Eric Wu built and ran Opendoor, one of the more ambitious real estate startups of the last decade, before stepping away in 2022 after fast-rising interest rates abruptly slowed down home sales. He spent a year resetting — he’d been running the company for eight years at that point — and could have easily jumped into investing when he felt done with his hiatus. But like a lot of founders, he’s so convinced that AI will be the defining tech platform of his lifetime that he more recently decided instead to dive back into company building. As he told me during a call earlier this summer, “I knew if I looked back in 10 years and didn’t do something related to it, I’d probably regret that.”

The opportunity Wu is chasing — building AI copilots for construction workers and other field laborers, something he describes as a hands-free expert coach for the people physically building things — isn’t exactly a secret. The construction industry is short workers by a wide margin.

The trade group Associated Builders and Contractors has said roughly 349,000 additional workers would be needed this year just to keep pace with construction demand, a challenge that looks to worsen with workers getting older, workers getting shipped out of the United States owing to more aggressive U.S. immigration enforcement, and also the growing number of major projects being planned, specifically data centers.

We’ve all heard now — from the entire AI industry — about the giant server farms that the AI boom demands. Those projects have grown enormous, and staffing demand has grown with them. Where a large data center campus once needed something like 750 workers at its peak, the biggest projects now require far more. Meta’s Hyperion campus in Richland Parish, Louisiana, for example, will reportedly require roughly 5,000 construction workers, and OpenAI’s Stargate site in Abilene, Texas has reported involved 6,400 workers.

Kelly, the global staffing company, has said that 90% of data center operators now name staffing shortages as a critical constraint on their ability to build or expand.

Wu’s new company, called NavigateAI, meant to address that people layer, officially launched in late May with $25 million in seed funding at a $225 million post-money valuation led by Elad Gil, with participation from Khosla Ventures, Fifth Wall, real estate giant Lennar, Tishman Speyer, electrical contractor Helix Electric, and a roster of angels including Tony Xu of DoorDash, Apoorva Mehta of Instacart, and Coinbase CEO Brian Armstrong.

None of those backers is particularly surprising, given Wu’s connections to the real estate industry, as well as to Silicon Valley bigs. Gil, for example, was an investor in Opendoor, and Keith Rabois of Khosla Ventures co-founded Opendoor. Khosla Ventures founder Vinod Khosla has also been outspoken in recent years about the dozens of startups in the firm’s portfolio that are building around AI “workers” of various stripes, whether oncologists, chip designers, or construction workers.

Wu’s core product runs on smartphones and, in hands-free mode, through Meta’s AI glasses. The idea is for a construction worker to point the camera at what he or she is building and ask, in plain language, whether it’s installed correctly, whether the torque is right, whether it meets code, and so forth. Wu says that NavigateAI can pull up building specs, manufacturer manuals, and company policy in real time. He also says the hands-free experience is measurably superior (you don’t want to be looking at a phone if you can avoid it) and that, in fact, the company is working with Meta to get the glasses safety-certified for environments where protective eyewear is required.

NavigateAI is also partnering with AIM, a Meta-backed fiber installation trade school that guarantees job placement to graduates, giving the company a channel to reach workers before they ever set foot on a job site.

Getting workers comfortable with AI-assisted work during training matters a lot, apparently. Wu says adoption curves split sharply between younger workers, who he says embrace the product, and 30-year journeymen, who trust their own instincts and aren’t exactly lining up to strap a computer to their face.

On the business model, the company started with a token-plus-margin pricing structure — akin to usage-based SaaS — but it has migrated its newer contracts to a share of value created. For example, if NavigateAI helps a builder reduce the all-in cost of a home from $300,000 to $280,000, the company captures roughly 20% of that $20,000 savings. Wu told me Lennar — one of the country’s largest home construction companies and another of NavigateAI’s investors — spends roughly $9 billion a year on labor, installation, and construction, so even a 5% to 10% improvement would represent hundreds of millions of dollars in potential value.

The longer-term play, which Wu is candid about, is the data. Every job completed with NavigateAI generates labeled egocentric video of field workers building and maintaining physical things correctly and incorrectly, and that’s a dataset that Wu believes will eventually be worth as much to robotics companies as the software business itself is worth to Navigate’s construction clients.

There are, of course, real challenges. Value-based pricing can create an attribution dilemma , as Wu openly acknowledges. Proving that a home was built faster because of NavigateAI and not because of sunny weather, or the particular crew on the job, or the availability of materials, requires A/B testing across divisions and is approximate by his own admission. You could imagine a client dispute over savings attribution getting complicated fast, even if that client is also an investor in NavigateAI.

That resistance from experienced workers would also seem to be a pretty serious problem. The veteran journeyman is whose industry knowledge would most make NavigateAI’s product smarter, and no trade school can fully replace that know-how. Then there’s safety liability. If NavigateAI’s software clears a connection that later fails, what happens? Defect liability is a notoriously litigious area, and it’s not yet clear how these issues will be handled as they invariably arise.

There’s also, as ever, the competitive question. When we talk, Wu mentions that the most common current alternative is a worker Googling something or asking ChatGPT, and that NavigateAI can go well beyond that. But all the big LLM companies have the model capabilities and, in Meta’s case, the hardware distribution, too. As unlikely as it is, they could spin up their own businesses.

Of course, the more likely threat is a similar player. NavigateAI’s defensibility rests on workflow integrations and proprietary data that take years to accumulate. Wu himself mentions Buildots and OpenSpace as the closest points of comparison, but he says that they’re “more focused on project management,” while Navigate is built around “the individual labor.”

Either way, in Silicon Valley, network matters sometimes as much as the product or the competitive landscape, and the combination of Wu, Gil, Khosla Ventures, and Lennar, among others, functions as strong signal. Besides, Wu doesn’t seem inclined to worry right now about competitors. He just seems excited to be building something new, at a time when not building would feel like a mistake to him.

Underscoring that point, he doesn’t have a board yet, and he says he is “going to try to go as long as I can without one,” so he can stay focused on customers instead of governance. You can draw a line between that observation and his time running Opendoor, which went public through a special purpose acquisition company back in late 2020. Being a public company CEO made the parts of the job he wanted to spend his time on harder to do, forcing a constant trade-off between building product and managing a board and shareholders. For now, it’s clearly a trade-off he doesn’t miss.