micro1 now has $1 billion to spend buying your company's operating manuals, CRM records and decision history, and the offer will probably reach a business-unit head before it reaches your general counsel. The money is real: its program page lists tiers of $100K+, $500K+ and $1M+ per company. What the public terms leave out is who ends up training on the data, whether you can sell it to anyone else, and what happens if it leaks. A lot of what micro1 asks for, starting with CRM records and customer contracts, is data your customers and vendors have contractual claims on. Before anyone answers the outreach, someone needs to own three things: an inventory, a rights check and a contract checklist.
What micro1 Announced on October 9
micro1 committed on October 9, 2026 to spend $1 billion over the next 12 months acquiring and licensing enterprise operational data through its Company Data Partnerships program, with Citi and Hercules Capital providing the capital. The company turns "de-identified" company data into reinforcement learning environments, simulated versions of real business operations where AI models and agents practice workflows and decisions. A reinforcement learning environment, in this sense, is a sandbox that replays your processes so a model can be scored on doing them.
TNW's report quotes micro1's pitch: "Real business operations involve incomplete information, competing priorities and exceptions that require judgment." In your company, that means the exception queue and the escalation path.
The buyer is well funded. Forbes Australia reported in September a raise of more than $100 million at a $4 billion valuation, citing two people familiar with the deal, and a $500 million gross annual run rate per TechCrunch. It names Microsoft, Amazon and the robotics company 1X among customers, alongside unnamed frontier labs. micro1 declined to comment on those figures.
What the Program Pays For, and Who Qualifies
micro1 wants the documents that describe how your company actually runs. The Data Partnerships page lists SOPs, knowledge bases, internal documentation, CRM data, project histories and QA processes, plus "decision-making patterns" and human feedback on AI outputs. Its example domains include finance and accounting, customer support, sales and CRM, legal and contracts, and operations and logistics.
The eligibility bar is low enough to cover most mid-market firms:
- "30+ employees" and established documentation.
- US companies first, then other Western markets, with demand strongest in the US, UK and Canada.
- Documentation in English.
Pay depends on dataset size, workflow complexity, quality, domain expertise and uniqueness. The headline tiers start at $100K, but the illustrative figure on micro1's payout estimator reads $50,284 to $92,776, and the page says the final number depends on "data sources, data size, data quality, access terms, and due diligence." A July 2026 micro1 forum session put the range at $100K to $2M+ across 50 partner companies.
Then there is the referral program. The partnerships page offers "up to $50k" to whoever refers a company, paid once the referred company hits revenue milestones with micro1. A consultant, a former employee or your own sales operations lead now has a cash reason to start the conversation, and nothing in the program requires that person to have authority over the data.
What the Public Terms Leave Out
The public terms are thin on the clauses that set your risk. Here is what micro1 does say:
- "Companies retain ownership of their underlying data."
- PII is removed "when relevant," and some datasets get "synthetic rewrites" that keep the workflow structure while loosening ties to the original records.
- Companies can review representative samples before prepared data is used.
- Data is kept for agreed periods and deleted on request or at the end of the engagement.
- Scope, security standards, redaction requirements and internal approvals are defined before participation.
None of it names the downstream buyer. The page says the data helps "train frontier AI models" and builds evaluations "for AI Labs," which tells you the category and not the counterparty. It says nothing about exclusivity, so you cannot tell from the public material whether licensing to micro1 bars you from licensing the same data to anyone else, or whether micro1 can resell access. It says nothing on liability, indemnity or a cap. The phrase "retain ownership" settles very little when the license grant carries the real terms, and the grant is not published.
These gaps may well be filled in the negotiated agreement. They are still the clauses your counsel has to see in writing before anything leaves the building.
Why Much of This Data May Not Be Yours to License
Owning a file and holding the right to license what is in it are separate questions, and CRM data fails the second one most often. Three sources of constraint apply to the categories micro1 lists.
Start with your customers' contracts. CRM records hold customer names, deal terms, pricing, support transcripts and the customer's own confidential information. Most MSAs and NDAs limit how you can use and disclose that information, often to performing the contract. Licensing it to a data lab for model training is a use your customer never agreed to, and de-identification does not automatically cure a confidentiality clause that covers the substance of a deal.
Privacy law comes next. Under GDPR, personal data must be "collected for specified, explicit and legitimate purposes" and "not further processed in a manner that is incompatible with those purposes". Commercial AI training is a new purpose for a sales contact you collected to sell to. In California, the CCPA's exemptions for employee and business-to-business personal information expired on December 31, 2022, so the B2B contacts and employee data in your systems carry consumer-grade rights, including the opt-out of sale.
Your vendors' licenses are the third constraint. Knowledge bases routinely contain material you do not own: analyst reports, licensed training content, vendor runbooks, software documentation under NDA. Those licenses rarely permit redistribution, let alone use as training data.
"PII removed when relevant" is a weak control on its own. A 2019 study in Nature Communications found that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes. Project histories and decision records are rich in exactly that kind of quasi-identifier: which account, which region, which quarter, which manager.
What Spirit Airlines Showed About Selling Operational Data
The closest public precedent is micro1's own bid for an airline's records, and it shows where the pressure lands. In Spirit Airlines' bankruptcy, Google's $10 million offer won the auction for Spirit's corporate data, and micro1 came back on August 19 with a late $12.5 million bid. Mercor sat as backup at $7.5 million, kept because its version used a third-party de-identification process. Ali Ansari, micro1's CEO, called Google's price "actually quite low" for decades of airline operations.
The package was the kind of material micro1's program now buys from going concerns: about 100 million emails, 500 million Microsoft Teams messages, 30 million lines of code across 516 repositories, pricing systems and project records. It also included, per Bloomberg Government, more than 175,000 employee records dating back to 1986 and 3.4 million payroll records. The Association of Flight Attendants-CWA objected that a pseudonymized dataset could still reveal which crew bases filed grievances and which employees were under investigation. Robert Drain, a former bankruptcy judge, said it was the first sale he had seen of "information for information's sake" and that consumer privacy rules may not reach employee records at all.
The court-appointed privacy ombudsman has since backed Google's deal. A solvent company signing a data partnership gets fewer safeguards than Spirit's estate had. A bankruptcy sale has a judge, an ombudsman and objecting creditors. A data partnership signed by a division head has whoever happens to read the contract.
The market has priced this kind of data before. Reddit's 2024 training-data license with Google was reported at about $60 million a year, and neither company disclosed the terms. Reddit negotiated with a public company's legal team behind it. A $100K offer for your operations data is an opening position, and you should negotiate it like one.
The Case for Saying Yes
The strongest argument for the program is that the money comes from assets you already paid to create. A 60-person services firm with a clean Confluence space and a decade of project retrospectives has never had a buyer for that material, and a six-figure check for documentation is real margin. micro1's stated controls (scope defined up front, sample review, synthetic rewrites, deletion at the end) go further than many SaaS vendors offer on data you hand them every day. micro1 sells into the same frontier-lab training-data market as vendors such as Scale AI, and Forbes reports its customers include Microsoft and Amazon, so the buyer has paying customers of its own.
The case holds for a defined slice: internal process documents you wrote yourself, containing no customer or employee detail, under a contract that names the use and bans re-identification. It weakens fast once CRM exports, contract repositories and support transcripts enter the scope, because the rights there belong partly to someone else.
What to Do Before Anyone Answers the Outreach
Decide who owns the answer before the first email arrives, because the referral rewards mean it is coming.
This Week:
- Name one owner for third-party data licensing requests, usually the CDO or general counsel, and send a two-line note to business-unit heads: no data-sharing conversation with a training-data buyer without that owner in the room.
- Ask procurement and sales operations whether anyone has already been contacted by micro1 or a referrer. Check inbound forms and partner channels too.
This Month:
- Build a three-column inventory of the categories micro1 lists (SOPs, knowledge bases, CRM, project histories, QA records, decision logs): who created it, whose information is inside it, and which contract or policy governs it. Anything containing customer confidential information, employee records or third-party licensed content goes in a "not licensable without consent" column.
- Have counsel pull your standard MSA and NDA confidentiality clauses and your privacy notices, and answer one question in writing: does any of it permit licensing derived data for AI training?
- If you still want to proceed, draft your own term sheet before you take theirs. It should name every downstream licensee or require notice before transfer; set exclusivity explicitly; prohibit re-identification and resale; cover evaluation as well as training; set deletion with a certificate of destruction; include an audit right; and carry an indemnity sized to a data incident, not to the fee.
Before Signing:
- Run the sample review micro1 offers against your own re-identification test: give the de-identified sample to someone who knows the business and see how many accounts, people or deals they can name.
- Get the license grant, not the ownership statement, approved by counsel and the data owner.
The Bottom Line
The last decade's data deals ran through companies with public data and large legal teams, such as Reddit and Spirit's bankruptcy estate. micro1's program brings the same transaction to companies with 30 employees and a documentation wiki, and it pays people to make the introduction. The data micro1 most wants, records of how customers buy and how staff decide, is the data where your customers, employees and vendors hold rights alongside you.
Put an owner and a rights inventory in place this month, and treat the first offer as the start of a license negotiation.
Continue Reading
- Training Data Provenance: Indemnity Stops at Your Fine-Tuning Data
- The DOE Wants Your Lab Data. You Write the Contract.
- Reddit's API Shutdown Spares Sprinklr, Not Your Python Scripts
- AI Vendor Exit Clauses: Standard Terms Give 0 to 90 Days to Get Out
- UK ICO Asks AI Agent Deployers for Evidence by November 20
- Fable 5 Needs Retention On. ZDR Was Never Zero.