As of November 2026, AI training data gains value when it records real work that is not on the public web: decisions linked to their outcomes, long specialist reasoning, cases worked to a resolution, records with a known right answer, several linked systems, and a fresh supply over time. Scarce domains such as chip and lab run records, lending and claims decisions, audit workflows, imaging with reports and expert screen recordings lead. Simple labeling, public documents, generic stock video and chat archives sold alone are becoming a commodity.
- Value follows scarcity: buyers pay for records that exist only inside companies and that a model cannot already see.
- The strongest records link a decision to what happened next, or a question to its verified answer.
- Connected systems and long history raise value; a single export from one tool lowers it.
- Simple labeling work is being automated, and public or generic material no longer earns a premium.
- Rights decide whether value can be realized: in September 2026 a federal appeals court found a developing market for licensing material to train AI.
Which AI training data holds its value in 2026?
The short answer: records of real work that a model cannot find anywhere else keep and gain value; material that is public, generic or easy to produce is losing it. AI developers moved from teaching models to recognize things to teaching them to do jobs. That needs examples of how professionals decide, check and correct their work inside real systems.
Buyers say so directly. micro1, one of the companies buying operational data, announced on October 9, 2026 that it would spend "$1 billion over the next 12 months acquiring and licensing enterprise operational data" (micro1). It says compensation depends on "dataset size and volume, workflow complexity, data quality, domain expertise represented, and the uniqueness of the operational knowledge contained within the dataset." (micro1). Protege's chief executive, in an interview summary, sees real-world data from companies doing their work as the most valuable kind, "while bulk text loses value and synthetic data keeps falling short." (Signalful).
What makes training data gain value
Springbolt Group, an advisory firm that prepares enterprise data for licensing, sorts the premium categories in a way most buyers echo (Springbolt). The shapes below hold across industries.
| Shape of data | What it means | Why buyers pay |
|---|---|---|
| Decisions linked to their outcome | A real decision, its context and what happened afterward, such as a claim approved and later validated | Springbolt calls this "Data that captures a real decision, the context around it, and what happened afterward." |
| Long-form specialist reasoning | Memos, case notes and root-cause analyses written by specialists | Material "produced by specialists whose knowledge does not end up on the internet." (Springbolt) |
| Multi-turn resolved cases | A problem worked through over several exchanges until it is solved | Shows how a task is completed, not only what a plausible answer looks like |
| Structured records with a known ground truth | Forecasts with actuals, process settings with measured results | "Data where the ground truth is known and provable." (Springbolt), so a model can be graded |
| Linked multi-system records | Tickets, email, chat, CRM and files that refer to each other | Appen: "Connected systems are worth more than isolated exports, because value compounds where records reference each other." |
| Fresh, ongoing supply | New records delivered on a schedule, not a one-off archive | Scale AI describes partnership value "scaling with cadence"; micro1 licenses agent trajectories "for $1–2M annually" |
Sources: Springbolt, Appen, Scale AI, micro1. Protege's chief executive adds that "Joined, longitudinal data covering many types is worth far more than its parts." (Signalful).
Scarce domains where demand is rising
Some fields have almost no public training data at all. In these, a company with years of well-kept records holds something buyers cannot get elsewhere. Each row links to Origin's guide for that record type where one exists.
| Domain | Evidence of scarcity or demand | Origin guide |
|---|---|---|
| CAD with feature trees, and expert screen recordings | Markov, a Y Combinator company that has "Sold more than 15,000 hours of data to frontier AI labs", says screen data "is especially scarce for expert workflows in domains like CAD, architecture, engineering, finance, and other professional software." (Markov) | CAD and engineering data; screen recording |
| Chip design, process and materials run records | "Semiconductor design has no equivalent public corpus." (ChipAgents) | Semiconductor process data; R&D experiment records |
| Lab notebooks with failed runs | Published buyer listings ask for failed and repeated runs, not only published results | R&D experiment records |
| Claims and lending decisions | Mercor lists "Finance & Insurance Statements, underwriting files, claims" among target data (Mercor) | Insurance claims; loan underwriting |
| Audit and finance workflows | Business Insider: "Fields that AI giants are trying the hardest to automate, like law and accounting, see especially high demand." (Business Insider) | Audit workpapers; ERP and accounting records |
| Medical imaging with reports | OneMedNet signed a deal worth "more than $11.5 million" under which it "will provide de-identified medical imaging and associated clinical reports" (Investing.com) | Medical imaging |
| Private code with tickets and reviews | Mercor takes "GitHub Repositories, commits, PRs, and issues" (Mercor); there is "lower demand for a codebase that a rival lab already has its hands on" (Business Insider) | Source code |
| Two-channel call audio in languages other than English | Speech technology "is restricted to about one hundred languages" of more than 7,000 (MMS paper) | Multilingual call recordings; call recordings |
| Real-workplace video | Versos scores real-world footage for world models at "demand index 85" (Versos) | Video footage for physical AI |
| Robot touch and failure-recovery data | A 2026 survey of robot training data says datasets should cover signals "that remain scarce, particularly failure and recovery trajectories", and lists "building large-scale tactile datasets" as an open challenge (arXiv survey) | Skilled work demonstrations |
| Private evaluation sets | Protege is "moving into benchmarks and evaluations" (Signalful); Trove reports "nearly every big broker released a benchmark" (Trove) | Data for AI evaluation |
Which AI training data is becoming a commodity?
Five kinds of material are losing value or never had a premium. The pattern is the same: either a model can now produce the work itself, or the material is already available to every buyer.
| Losing value | What the evidence shows |
|---|---|
| Simple labeling and rating work | TELUS, a large provider of this work, wrote in its June 2026 statements that legacy services for certain hyperscale customers "are being automated faster than anticipated", and its test "resulted in an impairment of $2.1 billion, of which $1.6 billion was an irreversible goodwill impairment" (TELUS). Appen "reported continued first-half losses as its non-China business contracted", although its third-quarter revenue grew 34%, mostly from China (Slator). |
| Public documents | Springbolt lists "Public information you happen to have. SEC filings, published research papers, patent filings, press releases. The labs already have this." (Springbolt) |
| Generic stock video | "Stock and B-roll sits at the bottom. Generic footage is abundant and largely commoditized." (Versos) |
| Household-chore video | Demand is real, but supply is crowded. NPR reports "there's thousands of people like Jegadish across India" recording chores, and one of them "earns about $120 a month" (NPR). |
| Closed-startup chat archives sold alone | One founder netted about $25,000 for a code base and "likely around $15,000" for workspace data in Google, Notion, Figma and Slack (MIT Technology Review). A micro1 executive told Business Insider: "If a company won't sell its Slack, customer relationship management, or email records, the rest of the corpus is less valuable." (Business Insider) |
Depth beats volume even within chat archives: Surge AI's Nick Heiner told MIT Technology Review that "one company's Slack history, if it captures conversations with variety and depth, can be worth more than another's Slack, email, and calendars combined" (MIT Technology Review). See selling Slack and email archives.
What raises and lowers the price
Buyers publish few prices, and those they publish are ranges, not offers. MIT Technology Review found that "Buyers quote wildly different numbers" for similar archives (MIT Technology Review). The factors below are the ones buyers name.
| Raises the price | Lowers the price |
|---|---|
| Decisions linked to outcomes or verified answers | Records with no result attached |
| Several systems that reference each other (tickets, email, CRM, files) | One isolated export |
| Long, unbroken history: Appen says "A decade of records is worth more than a single quarter of the same thing." | A few months of data, or gaps after a system change |
| A domain with no public equivalent | Material already public or widely sold |
| Exclusivity, or data no rival lab already holds | A data set already sold to several buyers |
| Clear ownership and permission to license for training | Unclear provenance: "The labs can't validate the data, so they won't pay for it." (Springbolt) |
| A schedule of fresh deliveries | A one-time archive with no updates |
| Specialist language and reasoning | Generic, scripted or templated content |
Sources: Appen, Springbolt, Business Insider. See also exclusive vs non-exclusive licenses and recurring data licensing revenue.
Published prices and ranges, November 2026
These are the figures buyers publish themselves. A published range is what a buyer says it may pay, not an offer for your records.
| Source | Published figure | What it covers |
|---|---|---|
| micro1 | "payments typically ranging from $100K to over $3M" | Company operational data; a top tier of "$3M+" for highly unique data |
| Handshake AI | "$100K to $4M+ payouts" | Company business records |
| Scale AI | "$10K–$1M+ illustrative value per data partnership, scaling with cadence" | Data partnerships |
| micro1 | "$1–2M annually" | Licensed agent trajectories, with the company's approval |
Treat pitches with care. WIRED's Steven Levy received cold emails claiming labs pay media companies "$350K to $1.8M as a one time fee for a copy of their old work records", and Anthropic told him that it "doesn't buy workflow from live companies like WIRED." (WIRED). Mercor publishes no range; it says "Compensation is based on data volume, the tools you connect, and the depth of your data." (Mercor). For what companies have actually been paid, see what company data is worth.
Rights decide whether value can be realized
A court decision this autumn made licensing more important, not less. "On September 29, 2026, the Third Circuit issued what appears to be the first federal appellate decision on whether training AI on copyrighted material is fair use." In Thomson Reuters v. ROSS Intelligence, "the court affirmed that case-law headnotes are copyrightable and that using them to train a non-generative legal-research tool was not fair use", and it found that "the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing." (Baker Botts). For owners, the practical point is that clean, documented rights to license are part of what a buyer pays for.
The questions to answer before any value conversation:
- Do you own the records, or do they belong to a customer, client, sponsor or partner?
- Do contracts, engagement letters or NDAs limit reuse?
- Did the people in the records agree to this use? A law professor quoted by Business Insider asks: "does the consent of the employees from whom they collected that data actually extend to the sale of the data?" See the legal guide.
- Which personal details must come out, and can the records still link after they do? See de-identifying company data.
- Do sector rules apply: banking privacy, professional confidentiality, export controls, recording laws? The guides for lending, audit, semiconductors and multilingual calls cover each.
How to prepare records that keep their value
None of these steps commits you to a deal. Each one makes the records easier to value.
- Stop deletion jobs on old systems until you know what the records are worth. See keeping company data sellable.
- Map one workflow end to end: the request, the decision, the check and the outcome. The company record map helps.
- Note which systems link to each other and how many years each covers.
- List who owns each record set and which contracts touch it.
- Keep failed, rejected and reworked cases. They carry the judgment buyers want.
- Ask whether you can supply new records each quarter, not only an archive.
Forms that keep their value beyond an archive
Evaluation sets, tasks, corrections logs and recurring refreshes are covered, with the evidence for each, in the guide to evaluations, tasks and corrections a company can license beyond its archive.
Next step
Compare record types side by side in what company data you can sell to AI and buyers by data type.
This page states what each rule says and where to read it. It is not legal advice: your counsel decides how the rules apply to your records.
Does your company hold records like these? See what your data is worth.
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
Which AI training data is most valuable in 2026?
Records of real work that are not on the public web: decisions linked to outcomes, specialist reasoning, cases worked to a resolution, records with a known right answer and linked systems. Scarce domains such as chip and lab run records, lending and claims decisions, audit workflows and imaging with reports lead.
Which AI training data is becoming a commodity?
Simple labeling and rating work, which TELUS says is being automated faster than expected; public documents buyers already hold; generic stock footage; household-chore video, where supply is crowded; and chat archives from closed startups sold on their own.
Is AI training data still worth selling?
Specific data is. micro1 committed $1 billion over 12 months to buying company operational data, and micro1 and Handshake AI publish company payouts of $100K to over $3M and $100K to $4M+. Generic or public material earns little.
What raises the price of a training data set?
Outcomes attached to decisions, several linked systems, long unbroken history, a domain with no public equivalent, exclusivity, clean rights and a schedule of fresh deliveries.
Does the Thomson Reuters v. ROSS ruling matter for data owners?
Yes, indirectly. On September 29, 2026 the Third Circuit held that training a legal research tool on Westlaw headnotes was not fair use and found a developing market for licensing headnotes as training data. Documented rights to license are part of what buyers pay for.
Is synthetic data replacing real company data?
Not yet, according to buyers. Protege's chief executive says synthetic data keeps falling short while real-world data is the most valuable kind, and Mercor says public data and synthetic datasets cannot replicate how organizations actually operate.
Sources and further reading
- micro1: commitment to spend $1B on enterprise data acquisition (October 9, 2026)
- micro1: Company Data Partnerships
- Handshake AI: data partnerships
- Mercor: data monetization
- Scale AI: data partnerships
- Appen: enterprise data partnerships
- Springbolt Group: What AI labs actually pay for
- TELUS: interim financial statements for the quarter ended June 30, 2026 (SEC Form 6-K exhibit)
- Slator: Appen Q3 2026 results (October 2026)
- Business Insider: AI labs are fueling a wild market for your Slack DMs (October 2026)
- WIRED: AI companies really, really want to buy your work files and emails (November 2026)
- MIT Technology Review: AI labs are buying data from dead startups (October 13, 2026)
- Versos: AI training data demand tracker (November 2026)
- NPR: AI companies are paying people in India to film themselves doing chores (October 15, 2026)
- arXiv 2607.24744: Data Pyramid for Embodied Manipulation, a survey
- Y Combinator launch: Markov, data for computer-use AI
- ChipAgents: collaboration with NVIDIA on agentic AI for chip design
- Investing.com: OneMedNet secures $11.5M deal to supply AI training data (July 2026)
- arXiv 2305.13516: Scaling speech technology to 1,000+ languages (MMS)
- Baker Botts: Thomson Reuters v. ROSS, the Third Circuit recognizes a licensing market for AI training data
- Signalful: why AI labs pay for data they could find themselves (Protege interview)
- Trove: Maturing data market (October 12, 2026)