The hottest business in AI right now is selling data.
Data brokers have sold credit histories, location data, consumer profiles, and marketing lists for decades. AI changed the buyer and the price. Model developers now want private records that capture how people make decisions and complete real work.
The companies supplying it are already moving serious money.
Polyshares says its typical operational-data licenses range from $100,000 to $2 million. Its
published case studies include $450,000 for a roofing company's records, $300,000 for a general contractor, and $110,000 for a marketing agency.
Troveo reports paying more than $20 million to over 7,000 rights holders and
says it works with more than 40 active buyers.
Those figures come from the companies themselves, so they are not an audited measure of the whole market. They still show that proprietary data licensing has moved beyond a theoretical business model.
Every established business has years of operational history sitting inside email, Slack, support tickets, CRM records, Drive folders, call recordings, field notes, and internal documents. Most of it was created as a byproduct of doing the work. AI companies increasingly see it as training material because it contains something the public internet does not: the messy path from a problem to an outcome.
That difference is what buyers are paying for. The people sitting on useful records often have no idea they own a sellable asset. Even when they recognize the value, they usually do not know which AI companies are actively buying it. Buyers have the opposite problem: they know what expertise or behavior they need, but they do not know which businesses have years of it buried in internal systems.
This is a connection game. The reseller's edge is knowing both sides: which companies hold scarce data and which buyers have live demand for it. Data companies find the records, establish who owns them, prepare them for training, and bring them to the right AI buyers.
The word "reselling" makes this sound like a contact-book business. The serious companies are underwriting the data. They have to determine whether the seller has the right to license it and understand what a model can learn from it. They may need to anonymize, label, segment, convert, or combine it with other datasets. Then they have to deliver it in a form that survives a buyer's technical and legal review.
Over the past week, our team has been speaking with people across the robotics version of this market.
What stood out was how little the hardest problem had to do with cameras. The data either exists already or can be collected. The harder problem is turning scattered supply into something a buyer can legally use, technically ingest, and trust.
|
What we found in robotics data
The emerging egocentric-data industry
Egocentric, or "ego," data is first-person video of people completing real tasks, usually captured with a head-mounted phone or camera. Robotics labs use it to learn task sequences and hand-object interactions at a scale robot fleets cannot yet produce. It does not replace robot-native demonstrations, but it is much easier to collect across diverse environments.
Companies such as Shift recruit people to record household work and license approved footage to AI labs. Amelia is building a similar collection network across homes and commercial workplaces, handling task briefs, permissions, quality review, metadata, licensing, and delivery.
To understand who was collecting, buying, and researching this data, our team launched data.bycrux.com and asked the market to show us what it was working on.
|
We are sharing the robotics example because it is how our team got close to this market. Mapping the ego-data industry made the broader opportunity tangible: companies can hold valuable data without recognizing it, while buyers struggle to find the right owners. Robotics is one branch of a much larger connection game. The same dynamic may apply to the operational records inside your business.
