Boon

The number of real plan sets in the world is finite. The amount of training data a vision model needs is not. That gap is the actual reason most "AI for construction" tools don't work past the demo. There aren't enough real plans, drawn the same way, labeled cleanly, to train a model to the accuracy a preconstruction team can trust. You can't scrape your way out of it. You can't licensing-deal your way out of it. So we make data up. Carefully. Synthetic plan sets that have to clear three bars before they're allowed near a training run. They have to look like real plans an architect would actually produce. They have to teach the model the specific thing we want it to learn next, a new symbol family, a new sheet style, a new revision pattern. And they have to reflect real building code. What a hospital needs is not what a high school needs is not what a data center needs. This is the unglamorous half of AI in the physical world. No chat UI. No agent. Just a team that decided the data problem was the actual problem, and built around it instead of around it. Full episode with @Reece Barnes: https://lnkd.in/gqgc9Kih Full clip: https://lnkd.in/gA7nmbpH

One line from last week's Builders, Budgets & Beers episode keeps surfacing in the conversations I'm having this week: "Something that takes two days, you can finish in two minutes." Customers ask me what's behind that number. Honest answer: it's not magic and it's not a single model. It's: Specialized vision models trained on construction plan sets, not on photos of cats. Integration with the estimator's existing system so historical data flows in automatically. An agent layer that lets the estimator review and correct, not start from zero. The 15 minutes of review still matters. The two minutes is what gives the estimator their Friday night back. Full episode if you missed it: https://lnkd.in/gqgc9Kih

Two weeks ago a VP of Preconstruction at a top-50 ENR contractor walked up to our booth and asked the question I now hope every estimating leader starts asking every vendor: *"What's your F1, by trade?"* Here's why that question matters more than any "% accuracy" claim you've ever been shown: *Precision* = of the things the AI marked, how many were actually right? Low precision = false positives = your team deletes work all day. *Recall* = of the things on the drawing, how many did the AI catch? Low recall = false negatives = your bid is short. *F1* = the honest single number. It balances both. You cannot game it. "90% accuracy" is a marketing number. F1 is an engineering number. And here's the part most vendors don't want you to ask: F1 has to be reported *per trade.* A vendor claiming "95% accuracy on floor plans" is telling you nothing about how they'll do on your MEP rough-in, your structural rebar, or your civil earthwork — which is where the bid risk actually lives. We publish per-trade F1 to every customer, every month. Movement up or down. Honest numbers. If your current vendor can't show you the same — that's the signal.

2026 is the year every estimator became a manager. Not of people. Of agents. The story I keep hearing in the field is the same one I lived running engineering teams. Three years ago the best engineers were the ones who could write the most code, fastest. Today the best ones are problem solvers who prompt, review, and direct AI to do the writing. Estimating is going through the exact same shift. The mundane work, count this, measure that, re-run after the addendum, is moving to agents. What's left is the part that actually pays: judgment, strategy, knowing which job is worth bidding and how to win it. The job didn't shrink. It moved up a level. The estimators I talk to who are leaning into this aren't worried about being replaced. They're worried about the ones who aren't leaning in. Because the bidding cycle is compressing, the data center work isn't waiting, and the firm that can put a strategic estimator on top of three or four agents is the one taking the work. If you've been told AI is here to replace you, that's the wrong frame. AI is here to report to you. Get good at managing it. Full episode with Reece Barnes on Builders Budgets & Beers: https://lnkd.in/gqgc9Kih Full clip: https://lnkd.in/g7CAcjCw

Mechanical takeoff is where AI vendors usually quit early. Ducts share sheets with structure, with electrical, with sprinkler mains. The takeoff is only useful if the model reads all of it at once. What that read looks like. Ductwork, fittings, equipment, schedules, all from the same pass on a full bid set. This is the work we said we'd ship at Advancing Precon. It's running now. https://lnkd.in/g_Ffudih