Besides a one-time copy of its records, a company can license five newer forms: evaluation sets written and graded by its own senior staff, tasks and environments seeded from its real work, its staff's time and judgment in company context, logs of humans correcting AI output, and recurring refreshes of any of these. Public evidence prices tasks and environments; evaluation sets, corrections logs and refresh licences have no published price yet. All of them need the same ownership and confidentiality checks as the records themselves.
- Buyers increasingly pay for tasks with a checkable answer, not only for documents.
- Interviewed builders put typical task prices at $200 to $2,000, with exclusive deals 4 to 5 times dearer.
- Tasks lose training value as models learn them, which is why refreshes can recur.
- An evaluation set needs a grader or rubric written by people who know the right answer.
- Client work, employee notes and regulated data need the same checks as an archive.
What can a company license besides its records?
An archive licence sells a copy of what already exists. The five forms below are built from the same company, but each adds something an archive lacks: a known right answer, a working environment, a person's judgment, a record of what a model got wrong, or a fresh delivery next quarter.
| Form | What the company provides | Public price evidence | Related guide |
|---|---|---|---|
| Expert evaluation sets | Real cases from its own work, with the answer and the grading rubric written by its senior staff | None for company-built sets; benchmarks built this way are published | Data for AI evaluation |
| Tasks and RL environments | Tasks seeded from real work, with a grader, and sometimes the systems and data to run them in | Yes: interview-based ranges per task and per contract | RL environments from company data |
| Staff time in company context | Experienced staff writing tasks, reviewing outputs or recording how they work, on the company's own systems | No price paid published; company offer ranges only | Workflow data, screen recordings |
| Corrections and human-in-the-loop logs | Records of staff approving, editing or rejecting AI output, with the final version | None found | Section below |
| Recurring refresh licences | New cases, runs or tasks delivered on a schedule, including fresh seed data for simulations | No price paid published; company offers only | Agent traces, pricing models |
Why are buyers paying for tasks and evaluations?
Training with reinforcement learning needs tasks a model can attempt and a grader that scores the attempt. Epoch AI, after interviewing 18 people at environment startups and labs, describes each task as "a prompt instructing the model to achieve an objective, and a grader that determines whether (or to what extent) the objective was met", and reports that enterprise workflows are "a major growth area" (Epoch AI).
Buyers want those tasks grounded in real companies. A Scale AI job post describes RL environments "built from the operating data of the companies that actually hold it" (Scale AI job post). Researchers who built an evaluation from "paid OR engagements totaling USD 3.68 million" found that across 11 frontier models "none matches" the expert reference on deployment-adjusted quality (OR-License, arXiv 2610.24087). Real work with a known good answer is what makes such tests hard.
What has been published about prices?
Each row is public evidence read on 14 November 2026, attributed to who stated it. Ranges are ceilings or typical figures as stated, not a price any company would receive.
| Who states it | What they say | Grade |
|---|---|---|
| Epoch AI (interviews) | "multiple interviewees agreed on a ballpark of $200-$2,000 per task"; "$20k per task would be rare but possible" (Epoch AI) | Independent research based on interviews |
| Epoch AI (interviews) | Contract sizes "are often six to seven figures per quarter"; "exclusive deals are roughly 4-5x more expensive than non-exclusive ones" (Epoch AI) | Independent research based on interviews |
What is not yet priced: no public price was found for an evaluation set built by a company's own staff, for a log of human corrections to AI output, or for a recurring refresh of company records or simulation seed data. Anyone quoting a figure for those is quoting an offer or an estimate.
Expert evaluation sets built by your own senior staff
An evaluation set is a group of real cases with the right answer and a way to grade an attempt. Benchmarks built this way are the hardest published: DAYJOB used "130 tasks built by professionals in healthcare (50) and finance (80)", each with "an expert rubric of binary criteria", and the strongest model passed "24.7% of healthcare and 23.9% of finance attempts" (DAYJOB, arXiv 2610.01306).
A company is well placed to build one when its senior staff can say what a correct answer looks like and why: an estimator who knows what the final cost was, an underwriter who knows how the loan performed, an engineer who knows which change fixed the defect. Buyers usually want the set kept private, so it is not published or reused for training, which keeps it useful for longer. The evaluation data guide covers evaluation-only licences and holdouts.
Corrections and human-in-the-loop logs
Where AI already drafts work inside a company, staff approve, edit or reject what it produces. A record of each correction, with the AI draft, the final version and who changed it, shows exactly where the model was wrong. Human review is still how most deployed agents are checked: a survey of practitioners behind 86 deployed systems found "74% depend primarily on human evaluation" (Measuring Agents in Production, arXiv 2512.04123).
In Origin's view these logs will matter more as more work passes through AI tools, but no public buyer price for them was found. Two checks come first: whether the AI tool's terms let you reuse its outputs to develop other models (see the provider clauses in the agent traces guide), and whether the drafts contain customer or patient content.
Refresh licences and simulation seed data
Tasks wear out. Epoch AI's interviewees describe discarding a task once a model passes it about 70% of the time: "Once a task reaches around 70% pass rate, you might discard it and move on to harder tasks" (Epoch AI). A company that keeps producing new cases, runs or tasks can license them on a schedule rather than once, and a recurring licence suits records that keep growing, such as closed jobs, claims or agent runs.
Simulation seed data is a related form: real starting conditions (an inventory position, a schedule, a deal room, a plant's operating limits) that a builder uses to make a simulated environment realistic, refreshed as the business changes. No public price for it was found. Price structures for recurring licences, revenue shares and subscriptions are compared in data licensing pricing models, and exclusivity in exclusive versus non-exclusive licences.
Rights and staff: what to settle first
The newer forms raise the same questions as an archive, plus two of their own.
- Client work: a task seeded from a client matter, claim or project carries that client's confidentiality terms. Use your own work, or get written consent.
- Regulated data: patient, financial or privileged content in a case stays regulated when it becomes a task or an evaluation. De-identify first; see de-identifying company data.
- Staff time: decide in writing who does the work, in or outside working hours, how it is paid, and that employees know how their notes and recordings will be used.
- Model terms: corrections and drafts produced by a hosted AI model can fall under the provider's restrictions on developing competing models.
- Exclusivity: an exclusive licence can stop you licensing the same tasks or cases to anyone else. Read the scope before agreeing to it.
What happens next: three steps
Nothing has to be built or sent for a first conversation, and you can stop after any step.
- Describe your company and records: which work your senior staff could grade, whether AI already drafts any of it, and how often new cases arise.
- Rights check: Origin works through client contracts, model terms, privacy and staff arrangements with you and your counsel before anything is built or shared.
- Buyer review and terms: Origin works with buyers, the AI developers and data companies that license records and tasks like these, and shares a description only with buyers you approve. You approve scope, use, price, delivery and any refresh schedule, and sign the licence yourself.
- Origin does not charge data owners; a buyer may pay Origin when a deal closes (how Origin is paid).
Related guides by record type
Sector guides where these forms come up most:
- Patient records and imaging data for practices
- Cybersecurity incident data for MSSPs and IR firms
- Linked company records of a 50+ staff firm
- Where AI still fails at professional work
- Which AI training data holds its value
- AI training data price index
Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.
Common questions
Can our senior staff build an AI evaluation set from our own work?
Yes, if the cases are your own or cleared with clients and the staff can write down the right answer and how to grade an attempt. Buyers usually want the set kept private. No public price for company-built evaluation sets had been published as of 14 November 2026.
What is a golden dataset in AI evaluation?
A set of cases whose correct answers have been checked by people who know the work, used as the reference that model outputs are scored against. It is only useful while it stays out of training data, which is why buyers ask for evaluation sets to be kept private.
Are logs of staff correcting AI output worth licensing?
Possibly, but no public buyer price for them was found as of 14 November 2026. They record where a model was wrong and what the right version was. Check the AI tool's terms on reusing outputs and remove customer or patient content first.
Why would an AI buyer pay again for the same kind of data?
Because tasks and evaluations wear out as models learn them. Interviewees told Epoch AI that a task is often dropped once models pass it about 70% of the time, so fresh cases from ongoing work can be licensed on a schedule.
What does an AI training task cost?
Builders interviewed by Epoch AI (January 2026) put most tasks at $200 to $2,000, with $20,000 rare, and said exclusive deals cost roughly 4 to 5 times non-exclusive ones. These are figures for tasks bought from builders, not prices paid to a company for its records.
Sources and further reading
- Epoch AI: an FAQ on reinforcement learning environments (12 January 2026)
- Scale AI job post: Data Acquisition Lead, Frontier Environments (read 14 November 2026)
- OR-License: evaluating LLM agents as operations research practitioners (arXiv 2610.24087)
- DAYJOB: a benchmark for long-horizon professional work (arXiv 2610.01306)
- Measuring Agents in Production (arXiv 2512.04123)