorrow: they have it, it's just unorganized.

It lives in folders, in file cabinets, in Excel, on a thumb drive in the back of a pickup truck. A national firm's LA office stores things one way, New York another, London another again. So the data can't make them more efficient, because the process wasn't efficient to begin with.

That has a cost when a firm brings in AI. Nigel says one of MeltPlan's real struggles is the groundwork of getting a client's data out and organized before any useful work can start.

He sees a second roadblock over the next five years: IT departments and the red tape around letting data near any software, even when, as with MeltPlan, the models are not trained on the client's data. Some of that caution is fear of malicious vendors, some is fear for jobs.

Nigel's view is that AI will assist people's jobs and lead to more hiring, not replace the thinking.

If you're planning an AI rollout, organizing the data is where it starts.

Guest kit

Post for Nigel

I joined Hayley Swords on Built For Tomorrow to talk about how I went from civil engineering in water infrastructure to building AI for pre-construction at MeltPlan. We covered Codex Civil, the tool I built after too many days compiling standards for preliminary design reports, and the LinkedIn comment that led to a call with our CEO and a very honest scorecard on my work. We also got into why frontier models out of the box sit around 60% on construction documents, why MeltPlan won't put anything in the marketplace below 95%, and why I think document review is the easiest return on investment in the industry. Thanks, Hayley, for a great conversation.

Pull quotes

"So you had software as a service, but now it's service as a software."

"there's no better way to gain trust than to do good work for someone"

Guest kit prepared by Show Copy.

Clip moments

Clip 1: The 60% nobody mentions

Hook: Out of the box, AI gets construction documents about 60% right. Most users don't know that.

Opens: "One of the best pieces being a vendor or just being a consumer of AI is understanding its limitations, where it breaks, how much it can do on its own autonomously, and where you need to step in."

Closes: "They're taking it as viable. They're taking it as this is accurate. And that's where a lot of the mistakes are happening."

Why it works: It gives a concrete accuracy range for raw models against tuned ones and names exactly how mistakes get onto projects. Any GC or engineer piloting AI tools will recognise their own team in it.

Clip 2: Workflow 10, AI 2

Hook: The CEO scored Nigel's understanding of AI two out of ten. Here's how he took it.

Opens: "He wanted to see what I had built. I wanted to see what he had built."

Closes: "I want to hear this is cool, this is where you need to improve."

Why it works: A specific, slightly painful scorecard told with good humour. It is a founder story with a lesson about wanting blunt feedback, and it doesn't need any context.

Clip 3: Two weeks of unpaid review

Hook: Five or six people, two weeks, the same set of plans, and nobody is paid for it.

Opens: "From my experience, you might have five or six different people working to review the same set of plans and specs, and it might take them two weeks."

Closes: "Who doesn't want 90% gains in productivity? Everyone does."

Why it works: It turns a vague AI promise into a specific cost every design firm already carries, and makes the case for automating it in under a minute.