AI labs, and the data companies that supply them, want records that show how real work gets done: the request, the discussion, the decision, the fix and the result, linked across the tools your team already uses. A folder of unrelated files, your public website or your customer list is worth far less.
Public web pages, books and open-source code are available to anyone training a model. What is hard to find is the inside of a working company: how a request turned into a finished job, and every step and argument in between.
Linked records from your business systems that show a task from start to finish, with the steps, the decisions, the corrections and the outcome intact. The buyer de-identifies it before it is used, so the people drop out and the pattern of the work remains.
A customer reports that invoices export with the wrong tax rate.
Support and engineering argue: bug or settings issue? Someone posts the root cause.
The fix is scoped, prioritized and assigned, with acceptance criteria written down.
The code change, the reviewer's comments and the revision that passed review.
The explanation sent back, and the ticket closed as resolved.
Illustrative example, not a real company.
Each piece is ordinary on its own. Linked by ticket numbers, issue keys and dates, they show a model how a problem moves through a business, who weighs in, what was tried first and what finally worked.
A folder of unrelated PDFs shows finished outputs with no sequence and no reasons. A model can read them, but it cannot learn from them how the work was done. That gap is what buyers notice first.
micro1's data partnerships page names the categories below, as published, checked 7 October 2026. Read them as signals of what a model needs to learn from a company, not as a shopping list where every item pays.
The written rule for how a task should be done. Next to real tickets and emails, it shows the gap between the rule and the practice.
The reference material staff use to answer questions: wikis, playbooks, how-to pages in Confluence or Notion.
Pipeline notes, call summaries and the reasons deals were won or lost. The reasoning matters; the contact list does not.
A project from kickoff to delivery, including scope changes, delays, handoffs and the final sign-off.
Reviews, checklists, rejected work and corrections. Records of what "not good enough" looked like are hard to find anywhere else.
How and why people choose: approvals with comments, trade-offs argued in threads, memos that explain a call.
Human feedback on AI outputs: staff rating, correcting or rejecting what an AI tool produced in their real work.
Most of these are stronger together. An SOP plus the tickets that followed it says more than either one alone.
Eligibility lists several years of records the company owns, 20+ full-time US office employees (accounting firms 10+, law firms 6+), with US-based teams the strongest fit. It buys "an agreed copy"; originals stay with the company.
30+ employees, mature operations, documented processes, modern software tools, primarily English; US prioritized, then other Western markets. "Documented processes" and "modern software tools" mean your work leaves an exportable trail.
Open to any vertical, and it also pays individual professionals for expertise. Expertise matters here: knowing the right answer in a field is part of what is being bought.
These are the sources named on the buyers' own pages. Each group captures a different part of the work. The value comes from overlap between them.
Shows the discussion. Who asked, who objected, what was agreed and which version got signed.
Shows the task. Status changes, assignments, escalations and how each item was resolved, all with timestamps.
Shows the commercial reasoning. Deals won and lost, month-end closes, reconciliations and approvals.
Shows the artifact. Drawings and designs with their revisions, and field jobs from dispatch to invoice.
Shows the change. Repositories with history: commits, reviews and issues, not just the final code.
All of these are systems of record with authors, timestamps and links. A company that has used three or four of them consistently for years usually has more to offer than one with a larger archive in a single shared drive. Chat is covered in Slack and Teams messages; written procedures in documents, SOPs and knowledge bases.
Practitioners describe a ladder. The same company records can sit on any step. What changes is how much work is built on top of them.
An agreed copy of your records, exported and de-identified. This is where almost every company starts: the buyer receives the history and does the rest.
Tasks with known correct answers, built from your data, used to test whether a model does the job right. They need people who know the right answer. Human feedback on AI outputs sits close to this tier.
A simulated workplace where a model practices tasks and is scored. Practitioners put these in the 6 to 8 figure range, but they need heavy engineering.
Figures are practitioners' estimates, not offers. For the published program ranges and the reported deal prices, see how much AI companies pay for data. For when a company should consider helping build evaluations, see raw data vs. evaluations vs. environments.
No one can price your data without seeing it. These are the traits that move records up or down a buyer's list.
The point about personal information follows from what buyers publish about privacy. micro1's page says the scope is agreed in writing, sensitive and confidential information is scrubbed, originals are deleted after processing, no customer information is exposed, and the company keeps ownership of its underlying data. Mode says it de-identifies before onward delivery. Both as published, checked 7 October 2026. If most of a dataset is names, addresses and account numbers, removing them leaves very little behind. The work patterns around customers are what survive de-identification.
Three things a model can only learn from linked records.
The order of steps: what was checked first, what was escalated, what waited on whom. A single document has no order. A ticket with its thread and its fix does.
Whether the work succeeded. Was the fix accepted, did the client approve the revision, did the invoice get paid? Without the outcome, a model cannot tell good work from bad.
Why a choice was made. "We went with vendor B because A could not ship by the deadline" is exactly the kind of decision-making pattern buyers name.
Gaps break the chain. Short retention settings on chat or email, a move from one ticketing tool to another that dropped the links, or work done in personal accounts all leave holes in the history. A buyer's review will find them. Check your own retention and export options before you describe your data to anyone, so the scope you state matches what you can actually deliver.
If your records hold client secrets (law, accounting, M&A, agencies), patient information or customer financial data, value is not the first test. Your client contracts, professional rules, attorney-client privilege and laws such as HIPAA, GLBA, GDPR and CCPA/CPRA come first, and some of this material can never be included. Read is it legal to sell company data before you scope anything.
General information, not legal advice. Talk to your own lawyer before you sign.
Headcount and storage size are easy to measure, so owners tend to lead with them. Buyers look past both to the shape of the records.
Fictional examples for illustration only. Not real companies, not an offer, and no price is implied.
Both fictional companies sit above the published headcount minimums (micro1 lists 30+ employees; Mode lists 20+ full-time US office employees), so size is not what separates them. The records are.
Practitioners advise never sending a full dataset before price. Share a manifest and a few samples, and get more than one offer. A manifest describes the data; it does not contain it.
| Field | What to write | Example (fictional) |
|---|---|---|
| System | Each tool the records come from, and the plan or edition you are on. | Jira Cloud, Slack, Figma |
| Years | The date range covered, and any gaps or deleted periods. | 2021 to 2026; Slack before 2023 deleted |
| Volume | Counts of items, not only gigabytes: tickets, threads, files, commits. | Ticket and thread counts per year, from admin reports |
| Linkage | How records connect across systems, and how consistently. | Job code in Jira key, channel name and file title |
| Language | The main working language and any notable share in others. | Primarily English; some Spanish client threads |
| Exclusions | What is out of scope from the start. | HR channels, direct messages, legal matters, one client under NDA |
| Ownership | Who owns the records and which contracts limit their use. | Working files company-owned; deliverables owned by clients |
| Export | Whether you have admin rights and an export path today. | Admin export available on current plans |
Illustrative fields and fictional examples, not a buyer's required format.
A good manifest answers the questions on this page before a buyer asks them: how connected the records are, how far back they go, what is excluded and who owns what. It also protects you. Writing the exclusions and ownership lines forces the internal conversations, with partners, IT and whoever manages client contracts, that you would otherwise have under deadline pressure later.
Tick what is true for your company. Nothing is sent or stored; this runs only in your browser and it is not a valuation.
If most boxes are ticked, the next question is which programs' published rules you meet on size, country and industry. The eligibility checker applies those rules.
Connected records of real work. Buyers want to see a task from request to result: the ticket, the discussion, the decision, the change and the reply, linked across systems and kept for years. micro1's page lists SOPs, knowledge bases, internal documentation, CRM data, project histories and QA processes, plus "decision-making patterns" and "AI performance feedback" (as published, checked 7 October 2026).
Usually not on its own. Buyers say they strip personal and customer details before the data is used, so a dataset that is mostly names, emails and phone numbers shrinks to very little. The reasoning around customers, such as deal notes, call summaries and support resolutions, is what carries value.
It is worth little. Anyone can collect public content, so it adds almost nothing a model trainer cannot already reach. Buyers ask for internal records that never appeared online: how work was planned, argued over, reviewed and finished.
None of the programs we checked publish an exact number of years. Mode lists several years of records the company owns in its eligibility, and micro1 lists mature operations and documented processes (both as published, checked 7 October 2026). A long, continuous history is generally more useful than one busy quarter.
Practitioners describe three tiers. Raw data, meaning exported and de-identified records, is the cheapest. Evaluations built on that data, which are tasks with known correct answers used to test a model, are worth roughly 10 times raw. Full training environments, where a model practices tasks and is scored, can reach 6 to 8 figures but need heavy engineering.
Practitioners advise never sending a full dataset before price. Share a manifest that describes the systems, years covered, volume, how records link, language, exclusions and ownership, plus a few de-identified samples, and try to get more than one offer.
Not automatically. Published rules set minimum sizes (micro1 lists 30+ employees; Mode lists 20+ full-time US office employees, with 10+ for accounting firms and 6+ for law firms), but above that, linked, continuous and documented records matter more than headcount or storage size. A smaller team with a connected history can look stronger than a larger firm with a scattered shared drive.
Each program reviews its own applicants and decides scope, price and acceptance. We cannot tell you what a buyer will pay or whether it will take your data. Start with the published rules, then apply where you fit.
Independent site. Some links are referral links: if your company signs with a buyer through them, the buyer may pay us a fee. You are not charged, and we never see your data.