Executive Summary
This article reads the $1 billion enterprise data purchase that the AI data company micro1 announced on 9 October through a single question. Does the seller hold the right to sell? The target is not text lying around on the web but the record a company builds while it works. Work standards and in-house knowledge bases, internal documents, CRM records, project histories, quality checks, and the traces a team leaves of what it decided and why are all on the list.
The window is the next 12 months, and Citi and Hercules Capital are funding it. The payments fall into three tiers. The phrase micro1 writes into its participation requirements, though, is "internal approval processes and permissions", and nothing in it asks a company to go and collect consent from the third parties tangled up in those records, customers among them. Establishing the right to sell stays on the selling side.
Everything to this point is written in the press release, in micro1's own program page and in the reporting that worked through them. The diagram in Section 3, which lays three parties over a single row of CRM data, and the reading in Section 5 that takes the announcement as the point where a price was attached to a conversation about quality standards, are this article's own interpretation.
Key numbers
Four cells. The first two are the figures micro1 put up. The third is the gap between those figures and the payout its own tool illustrates. The fourth shows how far the promise to strip personal data actually reaches.
Sources: TheNextWeb (9 October 2026), Beri.net analysis.
$1 billion
Committed over 12 months
What micro1 says it will spend buying or licensing enterprise operational data. Citi and Hercules Capital are funding it
$1 million
Where the top tier starts
The floor placed on highly unique, proprietary operational data. A basic partnership starts at $100K, and large multi-team datasets at $500K
$50K–$93K
What micro1's estimator shows
The illustrative figure on micro1's own payout estimator reads $50,284 to $92,776, short of even the lowest $100K tier
99.98%
Identities recovered from 15 attributes
The figure a 2019 Nature Communications study reports. Stripping personal identifiers out of a file does not amount to anonymity
micro1's $1 Billion Commitment on 9 October
micro1 builds and gathers the data used to teach AI models and agents. Until early 2025 it ran an AI recruiting service with about $7 million in annual revenue, then turned toward data, and it now gets named as a rival to Scale AI. On 9 October it announced a purchasing program called Company Data Partnerships, under which it will put $1 billion into buying or licensing enterprise operational data over the next 12 months. Citi and Hercules Capital are funding it.
The shopping list is specific. Standard operating procedures and internal knowledge bases, internal documents, CRM data, project histories, quality assurance processes, records of team decision-making, and the feedback people attach to what an AI produces. Finance and accounting, customer support, sales, legal and logistics are the kinds of work it gives as examples, all of them areas where the day-to-day practice is concrete. With the supply of web documents a model can still learn from running down, the new part of this announcement is that the suppliers being called on are ordinary working companies rather than companies that hold content, the way Reddit or Stack Overflow do.
The size of the sum shows up against the deals content holders have signed. Paying for training data is itself a pattern that has been running since 2024. Google's agreement with Reddit is reported at around $60 million a year, and a separate OpenAI agreement is said to be of a similar size. The $1 billion micro1 has put against 12 months would take more than ten deals of that size to reach. That is money you cannot spend by signing a few large agreements with content holders, which is why the supply has to come from many sources at once.
Training-data licensing deals, compared by annualized size
The sources are worth separating once here. What the press release itself states runs to the amount, the window, the de-identification, the funding and the name of the program. The price tiers, the eligibility requirements and the ownership sentence in Section 2 below come not from the press release but from the program page micro1 maintains, and secondary coverage carried them across. Reading this as a press release that announced a price list gets the facts wrong.
What micro1 Buys, and at What Price
Payment splits into three bands according to the scale and the scarcity of the data. The table below sets out the bands micro1 published and the condition attached to each.
| Band | Amount | Condition |
|---|---|---|
| Basic | $100K and above | Qualified enterprise data partnerships |
| Expanded | $500K and above | Large-scale datasets, or several teams taking part on an ongoing basis |
| Proprietary | $1M and above | Highly unique, proprietary operational data that is hard to find elsewhere |
Read on its own, the table looks like a guarantee of at least $100K. The illustrative figure on the payout estimator micro1 put up on the same page, though, runs from $50,284 to $92,776, short of the lowest band in the table. The company itself writes that the real amount moves with the sources of the data, its size, its quality, the access terms and the result of due diligence. The three bands are not what you will receive. They are the thresholds attached to a partnership that qualifies, and whether yours reaches that mark is settled once due diligence is over.
Nor is participation open to everyone. A company needs more than 30 employees and documentation with some structure to it, US companies come first and other Western markets after that, and the material has to be in English. Read from a Korean company's side, these requirements are somebody else's business for now, but the direction they point in is plain. Only a company whose everyday work survives as documents, and whose documents are organized well enough to be found, can get in line.
Ownership and the Right to Sell Are Not the Same Thing
micro1 has an answer ready for the first worry that surfaces when someone is asked to hand data over: companies retain ownership of their underlying data, personally identifiable information is removed when relevant, certain datasets go through synthetic rewrites meant to loosen the direct ties to the original records, a representative sample can be reviewed before taking part, and on request or at the end of an engagement the data is deleted.
Pull a single word out of that explanation and the shape of the problem appears. Ownership settles whose file it is. A license right settles whether you can pass it to somebody else to use. The two do not sit on the same level. A box in your own warehouse is still not yours to sell if the goods inside belong to someone else.
One row of CRM data makes that difference sharp. In the row that records a single deal, at least three parties overlap.
The promise to remove personal data does not close the lower two lines either. A 2019 study in Nature Communications reported that 15 demographic attributes are enough to correctly re-identify 99.98% of Americans. Project histories commonly carry attributes of that kind, such as job role, region, affiliation and duration. Synthetic rewriting, on micro1's own account, preserves the structure of the workflow. What goes is the name, and what stays is the shape of how that company works. Whether to read this as a risk or as the thing that makes the data valuable will divide people, but the identity of the goods on sale is the same either way.
The Checking Falls to the Seller
So who lines those three rights up against each other? The expression micro1 writes into its participation requirements is "internal approval processes and permissions". It takes as given that a participating company already holds the approval and the permission to hand that data over. Read the other way, micro1 does not itself demand consent from customers or trading partners. The platform only checks the paperwork at the door, and whether that paperwork is actually valid is judged inside the company selling.
What actually trips a company up has already been set out in some detail. These are the four things the Beri.net analysis picks out.
- • Confidentiality clauses in master service agreements and NDAs — a CRM record holding a client's deal terms is usually tied to use for performance of the contract. Licensing it for AI training sits outside that scope
- • The GDPR purpose limitation principle — personal data gathered in order to sell something needs a separate legal basis before it can be turned toward training
- • Expiry of the CCPA B2B exemption — the carve-out that kept employee and business-contact details outside consumer rights ended on 31 December 2022. Access and deletion requests now reach the contact details sitting in a CRM
- • No-redistribution terms in vendor licenses — an internal knowledge base has training material, software documentation and vendor operating guides from outside mixed into it. Handing the whole thing over touches those terms
Some things are missing from the contract as well. The same analysis points out that micro1 has not named the end buyer, that there is no published wording on whether the data will be used exclusively, and no published clause setting liability or indemnity if something goes wrong. A referral program sits alongside all this. Whoever introduces a company gets up to $50K, payable once the referred company reaches certain revenue milestones with micro1. This is where the incentive comes from to get things moving inside a company before the legal review is finished.
A deal of this kind is not a first. Pebblous has written separately about the corporate data of Spirit Airlines, which stopped flying this May after a second Chapter 11, going to auction and Google winning it at $10 million. micro1 put in a $12.5 million bid of its own after the deadline had passed. The legal channels of the two deals differ. A bankruptcy sale had a court looking at it and a place for the union to object. A program a functioning company signs up to voluntarily has no outside eye of that kind on it. Since no institution does the review for you, the deal where lining the conditions up is harder is the voluntary one.
Why Pebblous Is Watching micro1's Price List
Preparing data so that AI can use it has until now been discussed in the language of quality standards. How much is missing, whether the labels hold together, whether the schema lines up. What micro1's announcement changes is that a figure is now attached to that discussion. If the line between $100K and $1 million is drawn at whether the data is unique and proprietary, then the work records a company holds now get valued by price as well as by a quality score.
Due diligence follows anything with a price on it. And the first thing data due diligence asks about is not quality but provenance. Where this record came from, under what promise it was received, how far it was agreed it could be used. Data that cannot answer those three questions cannot be brought to a deal however good its quality is. The same questions remain for a company with no intention of selling. Data you cannot pass outside runs into the same wall when you feed it to an AI inside the building.
So the question of whether a company's work records are an asset or a liability turns out to be a question about how they were recorded rather than a question for the books. Where the terms of use and the scope of consent are written down beside the data, that data is an asset you can reach for. Where only the content has piled up and the terms live in somebody's memory, it becomes a burden that has to go through legal every time it is used. What decides which of the two a given file becomes is not the volume of the data but what is written down next to it.
Thank you for reading this far. The report covering micro1's announcement is at TheNextWeb, and the analysis going through the contractual and regulatory issues is at Beri.net. If your company handed over the CRM records it has built up over the past three years, how many people inside it could sign off on that decision? We would like to hear how you answer it.
References
Official announcements
- 1.micro1. (2026). "micro1 Announces $1 Billion Enterprise AI Data Commitment." Business Wire, October 9, 2026.
- 2.micro1. "Company Data Partnerships." micro1.ai program page.
Reporting and analysis
- 3.Constantin, A. M. (2026). "micro1 commits $1bn to buy company data for training AI agents." TheNextWeb, October 9, 2026.
- 4.Beri, R. (2026). "micro1 Offers $100K+ for CRM Records That May Not Be Yours to Sell." Beri.net, October 10, 2026.
Academic research
- 5.Rocher, L., Hendrickx, J. M., & de Montjoye, Y.-A. (2019). "Estimating the success of re-identifications in incomplete datasets using generative models." Nature Communications, 10, 3069.