Why every AI conversation in construction eventually circles back to the same uncomfortable question and what we heard from a $1B+ GC who is actually doing something about it.
When Last week I sat in a hotel ballroom in Phoenix and watched the VP of Preconstruction at a billion-dollar general contractor open his session with a slide about Christopher Columbus.
Stay with me on this.
His name is Rich Scopelliti. He runs preconstruction for Consigli Construction, a Northeast GC with 16 offices and roughly 200 estimators. He had just flown in from Boston that morning, somehow made it through a Spirit Airlines tarmac situation, and was about to walk a room full of preconstruction leaders through everything his team has gotten right and wrong, trying to build dynamic cost models.
And he opened with two stories.
Story one: Christopher Columbus. Relied on outdated maps, ignored geography, leaned on institutional knowledge from people long gone. Sailed to completely the wrong continent. Made history anyway.
Story two: The Mars Climate Orbiter. NASA, 1999. A $125 million spacecraft that disintegrated in the Martian atmosphere because one team used metric units and another used imperial. Same kind of bad data. Wildly different outcome.
Rich’s point: “Both relied on bad data. One got lucky. One didn’t. In our business, we don’t want to be either.”
I had just spent the previous Thursday co-presenting our AI in Construction webinar with Rowan Steel-Hall of Smartbuild. Same message, different audience. And here was Rich, in front of 200 estimators, hitting the same wall from the inside.
Every AI conversation in this industry, every single one, eventually circles back to the same uncomfortable question:
Is your data actually any good?
The lie we tell ourselves
Rich said something during his session that resonated with the audience.
He asked the room: “Who here has given a conceptual estimate to an owner and that’s the only number they ever remember?”
Every hand went up. Mine did too, different industry, same dynamic. Anyone who has ever quoted an MSP engagement before discovery knows exactly what he was describing.
In construction, the cost is real. A conceptual estimate gets thrown out in the first 60 seconds of a phone call. The owner anchors to it. Six months later, when the real number comes in higher, the GC is fighting the tide for the entire project.
Rich called it “the lie we tell ourselves.” The estimator believes the number. The owner believes the number. The number is built on data that nobody is sure of.
Garbage in, garbage out. The oldest line in computing. Still the truest line in construction.
Why this gets harder, not easier
Here’s where it gets uncomfortable for anyone running a preconstruction team in 2026.
AI is now sitting on top of all of this. Estimating platforms are advertising AI-driven forecasting. Procore, Autodesk, RSMeans, Centric, every one of them has a story about how their models are going to make your bids faster, sharper, more defensible.
And every one of those stories assumes you have clean, standardized, calibrated historical data underneath.
Most contractors don’t. Including the very good ones.
Earlier in the same event, the data team from Alberici Constructors, another billion-dollar GC, walked through what they found when they actually audited their own historical project data. Out of roughly 620,000 line items going back to 2010, only 45% were cost-coded correctly. Fewer than half. The rest were sitting under the wrong CSI codes, miscategorized as “subcontractor” with no division reference, tagged to people’s names instead of work types, or just floating in the system with twenty-character descriptions like “misc concrete stuff - Bob.”
Their solution? They had to hand-code every line item across a dozen projects by hand to build a training data set, then train a machine-learning model to go back and recategorize the other 615,000-plus line items so the data could actually be queried.
Read that paragraph again. A $2B+ GC with a dedicated data team and a Power BI dashboard had to manually retrain itself on its own history before any AI could touch it.
One of the GCs at our table summed it up perfectly: “We have 25 years of data. Most of it is in Bob’s spreadsheets. Bob is retiring in November.”
Translation: Your AI strategy is only as smart as the database underneath it. And the database is only as smart as the standards you used to build it. And the standards only matter if Bob actually used them.
What Consigli is actually doing about it
Rich was refreshingly honest. He didn’t pitch a product. He described a five-year roadmap, year one of which they have already burned just collecting and cleaning data.
Here is the framework, distilled. It applies whether you build hospitals or office TIs:
- Pick a standard. They chose CSI coding. Not because it’s perfect, it isn’t, but because it’s an industry standard with built-in update paths.
- Enforce it consistently. Across 16 offices and 200 estimators. This is where it falls apart for most companies.
- Backfill the historical data. The painful part. Re-coding old estimates so the new standard actually has a foundation to learn from.
- Standardize attributes between markets. A K–12 school in Boston and the same school in Phoenix don’t cost the same. The labor multipliers, the energy codes, the local fee structures all differ and they have to be parameterized, not assumed.
- Build the access layer. Make the data instantly queryable so the entire company can use it the moment a prospect calls.
And here’s the kicker. Consigli decided not to buy. They built it themselves.
Their lead developer is a senior estimator. He had never written code three years ago. He started ‘vibe coding’ with ChatGPT 3.5 in 2023, graduated to Gemini, then Copilot, and now runs a virtual machine environment with Claude doing the heavy analytical lifting. He has built three modules so far. Tens of thousands of historical cost records, queryable in seconds, with full source-document traceability.
Built by an estimator. Not a software company. Because the people closest to the pain built the cure.
The MSP angle nobody wants to hear
Here is where I have to be honest about my own seat at the table.
Computer Dimensions doesn’t build cost-modeling software. We support the construction companies who use cost-modeling software, AI-driven estimating tools, project management platforms, and everything else.
And what we see, every single day, is the exact same gap Rich Scopelliti described, just one layer below his.
The Consigli story is about preconstruction data being a mess. The story I see is that the data infrastructure underneath that preconstruction data is usually a mess too. Folder structures that nobody owns. Permissions inherited from people who quit in 2017. Backups that haven’t been tested. Project files scattered across three personal OneDrives, two SharePoint sites, and the desktop of a project manager who is on vacation.
Five years ago, that was an inconvenience. Today, it’s the difference between AI making you faster and AI making your bad data faster, which is just a more efficient way to lose money.
The order matters. Clean infrastructure, then clean data, then standardized data, then AI on top of all of it. Skip a step and the whole stack collapses.
Three questions worth asking on the way to your truck this week
If Rich’s session was a wake-up call for billion-dollar GCs, here’s the version for the rest of us:
- If our top estimator retired tomorrow, could we still bid a project at the same accuracy? If the answer involves a sigh, you have a data problem.
- When we look at a similar project we did three years ago, can we pull the actual costs in under 10 minutes? If the answer is “I’d have to ask Bob,” you have a system problem.
- Are our project files, financials, and estimating data sitting in one place with consistent naming, real backups, and access controls that match the people who actually work here today? If you flinched, you have an infrastructure problem.
None of these are AI questions. All of them are AI prerequisites. The contractors who win the next five years are going to be the ones who quietly fix this layer while their competitors are arguing about which AI tool to buy.
One more thing from the ballroom
After Rich wrapped up, he handed the room over to two of his estimators, the ones who actually built the system. Neither of them had a computer science background. Both spent a few minutes explaining that the magic of what they built wasn’t the AI. It was the underlying data structure.
One of them, almost as a throwaway, said this:
“The AI part is easy. The data layering is the sauce.”
If you took one thing away from our webinar last Thursday with Rowan, or from any of the AI conversations rolling through our industry right now, let it be that line.
The AI part is easy. The data is the sauce.
Get your sauce right.
For over 20 years, Computer Dimensions has been the trusted IT partner for Arizona's architecture, engineering, and construction industry. We help AEC firms communicate better, collaborate smarter, and actually use the technology they've invested in. Because in construction, the tools only work if your team does.
IT Built For Builders.
P.S. Part 2 of our AI in Construction webinar series is coming in mid-June. This one is the hands-on workshop: less framework, more rolling up sleeves on real Arizona contractor use cases.
AI In Construction Part 2 – The Live Build Workshop
