Ediphi founder Dustin DeVan has found a brand-new AI takeoff startup every single week for 16 straight weeks. Meanwhile, Trunk Tools CEO Dr. Sarah Buchner, who claims the strongest drawing AI in the market after tens of millions of dollars in R&D, refuses to sell a takeoff product at all. Her reasoning cuts through the hype: reading construction drawings properly requires three layers of intelligence most startups can’t afford to build, and per-trade takeoff accuracy is a squeeze that isn’t worth the juice.
Dustin DeVan keeps a counter. Every week, the Ediphi founder checks whether he can find a new AI takeoff tool he’s never heard of. On the latest Bricks, Bucks & Bytes, he reported the streak is alive at week 16, fed by a WhatsApp group where industry friends now submit fresh discoveries. Some weeks produce more than one.
Sitting across from him was the person you’d expect to be cashing in on that gold rush. Dr. Sarah Buchner runs Trunk Tools, which raised a $40 million Series B led by Insight Partners and has spent three years training its own models to read construction drawings. She says she’d back her drawing AI against anyone’s. And she still won’t ship a takeoff product.
That contradiction is the most useful lens on the takeoff frenzy you’ll find, whether you’re a GC drowning in cold outreach or a founder wondering if the 51st entrant can win.
Why drawings break generalist AI
Three layers of understanding, and most tools stop at one
Buchner’s framework for drawing intelligence has three layers. The first is text: title blocks, labels, notes. Large language models handle that fine, which is why so many startups can produce a convincing demo. The second is objects: can the model tell a single-swing door from a double, a single window from a pair, across 350-plus categories spanning MEP and architecture? The third is connectivity: understanding that a floor plan, a section, and a detail describe the same three-dimensional building, and that those drawings link outward to specs, RFIs, submittals, and finishing schedules.
“If you are doing only one or only two of these three layers, it’s useless,” she said. More than 80% of construction information lives in drawings, and drawings simply aren’t an LLM-readable file type. Getting past layer one demands computer vision models trained on labeled data that you cannot buy anywhere. Trunk Tools had humans label millions of construction drawings over years, then used a patented method involving BIM models to make the labeling scale. Others have landed in the same place: Primepoint just raised a $10M seed with Yann LeCun on the cap table to attack the exact same drawings problem with computer vision and knowledge graphs rather than an LLM wrapper.
When Martin recalled a founder claiming perfect drawing-reading would cost about 1.5 million euros, Buchner laughed. “Whoever did this, reach out to me, you’ll get a really highly paid job here.” Her company has spent far more than that on human labeling alone, and tens of millions of dollars on the drawings problem in total.
The company best placed to sell takeoff won’t
A TAM that isn’t worth the squeeze
So why no Trunk Tools takeoff product? Buchner’s answer is disarmingly commercial. Takeoff is the most obvious use case from a 30,000-foot view, which is exactly why every new founder lands on it. But getting accuracy good enough for real takeoff, for every trade, requires so much extra data labeling and layering that the market size doesn’t justify the effort. “The TAM is currently not worth the squeeze,” she said. If someone genuinely cracks it, she’s open to just buying them.
DeVan, who built BuildingConnected and sold it to Autodesk, sees a rerun of an old movie. When the iPad arrived, roughly 50 startups raced to put plans on mobile devices. People remember PlanGrid and Fieldwire. Almost nobody remembers the rest. We unpacked the venture logic behind funding the 50th entrant anyway in our piece on why YC keeps writing checks for AI takeoff tools. The short version: VCs are betting on teams, not products. That can be rational for a fund and still produce a graveyard of tools.
Don’t come on the market with something that Claude can already do and try to sell it. Do something real.Dr. Sarah Buchner, CEO, Trunk Tools
Cut through the demo-ware before it reaches your inbox
We track which construction AI is real and which is a wrapper, every week.
Join 3000+ ReadersThe flood is now a buyer’s problem
From 7 cold messages a day to 250 a week
The startup surge lands directly on the people expected to evaluate it. A couple of years ago, Bricks & Bytes asked construction leaders how many cold outreach messages they got from tech vendors; the highest answer was around seven a day. Last week, one executive told Owen he’s now receiving roughly 250 a week. AI didn’t just make it cheap to build demos, it made it cheap to spam the buyers too.
Buchner’s sympathy sits with the GCs and subs on the receiving end. “It’s almost impossible for them to actually assess what’s gonna work,” she said, because AI looks like a magic trick every single time. Trunk Tools now gives general contractors an assessment framework to separate genuine products from polish, and she was clear the goal isn’t to torch startups, just to protect an industry that can’t afford failed rollouts.
Her sharpest criticism was reserved for two categories: humans quietly masking as AI, which she called unscalable and misleading, and vibe-coded point solutions that repackage what a general-purpose model already does. Honest human-in-the-loop products get a pass; hiding the humans doesn’t.
What separates a demo from a defensible product
The checklist buyers can actually use
Strip out the noise and the evaluation criteria that emerged from the conversation are fairly concrete. They also explain why most of the week-16 counter entries won’t survive their Series A.
| Evaluation layer | What a wrapper shows you | What a real platform shows you |
|---|---|---|
| Drawing intelligence | Text extraction and a chat box | Object detection plus cross-drawing and spec connectivity |
| Training data | Whatever the foundation model saw | Years of proprietary labeled construction documents |
| Accuracy claims | A polished demo on cherry-picked sheets | Ground-truth eval datasets and confidence scores |
| Industry depth | Founders new to construction | Teams who can talk the talk and sell into it |
| Durability | Rebuilt by the next model release | Enterprise security, trust, and delivery at scale |
Takeoff is the most visible, most obviously automatable task in preconstruction, and modern AI tooling makes an impressive demo cheap to build. Buchner also pointed to a tough job market pushing people in their early twenties toward founding something, often without construction experience. The result is a flood of similar tools, one or more per week by DeVan’s count. (Source)
Because takeoff-grade accuracy has to hold up for every trade individually, and reaching that bar requires enormous additional data labeling on top of what general drawing understanding needs. Buchner judges the total addressable market too small to justify that investment, though she said Trunk Tools would consider acquiring a team that truly solves it. (Source)
Text (title blocks and notes, which LLMs can read), objects (identifying 350-plus element types like door and window variants, which needs trained computer vision), and connectivity (linking floor plans, sections, and details to each other and to specs, RFIs, and submittals). Buchner argues a product delivering only one or two layers is useless for real construction work. (Source)
Buchner is skeptical on the timeline. Drawings mix computer vision, spatial reasoning, and cross-document references in ways generalized models handle poorly, and no major lab is building construction-specific workflows around them. If a breakthrough lands, she says Trunk Tools works closely enough with the model providers to adopt it first, and the surrounding delivery, security, and trust would still need building. Primepoint’s founders, who include a Facebook AI Research veteran, made the same argument when raising their seed round. (Source)
DeVan’s reference point is the mobile-plans wave of the early 2010s, when about 50 startups chased the iPad opportunity. PlanGrid, which sold to Autodesk for $875 million, and Fieldwire are the names people remember; the rest largely vanished. Crowding didn’t prevent big winners, but it guaranteed a long tail of casualties. (Source)
Ask about all three drawing layers, demand ground-truth accuracy evidence rather than a demo, probe whether humans are quietly doing the work behind the scenes, and check enterprise security and data terms. Buchner’s team publishes an assessment framework for GCs precisely because polished demos and real products have become hard to tell apart. (Source)
Related Articles
Why YC Won’t Stop Funding AI Takeoff Tools (And Why That Might Be Smart)
Primepoint Raises $10M Seed to Build AI That Actually Reads Construction Drawings
