Tech firms’ hunt for data reaches HVAC, plumbing contractors

Tech companies are looking for help training robots and artificial intelligence models. 

The backdrop: As AI tools move beyond text and into the physical world, the once-obscure field of “data annotation” is booming. 

• Middlemen are gathering detailed information on how professionals across industries perform specific jobs, then organizing it into datasets, and selling or licensing it to AI developers and robotics companies. 

What’s happening: Minneapolis-based startup HUMXN Research Labs is one data licensor that’s specifically targeting the skilled trades, as well as farmers and manufacturers. 

• The company last week announced that it plans to spend $50 million acquiring real-world business information — service records, job-site photos, equipment manuals, training videos — from U.S. companies, including HVAC and plumbing contractors. (It remains unclear where the $50 million will come from.)
• It also said it’ll pay about $20 per hour of footage captured by technicians using wearable cameras to record their work and narrate as they move through jobs making diagnoses and repairs, explaining what they’re doing and why. 

Between the lines: “Before any purchase, [HUMXN] verifies that the business owns the data and has the right to sell it,” the announcement states. “All data is then de-identified: customer names, addresses, phone numbers, and other personal information are removed.”

Go deeper: Other companies are similarly buying up on-the-job data. 

• California-based micro1, for instance, pays Michigan HVAC company Vitality Mechanical a reported $1,000 per “labeling task” to understand how it handles vendor invoices, credit applications, and payroll, with Vitality in late July saying it expected to earn over $150,000 in a quarter. 
• micro1 announced on October 9 that it will spend $1 billion over the next 12 months acquiring and licensing enterprise operational data, with capital from Citi and venture lender Hercules Capital. 

What they’re saying: “OpenAI and Anthropic have run out of public data to train their models on,” HUMXN founder Thomas Unise told Homepros. “Like they’ve scraped the entire internet.”

• “So they’re offering companies like us that can take operational data from home service companies, manufacturing companies, bakeries, whomever, and de-identify that data and structure it,” he said.
• “If you’ve got 10 years of records, they’ll give you 500 grand, sometimes more,” he added. Contractors “could probably earn an extra 1,500 bucks a month per technician, literally doing nothing other than wearing Meta glasses.”  
• “These are trillion-dollar companies,” Unise said, “and it’s a data war.”

Of note: While Unise didn’t name specific clients, he said he works with “obscure robotics companies” and pointed to humanoid robotics maker Figure as the sort of company that would buy such datasets (but is not his client). 

Zoom out: The Washington Post in late September noted that tech companies have been on a “buying spree” for proprietary datasets, while the New York Times on Friday published video of a barber and nail tech recording themselves on the job for a company called Instawork.

• The Times has also reported that data annotation is booming in India, where gig workers earn about $2.50 per hour recording themselves from home. Time Magazine, meanwhile, has found U.S. jobs paying about $20 per hour for data-annotation tasks such as labeling photos.

Yes, but: While selling operational data may seem like a financial opportunity for businesses, tech companies aren’t paying “anything close to the actual value of that data for the company, because now that data is getting used to train this army of robots… that’s going to stick around forever,” Columbia University economist Suresh Naidu told the Times.

• “The companies want to get the data of how to do that job, and once they’ve gotten that, they don’t need you anymore,” he said. 
• The market “will probably be ultra-saturated in like a year and a half,” Unise told Homepros, “and then once they get enough data, it’ll be worthless.”

The big picture: Robots can perform roughly 75 percent of physical job tasks in the U.S., according to an Anthropic report released in September; however, they remain too expensive to displace humans.

• Go deeper