NavigateAI's new contracts price a saving that weather and crew composition also move

The invoice is where this gets engineering-shaped. Early NavigateAI contracts charged for tokens and took a margin under a usage-based software model; newer ones take a share of the economic value the system creates ^[6]. In the example Wu gave TechCrunch, as reported by mezha.net, a house that would have cost $300,000 lands at $280,000 and NavigateAI takes roughly 20% of the $20,000 difference ^[7]. That is $4,000 a house ^[8], on a cost reduction of 6.7% ^[9].

To bill it you need the $300,000. That number is a counterfactual, and the same account lists what else moves it: weather, crew composition and material availability ^[13]. Its stated remedy is a measurement program, comparisons between divisions plus A/B testing ^[14]. Read that as a product requirement. The unit of randomization is a crew or a community, the control arm is builds deliberately run without the tool, and the party staffing that comparison is the builder. Lennar invested in the round and is named by the company as a strategic backer ^[2], which puts one organization in three seats at once: buyer, measurement lab, and counterparty to any argument about what the baseline was.

The prize explains why anyone would carry that overhead. Wu puts Lennar's annual spend on labor, installation and construction at about $9 billion and says a 5-10% efficiency gain would be worth hundreds of millions ^[10]. That band is $450m to $900m ^[11]. Apply the house example's 20% share to it and one customer is worth $90m to $180m a year ^[12]. The 20% comes from an illustration rather than a disclosed Lennar contract, so that range is the shape of the bet, not a forecast.

The company's own dashboard splits along the same seam. Its homepage claims more than a billion video frames processed while the fields for labor savings and square footage on the platform both display zero, according to runtimewire ^[15]. Frames processed is a throughput counter and it costs nothing to earn. Labor saved and square footage covered are outcome counters, and they are the ones the pricing model depends on. The source reports the display state, leaving the reason unspecified, so an unpopulated field is a different claim from a measured zero. A counter reading zero is, at minimum, honest about not overstating the result.

The second revenue thesis has the same dependency in a different place. Every task done through the system produces first-person video labeled with correct and incorrect worker actions, which Wu argues could eventually matter to robotics companies as much as the software does to builders ^[17]. The worth of that corpus depends on whose hands are in frame. Wu told TechCrunch that younger workers pick the copilot up faster, while experienced journeymen lean on knowledge accumulated over decades ^[16]. Distribution pushes the same way: NavigateAI reaches new entrants through AIM, a Meta-backed fiber installation school that guarantees graduates employment, so trainees can meet the tool before they reach a jobsite ^[19].

Demand for any of this is real. The Associated Builders and Contractors estimate cited by mezha.net puts the 2026 US shortfall at roughly 349,000 workers ^[18], and a large data center campus that once peaked around 750 workers can now draw 4,000 to 5,000 on the biggest projects ^[21]. Scarcity like that buys pilots, not attribution. The first defensible number here will come from a division-level comparison inside a builder, run on jobs the vendor agreed to sit out, and it will be that builder's operations staff who produce it.