The most common reason estimators don't trust AI: they tried ChatGPT once, watched it hallucinate on a spec, and decided AI isn't ready.
That's a fair conclusion from a bad sample. The construction-trained version is a different product.
https://t.co/tHWtcT4NLo
In construction you do not get a second bid because your first one was close. You get a second bid because your first one was right and the GC relationship survived to the next project.
Every preconstruction lead in a pilot this quarter has the same question for AI vendors: what happens on the hard sheet. The one with the regional symbol convention and the addendum that contradicts the schedule. 🧵
A zoo of classifiers has no shared representation to anchor cross-trade relationships.
You can bolt heuristics on top. They fail on exactly the hard cases where the value is highest. The foundation model learns the topology in-distribution.
When the state of the art in vision encoders moves — and it has moved several times in the last 18 months — a zoo competitor runs 37 retraining cycles to absorb it.
A foundation model retrains the encoder once. The gap widens quarter by quarter.
Three things a 37-model construction AI vendor cannot do, no matter how much more data they collect:
1. Get cross-trade transfer
2. Amortize encoder improvements
3. Reason over physical connectivity
These compound. 🧵
Construction data is radioactive by contract. Sealed bid sets carry owner confidentiality. Federal jobs carry CUI rules. Many sub agreements explicitly prohibit third-party cloud handling.
The compliance boundary decides where the software has to live.
The drawings that teach a model real work live in NDA archives at GCs, in sub records, in permit offices that rate-limit downloads, on local servers in preconstruction trailers.
None of it is in Common Crawl. None of it is on GitHub.
Two tools, two points in the workflow:
AI on extraction.
Bluebeam on review, markup, collaboration.
Estimator on the judgment calls in between.
The problem isn't the markup layer. It's the 8 hours of clicking before it.
Bluebeam is what 2M+ AEC professionals use to mark up, review, and collaborate on docs. When a GC sends a plan set, they expect Bluebeam markups. It's the shared language.
AI removes the manual counting that has to happen before any of that work starts.
Every VP of Preconstruction asks me the same question about AI takeoff tools: do I have to give up Bluebeam?
No. And any vendor pitching that is solving the wrong problem. 🧵
Foundation > zoo. The math doesn't care which feels easier to build.
Deepti on why this is the architectural choice that decides the next decade:
https://t.co/rtVZsjKfKo
Three questions to ask any AI takeoff vendor:
1. One model for the whole sheet, or 37 stitched together?
2. Does adding data for one trade lift the others?
3. When a new vision encoder ships, how many of your models retrain?