They Ran Out of Internet

On October 9, micro1 announced that it will spend $1 billion over the next twelve months acquiring and licensing the operational data of ordinary companies. Not research data. Not the web. The CRM records, chat logs, support tickets, standard operating procedures, playbooks, and records of how teams make decisions inside businesses of a few dozen to a few hundred people, with the money coming from Citi and Hercules Capital.

Two weeks later I took a call with their data partnerships team, because I wanted to understand it from the inside. I won’t quote a private conversation, and I don’t need to. The public record explains the whole thing. And it should matter to anyone whose work product is expertise. Now they want your judgment, your lived experience, and every call you ever made before you could say why.

They Ran Out of Internet

For about five years, the recipe for a better model was more text. Scrape the web, filter it, train on it, repeat with more. It worked until the web was gone. Epoch AI estimated in 2024 that the stock of public, human-written text would be fully used by the frontier labs sometime between 2026 and 2032. The high-quality part went first. The same crawl, read again, teaches a little less each time.

This isn’t AI slowing down. It’s AI changing its diet.

What Replaced It

A model learns in three ways, and it helps to keep them apart. It can read. It can be rated. And it can practice.

Reading is what ran out. For years the labs fed models the written web and had them predict the next word, page after page, until they could write like it. The text was its own grader. No task, no environment, no human, just more pages. That’s the curve above.

Rating is what made the models pleasant to use. People looked at pairs of answers and picked the better one, millions of times, and the model learned what a good answer looks like. This was the first kind of reinforcement learning most people met, and the grader was human preference. It taught manners and helpfulness. It didn’t teach anyone how to do a job.

Practice is where agents come from. The model is dropped into a simulated environment that behaves like the real tool, given a task a person would be asked to do, and graded on whether the job got done. Then it tries again, thousands of times, and the behaviors that earn a good grade become habits. It isn’t learning what people said about the work. It’s learning to do the work. When the industry says it’s buying environments, it means the whole kit, the sandbox, the tasks, and the grader that scores each attempt without a human watching.

Code fell first because all three parts were lying around. GitHub was the sandbox, bugs and feature requests were the tasks, and a test suite is a perfect grader. Either the tests pass or they don’t. That’s why the first agents that could finish a job on their own were coding agents.

Back-office work had none of the three. There’s no GitHub for procurement. Nobody ever published how a company negotiates a supplier agreement or decides which exceptions to accept, so there’s no sandbox. And there’s no test suite for a good decision, because the right call on one contract is the wrong call on the next, and the difference is a fact nobody wrote down.

That’s what the operational data is for, and it does two different jobs. The exhaust, the tickets, the email, the CRM history, becomes the sandbox and the tasks. It’s the world the agent practices in and the requests it practices on. The decisions, what the team did and why, and the human verdicts on an AI’s attempts, become the grader. They’re the reference answer each attempt is scored against. Picture the procurement environment a lab would want. A simulated inbox with a renewal notice, a simulated CRM with the vendor’s history, the master agreement and the amendment nobody remembers signing, and one task, decide whether to renew and on what terms. The exhaust builds all of that. The grader is a record of what your best people decided in that situation, and why. Without it the agent can practice forever and never know whether it got it right.

So the labs started paying for environments. One frontier lab was reported to be discussing more than $1 billion a year on them, a supply chain of vendors appeared to build them, and the vendors went looking for the raw material. What goes in is a company’s operating history. What comes out is agents that can be handed a job inside your software and carry it through to the end. That’s what sits behind the word agentic.

What They’re Buying

Read micro1’s shopping list again. Standard operating procedures. Knowledge bases. CRM data. Project histories. QA processes. Records of how teams make decisions. Human feedback on AI outputs. The last two are the point. The rest is scenery.

The last two are the grader. Everything else on the list builds the sandbox. Sandboxes get cheaper by the month, because exhaust is everywhere. Graders stay scarce, because a grader has to know what a good decision looked like, and that’s what the premium is for. In micro1’s own words, the data builds environments where AI models and agents “practice tasks and decisions.”

The criteria tell the same story. Reported minimums run to about twenty employees, several years of history, English, and a handful of countries. Those aren’t volume filters. They’re realism filters. A company under twenty people doesn’t generate many handoffs between roles, and under a few years of history the outcomes of its decisions haven’t resolved. The buyers want the org chart in motion, with results attached.

micro1 is doing this in the open, with a press release and a bank behind it, which is more than most of this market can say. The problem isn’t them. It’s the deal.

The Deal

The published tiers start above $100,000 and run past $1 million for data the buyer considers unique. It’s paid once. You keep ownership of your files, which is true and beside the point, because what you sold was never the files. It was the judgment inside them, extracted somewhere else and dissolved into model weights you will never see, with no royalty, no audit right, and no way to call it back. There’s also a published referral bounty, so anyone with a network can become a broker, which is why the pitch may reach you through someone you’ve never heard of.

Who should take it? A company whose value is its relationships, its brand, or its physical assets can sell the record of its process and lose nothing it was relying on. A company whose value is the process, how it decides, what it knows that competitors don’t, is selling the moat for a check. Most sellers won’t ask which one they are. The check arrives before the question does.

And much of what’s in those folders isn’t the seller’s to sell. Customer records. Employee messages. The other side’s terms. Nearly every contract in the shared drive has a confidentiality clause. Read it before you sell the drive.

Exhaust and Fuel

Here’s the part that matters if your function runs on judgment. Procurement, legal, finance, underwriting, claims, anywhere the work is deciding.

The best-capitalized buyers in the world have decided that the record of how a company decides is worth six figures, cash, up front. Not the documents. The decisions, what was challenged, what was accepted, why, by whom, and what happened next.

They can only buy it one way, because of how it was kept. Judgment exists for one moment, when someone decides. After that it survives in one of two forms. The first is exhaust, the email, the chat, the ticket, the export, the trail the work left behind. The judgment is in there somewhere, and the buyers are betting that a model can dig it out, somewhere else, after the fact.

The second form is fuel, the decision captured at the moment it’s made, with its reason, its context, and later its outcome, held by the organization that made it and read by the next person who faces the same call. Nothing has to be inferred, because nothing was lost.

Most functions run on exhaust without meaning to. The decisions live in threads and in the heads of two people, one of whom is leaving. That’s what makes the record buyable and nearly worthless to its owner at the same time. Exhaust sells cheap because it was never built to be used. Fuel is the same judgment, kept on purpose. If a lab would pay six figures for your exhaust, ask what the fuel would be worth to you.

How We Think About It

We built Revver Contract Intelligence on four ideas, and none of them is specific to contracts.

1. Judgment is captured where it’s made, not reconstructed later.
2. The record belongs to the organization that made it, not to the tool that held it.
3. The data lives apart from the model, so the engine can change but the knowledge stays.
4. Kept that way, the record feeds the next decision. The same facts get the same reasoning, the exceptions carry their limits, and the judgment compounds instead of starting over.

For contracts, that means every review becomes fuel. The finding, the decision, the reason, who made it and on which version, kept in your workspace, never shared with another customer, never sold or licensed, never used to train a model. Frontier models do the reading, under terms that prohibit training on your data. They’re the engine, not the owner.

We’ve been asked whether that could change. Our answer is no. This is where we take our stance. micro1 isn’t the enemy. They’re where this is going. The record of how a company thinks now has a price, the price is paid once, and what’s bought is gone into models the company will never see. We’d rather the record stay where it was made, with the people who made it.

Our bet is the opposite of micro1’s and the others’. The expertise inside a function that runs on judgment is the most valuable thing in it, and the company that keeps it gets smarter every year while the company that sells it gets a check. So we help our customers capture their own expertise, apply it across every function that touches it, and keep it in their hands. The procurement lead who knows which supplier terms matter. The counsel who knows which exceptions are safe. The finance team that knows what a commitment costs. Their judgment becomes something the whole company can use, and it stays theirs. Not sold once. Not dissolved into someone else’s model.

Start with the last ten decisions your team made. Find the record of why. If it’s in an inbox, you don’t own it yet. Someone else may soon offer to.

If you want that record built across your whole contract book, with Revver and our Advisors working the decisions alongside your team, that’s a Portfolio Audit.

Revver Contract Intelligence is an AI-native platform that evaluates a company’s contracts from its own position, turning the judgment behind every decision into institutional knowledge that compounds over time.