On this page
Records that show a task from request to result
A workflow trace is the record of one piece of work: the request, the steps and messages in between, the review or approval, and the result. AI developers use company records like these to build training tasks and RL environments, where an agent practices a workplace task Gizmodo, (opens in a new tab). The value sits in the links between records, so keep the identifiers that join them.
| Record | What it shows of the task | What joins it to the rest |
|---|---|---|
| Tickets and issues | The request, who took it, each status change and how it was closed | Ticket IDs cited in commits, chat threads and documents |
| Code review | A proposed change, the reviewer comments, the rework and the approval | Commit and branch references, ticket IDs in titles |
| Approvals and sign-offs | Who asked, who decided, the decision and the reason given | Request IDs, timestamps and the record that was approved |
| Chat and email threads | How people discussed a problem, asked for help and agreed what to do | Links and IDs pasted into messages, thread timestamps |
| Documents with version history | Drafts, comments and the final version of a plan or spec | Links from tickets and messages, revision timestamps |
| Support conversations | A customer problem, the replies, the escalations and the resolution | Case IDs, masked account references and linked tickets |
Keep failed attempts, reverted changes and escalations in the trace. Reporting on the airline archive bid noted that a training task can be built from work that was done poorly TIME, (opens in a new tab).
Public company-records deals tied to agent training
The public Deal Tracker holds three company records transactions that public reporting ties to AI training. Each row below renders from its record. Open a record to read the value its sources reported and the dates behind it. A reported value belongs to that one transaction: do not turn it into a per-ticket, per-message or per-user rate for another archive.
| Tracker record | What the record shows | Value type |
|---|---|---|
2026-04-cielo24-archive | Reported sale | Reported |
2026-04-simpleclosure-startup-archives | Reported sale | Reported |
2026-08-google-spirit-airlines | Winning bid, pending court approval | Reported |
None of the three records sets a price for a private archive, and none says who will license the next one. For what public records can and cannot prove about a whole archive, read Workplace archive deals.
What de-identification removes and what it keeps
General information, not legal advice.
Reporting on the closed-startup archive sales said the records were processed to remove personally identifiable information before a sale Gizmodo, (opens in a new tab). For agent training, de-identification has two jobs: take out what identifies people or exposes secrets, and keep what shows how the task was done.
Removed before any sample
- Names, email addresses, phone numbers and other direct contact details, replaced with stable role codes
- Credentials, keys, tokens and other secrets
- Customer and client material outside the agreed scope
- Free text that points to one person, such as a rare job title together with a date and a place
Kept, so the trace still shows the work
- The order of the steps, from request to result
- Roles, such as requester, assignee, reviewer and approver
- Timestamps, coarsened where an exact time could point to one person
- The identifiers that join a ticket to its review, its messages and its documents
- Failed attempts, rework and the reason a change was turned down
A stable code in place of a name is pseudonymization, and the result can still be personal data. Do not call a scrubbed sample anonymous unless a review supports that conclusion. Leave out records about children: Rightmark does not accept data about children at launch. If the records include health information, stop here: Rightmark does not take health data yet (patient records, medical imaging or veterinary records), until a HIPAA program and vendor agreements are in place.
Next steps for AI teams and owners
AI teams can describe the task an agent should learn, the records that would show it and the evidence they need in a data request brief, or read the data lots open to AI teams.
Owners can start with the operational data guide to map systems, authority, exclusions and export state. Rightmark Trust says transfers use scoped, revocable connections approved by administrators or exports, and that passwords are never held.
Sources
- Gizmodo: Failed Companies Are Selling Old Slack Chats and Email Archives to Train AI (opens in a new tab)
- Epoch AI: An FAQ on Reinforcement Learning Environments (opens in a new tab)
- TIME: Why Google Bid $10 Million for a Failed Airline’s Data (opens in a new tab)
- Rightmark: Deal Tracker
- Rightmark: Trust