For established businesses
Sell Data for AI Training
You can sell data for AI training by licensing a copy of the records your business already keeps, such as email, chat, tickets and documents, to a company that uses them to teach AI models. You are paid for permission to use the copy. You still own the originals, and you decide which records are part of the deal.
Below we explain what training data means for a normal company, which of your records count, the different ways a license can be priced and paid, and the exact path from a folder of old files to money in your account. A free estimate sits at the top and bottom of the page.
- License, keep ownership
- Buyer pays our fee
- You set the exclusions
Plain definition
What training data means for a business
An AI model learns by studying examples, millions of them. Each example shows a situation and what a person did about it. Give a model enough examples of good work and it gets better at doing that work.
For a business, an example is something you made while doing your job. A customer writes in with a problem and your team replies until it is fixed. That pair is a training example. So is a long email chain where a quote turns into a signed job, or a how-to guide a new hire follows on day one. You have produced thousands of these without thinking of them as data.
AI companies want them because the free supply is drying up. A 2024 forecast from Epoch AI found that models could use up the stock of public human-written text somewhere between 2026 and 2032. Private records like yours were never on the open web, so no one has trained on them yet. For the wider picture, read our overview of licensing company records to AI firms.
What counts
Which business records count as AI training data
Here is what a buyer sees inside each common record type. Most deals use one to three of these, covering several years.
| Record | Where you probably keep it | The training example inside |
|---|---|---|
| Support tickets | Zendesk, Freshdesk, Help Scout, Jira Service Management | A customer's problem paired with the steps and reply that solved it |
| Outlook and Microsoft 365, Gmail and Google Workspace, PST archives | A request, the questions it raised, and the agreement that came out of it | |
| Chat | Slack, Microsoft Teams, Google Chat | Coworkers splitting up a task, asking for help and correcting course |
| Standard operating procedures | Confluence, Notion, SharePoint, shared drives | Numbered instructions for doing a task the right way |
| Proposals and reports | Google Drive, OneDrive, Box, Dropbox | Polished writing aimed at a client, tied to a real need |
| CRM activity | Salesforce, HubSpot, Zoho CRM | Follow-ups across weeks that show how a sale moves forward |
| Recorded calls | Zoom, RingCentral, Microsoft Teams recordings | Spoken back and forth, with natural pauses and clarifications |
| Code and commit history | GitHub, GitLab, Bitbucket | A change to a program plus the note explaining why it was made |
What does not count
Records that add risk or add nothing
Selling AI training data goes better when the weak and risky material is cut first. These are the categories we usually suggest leaving out. Treat this as general guidance rather than legal advice, and ask your lawyer to read any license before you sign.
Automated messages
System alerts, receipts, newsletters and auto-replies have no real person thinking through a problem. They pad the file without raising the price.
Personnel and payroll files
Pay slips, reviews, benefits forms and ID copies are about your staff as people. They raise your risk while showing a model almost nothing about how the work gets done.
Customer personal and payment data
Card numbers, account files and personal details about customers stay out. A deal like this is about how your business works, not about who your customers are.
Material under someone else's NDA
Attorneys at Frankfurt Kurnit Klein and Selz noted in 2026 that email and chat often contain other companies' confidential information. Those threads and folders can be excluded by name.
Records your old promises cover
The same note warns that an old privacy policy or handbook may not mention AI training, and the FTC has said that changing privacy promises after the fact can be unfair or deceptive. Read yours before you choose.
Regulated records
Patient records fall under HIPAA, and many professions owe clients strict confidentiality. If your field is regulated, ask your lawyer which records fall inside those rules before anything is listed.
Price and payment
How a training data license is priced and paid
There is no single price list for business records. Offers are built from a few moving parts, and knowing them helps you compare two deals that look alike at first glance.
| Part of the deal | How it usually works | What it means for you |
|---|---|---|
| Company size | More staff means more people writing, so more records and more variety. | Most buyers start at about 20 employees. Larger teams draw larger offers. |
| Years of history | Each extra year adds volume and shows how the work changed over time. | Three or more years is the usual starting point. Ten is better than three. |
| Record types | Some types are scarcer than others, and a buyer pays more for what it lacks. | Two well-chosen types can be worth more than six common ones. |
| Exclusive or not | Non-exclusive terms leave you free to offer the same records to another buyer later. | Exclusive terms should come with a much higher price. We flag them every time. |
| How it is paid | Often a set amount agreed before signing, paid once the buyer has received the records. | Buyers state in their terms that they pay within 7 to 60 days of delivery. |
| Who pays our fee | The buyer pays DillisData on top of the amount it pays you. | The number you accept is the number you receive from the buyer. |
Where the estimate comes from
How to put a number on your training data
AI data buyers publish pricing formulas, and the per-company payouts they list stretch from roughly $10K up past $4M. Our estimate takes four of those published formulas, feeds your answers through each, averages the results and rounds down. It is a starting point for talks, not an offer.
The real number comes from buyers competing. Two buyers can look at the same company and land far apart, because each is filling different gaps. That is why we collect several offers before you choose. If you want to go deeper on value first, our page on what AI companies pay for data goes through public prices and the factors that move them.
Price your training data in two minutes
Answer four questions. No upload, no login, no call needed to see the range.
Records to payout
How to sell AI training data, from first look to payment
The whole job takes roughly one hour of the owner's time. Everything else is handled by us or by the buyer.
Get a range
Answer four questions on size, age, industry and record types. You see a cautious estimate right away.
2 minutesLet us shop it
We describe your company to the buyers most likely to want it, collect their offers and press for better terms. Your files stay where they are.
1 to 2 weeksChoose and sign
On a single call we compare the offers line by line, including what is excluded and whether you can license again later.
One callHand over and get paid
Your admin runs a normal export and the buyer collects it directly. The buyer removes personal details, usually shows you samples, and pays you on the schedule in the contract.
7 to 60 daysOthers have done it
Companies are already paid for their training data
Public reports of real payments. None of them predicts your figure, but together they show the market is real and growing.
The reported range per company when around a hundred defunct startups were paid for their old Slack, email and Jira history.
Forbes, Apr 2026What Cielo24 reported receiving for 13 years of its own Slack, email and Jira history.
Forbes, Apr 2026Shutterstock earned this in a single year, 2023, from deals that let AI developers train on its library.
PetaPixel, Jun 2024The total value of Reddit's AI data licensing contracts, disclosed in its public filing before going public.
SEC S-1, Feb 2024Why act soon
The case for selling training data this year
We cannot tell you when prices will fall. We can tell you what pushes in that direction.
Scarce data pays best
A buyer will pay more for the first few companies in your line of work than for the tenth. Early sellers in each field get the strongest offers.
Training methods keep changing
AI companies are always looking for cheaper ways to train. Some may need fewer real business records over time, though nobody can name the year.
Stored files earn you nothing
You already pay to keep years of old mail and tickets. Licensing a copy puts that archive to work while it stays exactly where it is.
We limit new clients per month
Each search takes real work, so we cap how many owners we start with every month. A spot this month is not guaranteed next month.
Questions
Questions owners ask
Is selling data for AI training the same as selling my company's data outright?
No. What you sell is a license, meaning the right to use a copy. Ownership stays with your company, and you can usually license the same records again if the first deal is non-exclusive.
Do we have to clean or label the records first?
No. Your admin runs a standard export from the system you already use. The buyer handles the cleanup, including removing names and personal details.
I am an individual. Can I sell my own data to train AI?
We only work with businesses, usually with 20 or more staff and three or more years of records. This page is about company records, not personal data.
Do old records still count, or does training data need to be recent?
Old records count. Ten years of tickets or email is usually worth more than one year, because volume and change over time are part of what a buyer wants.
Our company closed. Can we still sell training data?
Often, yes. Closed and closing companies can qualify, as long as the records still exist and someone has the authority to sign. See our guide for closed company records.
What file formats do buyers accept?
In most cases the default export from your system works, such as PST or mbox for email, JSON for chat and CSV for tickets. The buyer converts it on its side.
Can I see what the buyer does with my training data?
The license spells out how the copy may be used, who may use it and whether it can be resold. We point out any loose wording before you sign.
What does DillisData cost?
Zero. The AI company that licenses your records covers what we charge, separately from your payment. With no signed deal, there is no charge to anyone.
Your first step
See what your records could earn as training data
Four questions, a cautious range and no commitment. If the figure is worth pursuing, we start finding buyers for you.
Free for your business. Your files stay put. Limited spots each month.