The short answer
What labs pay for is the record of how work gets done: briefs and their changes, rounds of feedback and response, hours logged by task, and the QA steps that decide when something is finished. micro1, Scale AI and Nyne all name project and process records on their own pages. Finished portfolios are not on any list.
On this page
Agencies tend to think their value is the work: the campaign, the identity, the build. To an AI lab the work is the least interesting part. What a lab cannot easily get is the path to it, and agencies have been logging that path for years without thinking of it as an asset.
Why process beats portfolio
The labs buying business data are training models to do multi-step work: take a request, plan it, act, take feedback, revise. A finished logo teaches very little about that. The sequence that produced it (brief, first concepts, feedback, second round, sign-off, hours at each stage) is a worked example of exactly the behavior they want.
That is why the buyers’ published lists read like an operations manual rather than a showreel. Look at the biggest business-data contest of the year: per Reuters in August 2026, Google put $10 million on the table for Spirit Airlines’ internal records (staff email, Teams chats, spreadsheets, calendars), ahead of Mercor’s $7.5 million, pending a judge’s approval. Nobody was bidding for the airline’s ads.
What the buyers name
| Buyer | Records named on its own page (read 7 Nov 2026) |
|---|---|
| micro1 | Process documentation, SOPs, product planning, operations records, internal knowledge bases, sales and CRM records, engineering workflows, bug tracking |
| Scale AI | Jira and Confluence work and knowledge records, documents and files with metadata, email threads, chat histories, calendar events, CRM and HR records |
| Nyne | Asana and Jira project and service records, CRM, accounting records, calendars |
None of the three lists “creative work” or “deliverables”.
The five agency records that matter
Briefs and change requests
The opening intent and every amendment, with dates. A brief that changed three times shows judgment under shifting requirements, which is what a lab wants a model to learn. It is also the record most likely to contain the client’s confidential plans, so it needs the most careful stripping.
Revision rounds
Feedback, the agency’s response and the next version, in order. This is where agencies are unusual: few other businesses hold thousands of documented feedback loops on the same kind of task. The comments themselves are the valuable part.
Timesheets
Hours by person, task and phase, often across a decade. Structured, tabular and generally the agency’s own operating record. Remove staff names, and rewrite task descriptions that name the client.
QA and sign-off
Checklists, bug tickets, approval steps. How a team decides a thing is done is hard to observe from the outside and easy to export from a ticketing tool.
Process documents
SOPs, estimating templates, onboarding guides. These are the least risky records you hold: plainly yours, usually covered by the pre-existing materials clause of your contracts, and directly on micro1’s list.
What to take out before anyone sees a sample
- Client names, product names and anything a reader could trace back to an account
- Staff names and personal details in timesheets, comments and email threads
- Client customers’ personal data of any kind
- Final deliverables and source files the client owns
Scale AI rules out staff and customer identities altogether and does its anonymizing on the seller’s side, before export. The other buyers promise de-identification without saying where it takes place. Run your own pass first either way.
Does the tool matter?
Somewhat. Records in a project management tool with an export (tickets, comments, timestamps, assignees) are easier to scope and de-identify than the same history spread across inboxes and shared drives. Scale AI wants mature systems of record and one named person inside the company to own the project. If your history lives in email, expect more work before a buyer can price it.
How much history is enough?
No buyer sets a minimum number of years of records. micro1 says only that a longer history tends to fit better; Nyne wants the business itself to be three years old. In practice, depth matters more than breadth: five years of consistently logged projects in one tool says more than fifteen years spread across three tools, two inboxes and a shared drive nobody has opened since a migration.
Before you talk to anyone, it is worth an afternoon to map where each record type lives, for which years, and in what format. That map is what a buyer will ask for first, and making it will also tell you how much of the archive survives the contract review.
Packaging records for review
Buyers price what they can inspect. A small, representative set (one project’s brief, revision rounds, timesheet and sign-off, fully de-identified) does more to start a serious conversation than a description of the whole archive. Keep the scope of any sample tight, and agree in writing what the buyer may do with it before it leaves your hands.
Where to go next
For the full picture of ownership, buyers and money, the front page on how to sell agency data brings it together. To see each buyer’s terms side by side, read buyers compared.
The founders who do well here are the ones who kept their own records tidy for their own reasons long before anyone offered to pay for them.