Handoff H1: a takeoff model that outperforms human estimators

TL;DR: H1 reads construction blueprints and produces a complete material takeoff autonomously, from the raw PDF. It scores 81.6% on our benchmark — ahead of human estimators (77.6%) and ~20 points clear of the best frontier model. It runs in ~2 hours vs. up to a week for a human.

The problem

Before a contractor can bid a job, someone has to do a takeoff: read every sheet, count materials trade by trade, measure walls and roofs, cross-reference dimensions that live on different pages. It takes an experienced estimator 8–20 hours per project and years of training to do at all. It's the tax on every bid in a huge, fragmented industry.

You'd think a frontier model could do this now. It can't. Foundation models today aren't capable of understanding construction drawings to a level that is actually useful for a construction business. So this massive $2 trillion industry still runs on blueprints that are being analyzed by humans pretty much the same way they have been for the last few hundred years.

What we built

Handoff has developed H1: an AI model that understands and analyzes blueprints better than a human estimator, thus resolving this manual bottleneck once and for all.

H1 has three layers:

• Scope analysis — a detailed, trade-by-trade breakdown of what the project actually involves, read straight from the drawings.
• Takeoff with quantities — a complete material quantity list per trade: framing members, sheets, linear and square footage, counts, all with units.
• Localized construction costs — accurate pricing on those quantities, localized to your ZIP code.

All of it programmatic and autonomous. No human needs to be involved in the process of creating the entire takeoff. You can simply upload plans and get structured scope, quantities, and costed output back.

The benchmark (TakeoffBench-V1)

As part of this, we are also releasing the industry's first takeoff benchmark that enables us and anybody else in the research community to compare models against 15 real, permissioned, PII-stripped residential blueprint sets, each paired with a consensus-validated expert takeoff.

Here's how we performed on the benchmark:

• Handoff-H1: 81.6%
• human estimator (with takeoff software): 77.6%
• best frontier model (Claude Fable): 61.4%

The rest cluster in the low-to-mid 50s (gpt-5.6, Claude, Kimi) down to 34.7% (DeepSeek-V4-Pro).

Try it

H1 is live in the Handoff product and on the API. We have one of the world's largest suppliers already using our model in every one of their locations across the country, including thousands of projects every month.

If you're looking for an intelligence layer to extract detailed structured data from blueprints with construction costs priced out to the nearest cent, hit us up at api@handoff.ai