❓ Problem

AI labs have run out of internet. By late 2024 they had already hoovered up most of the public web, and the agents they're building now need something the web can't give them: how real companies actually work. Slack threads, Jira tickets, SOPs, support queues. That's the raw material for "RL gyms", simulated workplaces where agents practise being employees, and Anthropic has reportedly discussed spending more than $1bn on those environments.

So the buyers have gone shopping. micro1 advertises data partnerships from $100k to $1m+ and reportedly offered a Michigan HVAC company $150,000 for its records. Mercor emailed Warmly's CEO eight days after the HubSpot dealasking to buy or license its code and task records, with offers reaching $300k. That onboarding doc nobody has touched since 2022? It's an asset now.

Here's the catch. Almost everyone brokering these deals works for the buyer. micro1 even pays referrers up to $50k for bringing companies in. Wind-down platform SimpleClosure has done nearly 100 deals and recovered over $1m for founders, which works out at roughly $10k a company.

Now look at what happened when Spirit Airlines' data went to auction. Google opened at $5m, Mercor came in at $7.5m, and Google won at $10m. Same kind of asset, double the price. The difference was competition.

Most founders have never priced operational data, negotiated permitted uses or exclusivity, or worked out which parts they're even allowed to share. When an offer lands, they're negotiating against people who do this every week.

✅ Solution

A sell-side brokerage that represents the company selling access, not the lab buying it. Think M&A adviser, but for your operational exhaust.

• Start with an approved sample. The owner shares a small, pre-cleared slice (a folder of SOPs, a month of anonymised support tickets) and the brokerage assesses what buyers will pay for: workflow depth, domain rarity, volume, and how messy and real it is.
• Define the package. Help the owner decide what's in, what's out and what never leaves the building. Customer data, HR files and anything under NDA get ring-fenced from day one.
• Run a competitive process. Take the package to labs, data vendors and RL gym builders, seek competing bids, and negotiate scope, permitted uses, exclusivity, retention, deletion and payment terms. One-off sale vs recurring licence is one of the biggest levers on price.
• Coordinate specialist review. Bring in privacy counsel and a redaction partner before anything changes hands, so personal data is scrubbed and the rights are clean.
• Vet the buyers. Know exactly who is getting the data and what they'll do with it.

Start as a hands-on service for companies that already have an offer on the table. Then productise what repeats: a data-readiness audit, a valuation model, standard licence templates and a deal room.

📊 Key Numbers

Market size

• The AI training dataset market was worth around $3.2bn in 2025 and is forecast to reach $16.3bn by 2033.
• The top end of licensing is huge: Reddit's Google deal is worth about $60m a year, with a separate OpenAI deal estimated at around $70m a year.