Last fall, a mid-size civil engineering firm I work with tried something that sounded great on paper. They wanted to feed two years of project closeout reports into a large language model and have it surface patterns in cost overruns, schedule slips, change orders. The kind of institutional knowledge that usually disappears when a senior PM retires or moves on.
They spent six weeks on it. Burned through about $40,000. Then killed the whole thing.
The technology wasn’t the problem. The data was. Their closeout reports were spread across three different formats, living on two SharePoint sites and a shared network drive nobody had bothered to migrate. Half the PDF files were scanned images with no searchable text at all. Naming conventions had changed twice in 18 months, so the model couldn’t distinguish a $2-million highway interchange from a $200,000 drainage repair. Both showed up as “Infrastructure — Misc.”
I’ve been consulting on AI and data strategy for AEC and manufacturing firms for about five years now. This story isn’t unusual. I’d say some version of it plays out at nearly every company that tries to jump straight to the exciting part without doing the boring work first.
I saw something similar at a regional MEP contractor, about 200 people. They wanted AI to speed up submittal reviews, comparing incoming submittals against project specifications and flagging discrepancies. Solid use case, but their specs came from a dozen different architects who all formatted things differently. Submittals lived in Procore, in emailed PDFs at that, and in one case, as photos of a whiteboard someone snapped on their phone. The AI choked on it, and honestly, so did the team trying to make it work.
Harder Than it Looks
What’s driving this? The pitch from AI vendors sounds incredible. Drop your documents in, ask questions in plain English, get answers instantly. After years of digging through Procore exports and buried email chains, that promise feels like it might actually fix something. I get why people buy in.
However, what the demo doesn’t show you is the gap between clean sample data and your actual files. In the demo, everything is formatted, digitized, consistent. In your firm, the data is whatever your team produced under deadline pressure over the last 10 years. That’s a very different animal.
Looking for quick answers on construction and engineering topics?
Try Ask ENR, our new smart AI search tool.
Ask ENR →
I’ve started calling this the “foundation problem.” It’s the single biggest reason AI pilots stall in this industry. Everyone focuses on the technology layer. Almost nobody focuses on the data layer underneath it. Vendors aren’t going to bring this up, because their sales process depends on you believing the tool works out of the box. Some of them actually think it does. They’ve just never tried running it against a real AEC firm’s file structure.
Think about what a typical contractor is actually sitting on. Specs in PDF, a lot of them scanned from paper originals that are 15 or 20 years old. RFIs living in email threads, in Procore or some other cloud platform, sometimes still on paper logs in a filing cabinet. Submittals are tracked in spreadsheets where every project manager uses a different format. Lessons learned documents that, theoretically, exist but that nobody can actually find. Drawings scattered across Autodesk Construction Cloud, Bluebeam sessions, and local folders that were supposed to sync but didn’t.
Asking an AI to make sense of that is like handing someone a shoebox full of receipts from five differ