Workflows

AI Training Data: Why AI Companies Want Business Records

AI already read the internet. What AI training data it still lacks is how real businesses work. Why labs now pay for business records, and how the market works.

Rows of old labeled filing-cabinet drawers

Photo by Mike Stoll on Unsplash

On this page

AI training data is the material AI models learn from, and for most of the last decade that meant the public internet: websites, books, articles and forums. The models have already read most of it. What they never saw is how a business actually runs day to day: the email thread where a contract got renegotiated, the maintenance log that shows which repair worked, the report a bookkeeper rewrote three times before it went to the client. That record of real decisions is now one of the scarcest kinds of AI training data, and AI labs are paying for it, both by licensing company records and by hiring experts to create it. Here is why that is happening and how the market works.

Disclosure: we are a micro1 referral partner. Links marked "sponsored" are referral links, and we may be paid if you sign up through them. The explanation and the risks below are ours either way.

TL;DR

AI models learned to write and reason from public text, and researchers at Epoch AI estimate the supply of high-quality public human-written text could be fully used somewhere between 2026 and 2032. Public text also says little about how professionals do real work inside a company. So AI labs now pay for two things: licensed copies of real business records, and experts who create examples and grade AI answers. Go-between companies such as micro1, Mercor and Handshake AI connect labs with the businesses and people who supply this data. Most small businesses hold useful records without knowing this market exists. Any sharing needs the right permissions and must protect customer and employee information. If you want the practical steps and the buyer list, read our guide on how to sell data to AI companies. If you want to get comfortable working with your own files and AI first, start with the free Claude Cowork course.

Where AI training data comes from

Large language models, the technology behind chatbots like Claude and ChatGPT, learn by reading huge amounts of text and picking up the patterns in it. Until recently, almost all of that text came from public sources:

  • Websites and forums, collected by automated crawlers.
  • Books, articles and research papers.
  • Code from public software projects.
  • Licensed content from publishers and platforms that agreed to share it.

That gave AI a broad education. It can explain a lease clause or summarize a tax rule because millions of public pages discuss them. But a public page about lease clauses is not the same as watching a property manager handle a tenant dispute from first email to signed settlement. The first is general knowledge. The second is skill.

Why AI companies now buy business data

1. The public internet is running low

In June 2024, researchers at Epoch AI, an independent group that studies AI trends, published a widely cited estimate in its paper "Will we run out of data?". They put the stock of high-quality public human-written text at roughly 300 trillion tokens (a token is a chunk of a word) and projected that, if trends continued, AI training would fully use it sometime between 2026 and 2032.

That is a projection, not a cliff, and labs have other tricks, such as having AI generate practice material. But the direction is clear: the easy, free source of new text is drying up, so new data has to come from somewhere else.

2. Labs already started paying for content

You can see the shift in the licensing deals of the past two years. In February 2024, Reuters reported that Reddit had signed a deal letting Google train AI on its posts, worth about $60 million a year. In May 2024, News Corp, the owner of The Wall Street Journal and the New York Post, announced a multiyear deal letting OpenAI use its current and archived articles; the Journal reported it could be worth more than $250 million over five years. The Associated Press, Axel Springer and the Financial Times have signed their own licensing agreements with OpenAI.

Those deals are about public writing. The newer wave goes one step further, into private business records.

3. What AI still cannot learn from the internet: how work gets done

The goal for AI labs is moving from chatbots that answer questions to AI "agents" that can complete real tasks: review a contract, prepare a monthly report, sort out a billing problem, schedule a repair. To learn that, a model needs examples of the whole job, not just the finished product.

Business records hold exactly that. A ten-year archive from a property management company, for example, contains lease negotiations, tenant complaints and how each was resolved, vendor quotes, maintenance logs, and owner reports. Together they show how experienced people make real decisions under real constraints, which is the thing public text almost never captures.

In April 2026, Forbes reported, and Gizmodo summarized, that startups shutting down were licensing their old Slack messages, emails and documents to AI companies through SimpleClosure, a firm that handles company wind-downs. The buyers wanted the material to train AI agents on workplace tasks. SimpleClosure's CEO described a "gold rush" for real-world data, and the reported payouts were $10,000 to $100,000 per company. The same coverage noted privacy concerns raised by the Center for AI and Digital Policy, which is a fair warning about how this market can go wrong.

4. Expert knowledge is the scarcest piece

The most valuable training data is the kind only a trained professional can produce: how a nurse decides what to escalate, how an accountant spots a misclassified expense, how a lawyer redlines an indemnity clause. There is very little of it in public, and a non-expert cannot fake it. That scarcity is why expert roles on AI training platforms list some of the highest hourly rates, and why buyers ask most for records from law, accounting, healthcare, finance and engineering firms.

What makes business records valuable, by industry

These are illustrations of the kinds of tasks each industry's records can teach, not a list of what any particular buyer has purchased.

IndustryRecords that show real workWhat an AI could learn from them
Property managementLeases, tenant emails, work orders, maintenance logs, owner reportsHandling tenant issues, scheduling repairs, preparing owner statements
Accounting and bookkeepingWorkpapers, client emails, reconciliations, review notesCategorizing transactions, spotting errors, preparing reports
Law firmsDrafts and redlines, closed matter files, client correspondenceReviewing and negotiating contracts, organizing case files
Trades and field servicesQuotes, job notes, photos, invoices, customer messagesEstimating jobs, diagnosing problems, scheduling crews
Healthcare adminScheduling and billing workflows, prior authorization paperworkHandling claims and admin problems (with strict privacy rules)
Any office teamYears of email, chat, shared documents, SOPsFollowing procedures, writing routine documents, handing off work

Two qualities matter more than the industry: volume (many people over many years) and completeness (the whole trail from request to finished work, not just the final PDF).

How the AI data market works

There are three main players.

  1. AI labs build the models and need the data. Most do not want to negotiate with thousands of small businesses one by one.
  2. Intermediaries sit in the middle. Companies such as micro1, Mercor, Handshake AI, Mode and datasupply find businesses and experts, check that the data can legally be used, help remove personal information, and package it into a form a lab can train on.
  3. Suppliers are the businesses with records and the individual experts with know-how.

We are not naming which labs buy from which intermediary, because those contracts are mostly private and we have not seen them confirmed.

Two ways to get paid

License your records. A company licenses a de-identified copy of its records, such as emails, documents, tickets and procedures. It keeps the originals. Buyers advertise ranges like $100K to $3M or more for larger firms, but those are advertised ranges, not offers, and most want 10 to 30 or more employees. Our how-to guide lists the buyers, what each one wants, and the steps.

Supply expertise. Not every intermediary buys datasets. Many mainly hire experts, such as lawyers, accountants, nurses and engineers, to write hard questions, create example answers and grade what the AI produces. That is the same market seen from the individual's side. It pays by the hour; see our guide to AI training jobs, or our review of micro1, one of the larger platforms.

Some programs combine the two. micro1, for example, runs a company program for larger firms and a separate video program for businesses with three or more employees that pays per approved hour of footage of everyday work. If you run a company and want to see whether you qualify, you can check micro1's company program (sponsored).

What this means for a small business

For most of history, a business made money from its information in one way: by using it to operate. Your files helped you serve customers, price jobs and avoid repeating mistakes. Now some of those same files may also be worth something as AI training data, because AI has to learn from real businesses.

A few honest takeaways:

  • Most small businesses do not know this market exists. Buyers have only recently started advertising to business owners.
  • Size matters today. Most record-licensing programs want a real team and several years of history. Very small teams have fewer options, though some programs start at three to six employees depending on the industry.
  • Location varies. Some buyers say they work only with US-based teams; micro1's video program says it accepts companies worldwide. If you are outside the US, check each program's requirements before you spend time on it.
  • Knowing what you have is useful either way. Taking stock of your records helps you run the business better, whether or not you ever license a file.

Risks and permission come first

Selling access to business records is not like selling old office furniture. Before any data leaves your systems:

  • Make sure you have the right to share it. Customer contracts often promise confidentiality. Licensing their information, even with names removed, can breach those promises.
  • Protect personal data. Records are full of customers' and employees' names, addresses, health details and account numbers. Ask exactly who removes that information, how, and whether you can inspect the result.
  • Respect regulated records. Health, legal and financial data come with their own laws. A buyer's promise to "de-identify" does not move your legal duty to them.
  • Tell your team. Employee privacy rules vary by state and country, and some require notice or consent.
  • Watch for scams. A real buyer pays you. Anyone asking for an upfront fee, your passwords or admin access before a signed agreement is not a buyer.

I am not a lawyer, and this is not legal advice. Talk to one before you share anything. Our how-to guide has the full checklist.

Frequently asked questions

What is AI training data?

AI training data is the information an AI model learns from. For chatbots, that has mostly been public text from websites, books, articles and code, plus licensed content from publishers. More and more, it also includes licensed business records and examples written by experts, because those show how real work gets done.

Where does AI get its training data?

Mostly from the public internet, collected by automated crawlers, along with books, research papers and public code. AI companies also license content from platforms and publishers, such as Reddit and News Corp, and now pay businesses and experts for private records and specially written examples.

Why do AI companies buy data from businesses?

Because public text is running low and says little about how professionals actually do their jobs. Business records such as emails, documents, tickets and procedures show real decisions from start to finish. That is what AI labs need to train models that can complete real work tasks, not just answer questions.

Is AI running out of training data?

Researchers at Epoch AI estimated in 2024 that, if trends continued, AI training would fully use the stock of high-quality public human-written text sometime between 2026 and 2032. That is a projection, not a hard limit, but it is one reason labs are paying for new sources such as licensed business records and expert-written examples.

Can a small business sell its data to AI companies?

Sometimes. Most record-licensing programs want 10 to 30 or more employees and several years of records, though a few start smaller, including a video program for companies with three or more employees. Whether it is legal depends on your contracts, privacy rules and industry regulations, so talk to a lawyer before sharing anything.

Know what your business already holds

Whether or not you ever license a file, the records your business has built up are a picture of how you work, and AI can help you use them better today. The free Claude Cowork course teaches you, hands-on, to work through your own documents with AI, no technical background needed. When you are ready to look at the market itself, our guide on how to sell data to AI companies walks through the buyers and the steps.

← All guides


MIT 2026 © Nextra.