Key takeaways
- AI models are trained in three stages: pretraining on public text, fine-tuning on worked examples, and reinforcement learning on scored practice tasks.
- Public human-written text is running short, and models trained mostly on AI-written synthetic data lose rare cases and degrade over generations.
- Company records show employees completing real tasks step by step, which is what AI agents need to learn long tasks that span several apps.
- AI labs and data vendors turn anonymized company records into RL environments, worked examples and tests, paying $100K to $2M or more per dataset.
01
What is AI training data?
The short answer
AI training data is the collection of examples an AI model learns from, including written text, worked examples of tasks done well and practice tasks the model is scored on. A model can only learn the kinds of work its training data shows it. Most training data so far has come from the public internet, and AI labs and data vendors now want records from inside companies that show employees doing real work from start to finish.
How an AI model learns from examples
An AI model is a program that learns patterns from examples instead of following rules a programmer wrote. During training it studies millions of examples, compares its own output with each one, and adjusts its internal settings a little each time until its output closely matches the examples.
The examples it studies are called training data. A set of examples gathered for one purpose, such as teaching a model to answer support tickets (the written help requests customers send to a company's support team), is called an AI training dataset.
Training data comes in three kinds
AI models learn from three kinds of training data, one for each stage of training explained in the next section.
- Raw text
- Trillions of words of public writing, such as websites, books and computer code. The model reads it to learn how language works and what is generally known about common topics.
- Worked examples
- Pairs of a request and a good response, written or checked by people. A customer's question paired with the answer an experienced support agent gave is one worked example.
- Tasks with a score
- Jobs the model attempts on its own, such as fixing a bug in a piece of code, each with an automatic check that says whether the attempt succeeded. The model improves by repeating the jobs and keeping what earned a pass.
A fourth kind, evaluation data, isn't used for learning at all. It is a set of tasks kept out of training and used afterward to test how well the finished model performs. It works like a final exam, because the model has never seen the questions before.
Who makes and buys training data
Two kinds of companies build and buy training data. A business owner who licenses company records usually deals only with the second kind, the data vendor.
- AI labs
- Companies that build and train AI models, such as OpenAI, Anthropic and Google DeepMind.
- Data vendors
- Companies that collect, clean and package training data for AI labs. Some buy records from businesses, anonymize them (replace every name and personal detail with a stand-in) and turn them into worked examples and practice tasks.
02
How AI models are trained: pretraining, fine-tuning and reinforcement learning
The short answer
AI models are trained in three stages. In pretraining the model reads public text, in fine-tuning it studies worked examples, and in reinforcement learning it practices tasks and is scored on each attempt. Company records are used in the second and third stages, which are the stages AI labs are now spending more on.
Pretraining: reading the public internet
Pretraining is the first and largest stage. The model reads trillions of words from websites, books, articles, forums and open-source code, and learns to predict the next word in a passage. To predict the next word correctly across that much text, the model has to pick up grammar, facts about the world and the usual steps of an argument.
Fine-tuning: studying worked examples
Fine-tuning teaches the pretrained model to respond helpfully to requests, instead of just continuing whatever text it is given. The lab shows it thousands of worked examples, each a request paired with a good response, and the model adjusts itself to produce responses like them.
Fine-tuning uses far less data than pretraining, yet it decides much of how useful a model is to the people asking it questions. In OpenAI's 2022 InstructGPT research, people preferred answers from a fine-tuned model with 1.3 billion parameters over answers from the original GPT-3, a model more than 100 times larger (Ouyang et al., 2022). A smaller model trained on well-chosen worked examples gave people better answers than a much bigger model without them.
Reinforcement learning: practicing and being scored
Reinforcement learning (RL) is training by trial and score. The model attempts a task on its own, an automatic checker scores the result, and the model's settings are adjusted so it becomes more likely to repeat the actions that earned high scores. After thousands of attempts, the model can do tasks that no one ever wrote out for it step by step.
Reinforcement learning became the main focus of AI training research in 2025. DeepSeek reported that its R1 model learned to reason through math and coding problems "through pure reinforcement learning", without people writing out the reasoning first, on tasks whose answers can be checked automatically (DeepSeek-AI, 2025).
Where company records enter
Company records don't go into pretraining, which runs on public text. They enter at the two later stages, as the diagram below shows. A resolved support ticket, from the customer's first message to the fix, becomes a worked example for fine-tuning. A company's anonymized inbox, ticket queue and CRM (customer relationship management software, where a company tracks its customers and deals) become the contents of a practice copy of its software, where a model attempts tasks during reinforcement learning.
03
Why AI is running out of public training data
The short answer
Epoch AI, a research group that tracks AI trends, estimates the stock of public human-written text at about 300 trillion tokens and expects AI training to use all of it between 2026 and 2032. At the same time, more websites are blocking the programs that copy their text for AI training. AI labs therefore can't get the data for their next models from the public internet alone.
There is only one internet
Every new generation of AI model has been trained on more text than the last. The amount of public text people have ever written is finite, though. Epoch AI estimates the usable stock of public human-written text at about 300 trillion tokens, a token being a word or piece of a word. Its 80% confidence range for when AI training uses all of it runs from 2026 to 2032 (Epoch AI, 2024).
Leading AI researchers have reached the same conclusion. At the NeurIPS research conference in December 2024, Ilya Sutskever, a cofounder of OpenAI, told the audience that pretraining as it is done today will end, because computing power keeps growing while the supply of data doesn't. "We have but one internet," he said (Maginative).
The open web is closing to AI crawlers
Even the public text that exists is getting harder to collect. A study by the Data Provenance Initiative found that in a single year, 2023 to 2024, websites restricted AI crawlers (programs that visit websites and copy their text for AI training) from about 5% of all the text in C4, a widely used training dataset built from public web pages, and from more than 28% of its most actively maintained sources (Longpre et al., 2024).
Public text shows results, not the work behind them
A bigger limit is the kind of information public text contains. A business puts its product pages and blog online, and its customers post reviews. The business doesn't publish the support tickets, internal emails and decisions behind those pages and reviews. A model trained only on the public web has therefore seen the results of millions of jobs but almost never the steps employees took to do them.
That gap mattered little while AI was mostly used to answer questions. It matters now that AI labs want models that carry out whole jobs, such as handling a customer's repair from the first phone call to the final invoice. Our guide on what company data is worth follows one such job at an HVAC company, a business that installs and repairs heating and air conditioning.
04
From chatbots to AI agents: why the next step needs new training data
The short answer
AI labs are moving from chatbots, which answer questions, to AI agents, which carry out whole jobs across several software apps. Today's agents fail most often on long tasks with many steps. The records companies keep are the only large source of examples showing people completing those tasks, with the result attached.
What an AI agent is
An AI agent is AI software that carries out a job step by step, using the same apps an employee uses, instead of only answering a question. An agent can open an inbox, read a customer's email, look up the order in another system, issue a refund and reply.
A chatbot only needs to know facts in order to answer questions. An agent also needs to know how a company's work gets done. It has to know which app holds the customer's order, what to check before issuing a refund, and when a case is unusual enough to hand to a person.
Where agents break: long tasks across several apps
Today's agents usually complete short tasks correctly and often fail on long ones. AI researchers call a task that takes many steps, several apps or a long stretch of time a long-horizon task. On those tasks an agent tends to lose track of what it has already done, or makes one wrong step that spoils everything after it.
micro1, the data vendor we typically refer companies to, gives an example of the kind of task AI labs want agents to learn. A salesperson checks the CRM for the contacts on a deal, looks up the contract's status in a second system, then reads the latest team-chat thread with the legal team to decide whether to follow up. The task uses three apps and ends in one decision, and a mistake in any of the three lookups leads to the wrong decision.
Why "long horizon" matters to buyers
AI labs and data vendors value records of long, multi-app tasks most, because those are the tasks agents fail at today. A company's archive of email, chat and CRM history holds thousands of those tasks, each already finished and showing how it turned out.
05
Why synthetic data isn't enough
The short answer
Synthetic data is training data written by an AI model instead of people. It works well for tasks with a checkable answer, such as math and code, but a model trained mostly on AI-written data gradually loses the rare cases and starts producing errors. Data vendors start synthetic examples from real company records, so the generated examples still match how real customers and systems behave.
What synthetic data is
Synthetic data is training data generated by an AI model rather than collected from people. A lab can ask one model to write ten thousand customer complaints and another to answer them, and train on the result.
Synthetic data is cheap, and a model can generate as much of it as a lab asks for, so AI labs use a great deal of it.
Model collapse: what happens when AI learns from AI
Synthetic data has a known failure mode. A 2024 study in Nature found that models trained on data generated by earlier models lose the rare and unusual cases first. Over several generations their output drifts into nonsense, which the authors call "model collapse" (Shumailov et al., 2024).
Rare cases are where AI agents make the most mistakes at work. An angry customer with an unusual billing history, an order split between two shipping addresses and a contract with a non-standard clause are all cases that trip agents up. An AI model writing practice examples from what it already knows tends to leave such cases out, because it reproduces the common pattern.
Where synthetic data works, and where real records are needed
Works well
Tasks with a checkable answer
A math problem has one right answer and code either runs or doesn't, so a lab can generate millions of practice problems and check every result automatically.
Reasoning models, which work through a problem step by step before answering, improved quickly in 2025 this way.
Needs real records
Business work with judgment
Whether a refund was the right call depends on the customer's history, the company's policy and what happened afterward. Only records of real cases show those details and how each case turned out.
Data vendors start from real records and generate variations of each case, so the synthetic examples keep the real details.
06
What company records add to AI training data
The short answer
Company records show how employees actually did their work, from a customer's first request through each step, check and correction to the final result. Public web data shows only what a business says about itself. AI agents learn to do jobs from company records.
The same business, seen from outside and inside
Every business produces two kinds of information. One is what it publishes for customers, and the other is what its staff record while doing the work. Until now, AI has mostly been able to read only the first kind.
A small accounting firm is a good example. Its website, blog and job listings tell the public which services it offers. Its internal records, from the client's email and the accountant's working notes to the manager's review and the CRM entry, show how one client's tax return was actually prepared. Those records include the missing mileage log the manager caught before the return was filed, as the diagram below shows.
Which company records AI labs and data vendors want
AI labs and data vendors most want records written while the work was happening, because those show each step. micro1 lists standard operating procedures (SOPs, the written instructions for how a company does a task), process documents, internal knowledge bases, CRM records, support tickets, internal communications and project histories (micro1).
| Record | Who creates it | What it teaches an AI agent |
|---|---|---|
| Support tickets | Support staff, each time a customer asks for help | How a problem is diagnosed and resolved, from first message to fix |
| Team chat and email | Employees across the company, while they coordinate work with each other | How people ask for help, hand off tasks and make decisions |
| CRM history | Sales and account staff, as deals move forward | Which steps lead from a first contact to a closed deal or a lost one |
| SOPs and knowledge bases | Managers and senior staff | The rules and checks a company expects its staff to follow |
| Project and task histories | Teams using project-tracking software such as Jira or Asana | How large jobs are broken into steps, tracked and finished |
07
How AI labs and data vendors turn company records into RL environments
The short answer
An RL environment is a working copy of a company's software, filled with anonymized records, where an AI agent practices a task and is scored on every attempt. Data vendors build RL environments from company records so that the tasks an agent practices match the tasks real employees handle.
What an RL environment is
An RL environment (reinforcement learning environment) is a simulated workplace where an AI agent practices tasks and is scored automatically. TechCrunch describes RL environments as "training grounds that simulate what an AI agent would be doing in a real software application" (TechCrunch, 2025).
Every RL environment has three parts. The world is the content, such as the messages, documents and customer records. The apps are working copies of the software a company uses, such as an inbox, a ticket queue and a CRM. The tasks are the jobs the agent must complete, each with a scoring rule that checks whether the job was done right.
Inside an RL environment, an agent can practice a billing dispute (a customer's complaint about a charge) a thousand times without touching a real customer or a real bank account. Trainee pilots practice in flight simulators for the same reason.
How an agent learns inside an RL environment
The agent learns by repeating a loop, shown in the diagram below. The agent receives a task and works through the copied apps. A scorer, a program that checks the result against the task's rules, then grades the attempt. The AI lab uses the score to adjust the agent, and the agent tries the task again.
How a company's records become an RL environment
Data vendors turn a company's records into an RL environment in roughly five steps, though the details differ by vendor and task.
Anonymize the records
The data vendor replaces names, emails, phone numbers and other personal details with consistent stand-ins, so conversations still make sense but no real person can be identified. Our guide on how data is anonymized shows how.
Rebuild the world
The data vendor loads the anonymized emails, tickets, documents and CRM entries into working copies of the apps. micro1 calls this snapshot of a company "seed data" and wants it as complete as possible. If a document refers to a chat thread that is missing, the agent can't find the information the task depends on.
Write tasks from real cases
The data vendor turns real jobs from the records, such as a customer charged twice or a sales deal that stopped moving, into tasks for the agent. The vendor can then generate variations of each real case, such as the same double charge on a different order.
Write the scoring rules from real outcomes
The records show how the company's staff resolved each case and whether the customer was satisfied. The data vendor turns those outcomes into scoring rules, such as "refund the extra charge once and reply to the customer."
Let the agent practice
The AI lab runs its agent through the environment thousands of times and uses the score from each attempt to adjust the agent.
Why realism decides whether an RL environment works
An agent in an RL environment learns whatever behavior earns a high score, even when that behavior isn't what the builders intended. Ross Taylor of General Reasoning, an AI research company, told TechCrunch that RL environments are prone to reward hacking, where a model finds a way to collect the reward without really doing the task.
A simple simulation with a few made-up emails might let an agent score well by replying "refund issued" without ever checking the billing system. An environment built from a real company's connected records lets the scorer check that the refund was actually made, against the right charge. That protection against shortcuts is a large part of why AI labs and data vendors pay for real company data.
08
The other two uses: worked examples and evaluations
The short answer
Besides RL environments, AI labs and data vendors use company records as worked examples, which show a model a task done well from start to finish, and as evaluations, which test where agents still fail on real work. All three uses need records that cover each task from start to finish and link related messages and files together.
Use 2
Worked examples
A complete record of a task done well, such as a support ticket from the customer's first message to the fix, is used in fine-tuning. The model studies each step the employee took and learns to take the same steps on similar requests, much as a student checks work against an answer key.
A ticket that is missing the final fix can't teach the model how the problem was solved.
Use 3
Evaluations
Real tasks from company records, kept out of training, become a test the agent takes. The results show AI labs which kinds of tasks their agents still fail, so they know what to train next.
Evaluations and RL environments are often built from the same records.
Recorded workflows: a newer kind of worked example
Some data vendors go a step further and pay a company's employees to do their real tasks with AI tools while software records each step. Each finished task becomes a trajectory, a step-by-step record of an agent-assisted job, with the outcome attached. micro1 runs this as an ongoing, paid program alongside one-time purchases of a company's archive.
micro1's founder has described the goal as moving AI training "from simulated environments to the actual work environment." In practice, that means agents learn from real employees doing real tasks, instead of only from practice copies.
09
What makes good AI training data
The short answer
Good AI training data covers each task from start to finish, links related records across systems, shows how each task turned out, includes unusual cases and comes with clear rights to use it. We also believe good training data should show work done well and with care, because an AI model copies the habits shown in the records it learns from.
The qualities AI labs and data vendors check
- CompleteThe whole task is there, from the first request to the result, with no missing threads or attachments.
- ConnectedRecords link to each other across systems, so a ticket can be followed to the fix and a chat thread to the decision.
- Shows outcomesThe record says how each task turned out, such as whether the deal closed, the customer came back or the repair held.
- Covers unusual casesYears of history include the rare problems that agents fail on and synthetic data leaves out.
- Clear rightsThe company owns the records and is allowed to license them. See whether it's legal to sell company data.
- AnonymizablePersonal details can be replaced with stand-ins without breaking the meaning of the records.
The amount of data matters less than most owners expect. A support history that links each customer's request to the fix can teach an agent more than a much larger pile of unrelated files.
Good data, in both senses: our view
We believe the data an AI learns from shapes how it behaves at least as much as the lab that trains it. An AI trained on a company's records picks up that company's habits, good and bad. A new hire picks up habits from the people around them in the same way.
That is why we think the companies supplying data, and the vendors building environments, carry real responsibility. An environment that rewards one narrow goal, such as closing tickets fast, can teach behavior nobody wanted, such as closing tickets without solving them. We'd like AI to learn from records of work done well and with care. We prefer data vendors that choose carefully what goes into training and that check what an agent's behavior leads to, not only its score.
10
How much AI labs and data vendors spend on AI training data
The short answer
AI labs spend billions of dollars a year on training data, much of it through data vendors. A company licensing its own records is typically paid $100K to $2M or more per dataset. The largest archives draw eight-figure bids: Spirit Airlines' internal records drew a $10M winning bid from Google and a later $12.5M bid from micro1 in 2026.
The money flowing to data vendors
The clearest signs of what AI labs spend on training data are the sizes of the vendors they buy from. Meta invested $14.3 billion for a 49% stake in a large data-labeling vendor, a company that pays people to label and write training examples, in June 2025 (The Batch). Mercor, a data vendor whose customers include Anthropic and OpenAI, reached $2 billion in annual recurring revenue in June 2026, according to Fortune (Fortune).
Spending on RL environments is growing fastest. Leaders at Anthropic have discussed spending more than $1 billion on RL environments over a single year, The Information reported, according to TechCrunch. micro1 has committed to spend $1 billion over 12 months acquiring and licensing company operational data to build RL environments (micro1).
Meta's investment in a data-labeling vendor, June 2025
Mercor's reported annual revenue run rate, June 2026
micro1's 12-month commitment to buying company data
Why AI labs pay for data instead of scraping it
Company records can't be scraped (copied automatically from public websites), because they sit behind logins. AI labs have also learned that using data without permission is expensive. Anthropic agreed in 2025 to pay $1.5 billion to settle a lawsuit by authors over pirated books, about $3,000 per book (NPR).
A license is a written agreement that lets a buyer use data for set purposes while the company keeps ownership. Buyers pay more for anonymized company data that comes with a license and clear proof that the company owns it, because it carries little risk of a lawsuit. Our guide to data licensing explains in detail.
What a company is paid for its records
Companies that license their records to AI labs and data vendors are paid $100K to $2M or more per dataset. micro1 breaks its own payments into three tiers: qualified company data at $100K–$500K, large-scale datasets or ongoing participation across several teams at $500K–$2M, and highly unique data at $2M+ (micro1).
The largest archives draw eight-figure bids. In August 2026 Google won the bankruptcy auction for Spirit Airlines' internal emails and Teams messages with a $10M bid, as Decrypt reported. micro1 then filed a late bid of $12.5M (Fortune), and at the latest reports the bankruptcy court had not yet approved a buyer.
The data value estimator gives a range for your company, and the data monetization examples guide lists real, dated deals.
11
AI training data news: the milestones that changed the market
The short answer
From 2022 to 2026, AI training moved from scraping public text toward fine-tuning, reinforcement learning and licensed company data. The dated milestones below show how quickly that happened.
| Date | What happened | Why it matters |
|---|---|---|
| Mar 2022 | OpenAI publishes InstructGPT research (arXiv) | Shows that a small model trained on good worked examples beats a far larger one |
| Jun 2024 | Epoch AI estimates public text runs out between 2026 and 2032 (Epoch AI) | Puts a date on the end of public training data |
| Jul 2024 | Study finds websites restricted AI crawlers from 28%+ of key sources in a year (arXiv) | Websites begin blocking AI from the text it trained on |
| Jul 2024 | Nature publishes the "model collapse" study (Nature) | Shows the limits of training AI on AI-written data |
| Dec 2024 | Ilya Sutskever tells NeurIPS that pretraining as done today will end: "We have but one internet" (Maginative) | A leading researcher calls the end of the public-data era |
| Jan 2025 | DeepSeek releases R1, trained to reason with reinforcement learning (arXiv) | Reinforcement learning becomes the main focus of AI training |
| Jun 2025 | Meta invests $14.3B in a data-labeling vendor (The Batch) | Shows how much an AI company will pay for a stake in a data vendor |
| Sep 2025 | Anthropic settles with authors for $1.5B (NPR) | Using data without a license now carries a billion-dollar cost |
| Sep 2025 | TechCrunch reports the rush to build RL environments (TechCrunch) | Building RL environments for AI agents becomes a business of its own |
| Jun 2026 | Mercor reaches $2B in annual recurring revenue (Fortune) | Shows the scale of lab spending on data vendors |
| Oct 2026 | micro1 commits $1B to buying company data (micro1) | A data vendor budgets $1B to buy company records for RL environments |
12
What this means for a company with years of records
The short answer
A company that has run its work through email, chat, CRM and ticketing software for years holds records of real work that AI labs and data vendors can't get from the public internet. The sensible next steps are to check whether the company qualifies, estimate what the records are worth and understand the privacy and legal terms before sharing anything.
Why the current premium probably won't last
We think AI labs and data vendors pay a premium today partly because few owners know their records are worth anything. As more companies start licensing their records, we expect prices to fall toward a steadier level. micro1 has said as much, telling owners that "commercial conditions will probably change."
Four steps before you talk to a buyer
- Check whether your company fits what buyers look for, using the qualification checker.
- Get a range for your records with the data value estimator.
- Learn how your records would be anonymized and what a data license lets the buyer do.
- Compare buyers in our guide to where to sell data, and have a lawyer review any agreement before you send a sample.
Questions
Frequently asked questions
What is AI training data?
AI training data is the set of examples an AI model learns from. It comes in three main kinds. Raw text, such as websites and books, teaches language; worked examples pair a request with a good response; and practice tasks let the model attempt a job and be scored. Most early AI training data came from the public internet. AI labs and data vendors now also buy anonymized records of real work from companies to train AI agents.
Where do AI companies get their training data?
AI companies get training data from the public web, from licensing deals with publishers and platforms, from experts hired to write and grade examples, from synthetic data generated by other models, and increasingly from companies that license their internal records. Most company records reach AI labs through data vendors such as micro1, Mercor and Polyshares, which anonymize the records and turn them into worked examples and RL environments.
Is AI running out of training data?
AI is running short of public human-written text. Epoch AI estimates the stock at about 300 trillion tokens and expects AI training to use it all between 2026 and 2032, while more websites block AI crawlers. AI labs are responding with synthetic data, reinforcement learning and licensed company records, which contain the step-by-step examples of real work the public internet never had.
Why would an AI company want my company's emails, chat and tickets?
AI companies want company emails, team chat and support tickets because those records show employees doing real work, from a customer's request through each step to the result. AI labs and data vendors anonymize the records and use them to build RL environments where AI agents practice jobs, to create worked examples of tasks done well, and to test where agents still fail.
What is an RL environment?
An RL environment, short for reinforcement learning environment, is a simulated workplace where an AI agent practices tasks and is scored automatically on each attempt. It combines content such as emails and customer records, working copies of business software such as an inbox and a CRM, and tasks with scoring rules. Data vendors build RL environments from anonymized company records so the practice resembles real work.
Why can't AI companies just use synthetic data?
Synthetic data, written by AI models, works well for tasks with a checkable answer such as math and code. For business work it falls short, because a model writing examples from memory leaves out the rare, messy cases that trip agents up. A 2024 Nature study found models trained repeatedly on AI-generated data degrade, which it called model collapse. Data vendors therefore start from real company records and generate variations of real cases.
Will an AI model learn my customers' names or trade secrets?
No. Data vendors replace names, email addresses, phone numbers and other personal details with consistent stand-ins before any AI lab sees the records, so the conversations still make sense but no one can be identified. Buyers also require companies to leave out trade secrets and legally protected material, and a company can exclude other categories, such as HR records, though excluding more may lower the price.
How much do AI companies pay for company data?
AI labs and data vendors pay from $100K to $2M or more per dataset for company records. The largest archives draw eight-figure bids: Google bid $10M for Spirit Airlines' internal records in 2026, and micro1 later bid $12.5M. The price depends on how complete, connected and unusual the records are, how many years they cover and whether the company can show clean rights to license them. A data value estimator gives a starting range before any data is shared.
Bottom line
The gap is inside companies, and buyers are paying to fill it
If your company has run on software for years, its records show exactly the kind of work AI labs and data vendors can't get anywhere else. Find out what they're worth while buyers are still paying a premium.
Not legal, tax or financial advice. Figures come from micro1's published pages and public reporting as of November 2026.