Business data licensing
Frontier AI labs pay for real examples of how businesses work to train better models. We license the emails, sales records and support tickets you choose to share, with identifying details removed and on terms you approve.
What frontier labs already pay for data
How long has the company been around?
Illustrative estimate in USD, not an offer. Actual value depends on your data and buyer demand. No sale or payment is guaranteed.
What AI can't do yet
Tested on real paid freelance projects, the best AI agents still can't deliver about 4 in 5 to a professional standard. To close that gap, models need real examples of people taking a task from start to finish. Source: Remote Labor Index, Center for AI Safety & Scale AI
Your private records
Never scraped and never trained on. They show the steps your team took, the judgment calls along the way and the checks before anything was signed off.
The public web
Scraped and trained on many times over. It shows what people say about work, rarely how the work actually got done.
What business data means
Useful data shows how your business handles a request, the judgment calls along the way and the result. Think email threads, sales histories and solved problems, not a contact list.
Slack · Microsoft Teams · Google Chat
The questions your team raises, the replies and the decisions that follow.
What a record could look like
Made-up record, not real customer data.
Gmail · Outlook
The back-and-forth behind a customer request, an approval or a project.
What a record could look like
Made-up record, not real customer data.
Salesforce · HubSpot
How a sale happened: what the customer needed, the follow-ups and why they bought.
What a record could look like
Made-up record, not real customer data.
Zendesk · Intercom
A customer's problem, the steps your team took and how it was resolved.
What a record could look like
Made-up record, not real customer data.
Notion · Google Drive · SharePoint
The instructions and internal knowledge your team relies on to get work done.
What a record could look like
Made-up record, not real customer data.
Zoom · Approved business calls
Written records of business calls: questions, answers and agreed next steps.
What a record could look like
Made-up record, not real customer data.
Asana · Jira · ServiceNow
Work captured end to end: the brief, the steps taken, the back-and-forth and the finished result.
What a record could look like
Made-up record, not real customer data.
GitHub · GitLab
Commits, code reviews and issues that explain why your software changed, not just what changed.
What a record could look like
Made-up record, not real customer data.
QuickBooks · Xero
Invoices, orders and adjustments that show how money and goods move through the business.
What a record could look like
Made-up record, not real customer data.
Threads, participants, timestamps and reply order stay intact. Only identifying details change.
Do not upload raw records during onboarding. We start with counts and screenshots; data moves only after a buyer deal is agreed.
Privacy, made simple
We remove or replace identifying details before any AI training, working with a specialist de-identification partner. Messages and documents are also reviewed for sensitive information beyond names. Your team doesn't redact anything.
An example from a work email
Jim Smith resolved a delivery issue for Northwind Ltd. Send the revised plan to [email protected].
Jim Smith confirmed the customer accepted the new date.
The same work, with replacement labels
Person 14 resolved a delivery issue for Company 08. Send the revised plan to Email 14.
Person 14 confirmed the customer accepted the new date.
Jim Smith becomes Person 14 consistently across the dataset. Company names and email addresses get labels too.
Passwords, payment credentials and other private fields are excluded, never offered for sale.
The problem, the steps taken and the result stay connected. That's what helps AI learn.
Made-up example. Replacing names is only one step: other details may still identify someone, so the agreed process checks context and sensitive fields before training.
Buyers license the records, never your name. They may use the data for model training and evaluation only. It is never published or made publicly accessible.
Why buyers want your data
Frontier AI labs look for the problem, the response and the result. Your records can help their models in three ways, limited to the uses you agree in the licence.
Show models how people solve problems, make decisions and finish real tasks.
For example
A support thread shows the problem, the team's response and the fix.
Make an existing model better at one specific job. This is called fine-tuning.
For example
Sales records help a model write more useful follow-up messages.
Use patterns from real work to generate new, made-up examples, called synthetic data.
For example
One delivery issue becomes many practice scenarios with different details.
The agreement sets when the raw dataset must be deleted and when future licensing stops. Training permission does not automatically allow public release or resale of the raw data.
Models already trained, their outputs and derivatives may remain. Deleting the dataset does not reverse completed training.
What actually happens
Your records stay in your systems during onboarding. Start with an assessment, then verify counts with screenshots. Data moves only after a buyer deal is agreed.
Your team's workloadAbout 2 hours
We guide you. Allow about two hours for your team to verify record counts. Older systems or several data sources can take longer. No raw files are uploaded at this stage.
Tell us about your company and the systems you use. No raw data upload.
Your team confirms record counts with screenshots. Records stay in your systems.
Choose the systems, folders and records to include, plus permitted uses and buyer restrictions.
The opportunity is presented to buyers. A listing is not a guaranteed sale.
After a deal, we de-identify the data and deliver it to the agreed controlled environment. Your company is never named.
You get paid for every approved sale under the agreement. A signed licence does not guarantee a sale.
Who we work with
Buyers pay for depth: years of connected records from a team doing real work. Beyond size, we need one person who can approve the scope and keep reviews moving.
Full-time staff whose daily work leaves a record behind.
Operating history in the tools your team already runs on.
An owner, CEO or president who can approve the scope and sign.
Before you decide
Straight answers
Start with counts and screenshots, not a data upload. These answers cover what to check before you commit.
Their models can explain how work should be done but still fail at most real jobs. To improve, they need real examples of people taking a task from start to finish, and that record lives inside companies, not on the public web.
No. Frontier labs aren't entering your industry. They use the records to teach models how work gets done in general. You can also exclude named buyers and competitors in your agreement.
No. Your company is not named to buyers, and the delivered dataset is de-identified. Buyers may use it for model training and evaluation only, never publication.
We do. Aba Compute manages de-identification with a specialist partner named in your agreement, before any buyer can train on the data. Your team doesn't have to clean or redact anything.
Both are possible. A dataset can be licensed to more than one buyer if your agreement allows it, and new records can be packaged as your business keeps running. Neither is guaranteed.
Often, yes. A logistics, services or retail business has delivery problems, customer conversations and sales histories that show how real work gets done. We assess the records and your right to share them, not just your industry.
No. You can exclude specific systems, folders, customers or sensitive workflows. The agreed scope defines exactly which records are offered.
Never send passwords, API keys, payment details or other sensitive records through this website. Tell us which systems you use first so we can agree exclusions.
Information about your systems, record counts and screenshots for verification. You do not upload raw business records at this stage.
Not on its own. A message or a combination of details can still identify someone, so the review goes beyond names and email addresses. Replacing names is not a guarantee that every privacy risk is removed.
Yes. Raise excluded buyers, competitors, customers or uses before signing. Those restrictions belong in the agreement.
Yes. A licence grants specified rights to use the data, not ownership. You must hold the rights to license it, including any third-party permissions.
Nowhere until a buyer deal is agreed. Then the agreed records go to a de-identification partner or the buyer's controlled environment, as set out in the agreement, and are prepared before training. Buyers never get access to your live systems.
Plan on about two hours for verification. Older systems, several data sources or extra privacy checks can take longer. We explain steps and any costs up front.
It depends on useful history, volume, quality, context, rights and buyer demand. Headcount alone can't set a price. The calculator is an illustration; an assessment gives a better-informed range.
Detailed assessment
Tell us about your company and the tools you use. We'll reply with a better-informed range and the next steps. No data upload, no passwords, no commitment.
Questions first? [email protected]
Advise companies like this? Accountants, IT providers and fractional executives can introduce a client and earn a referral fee, agreed in writing before the introduction. Become a referral partner
Buying data for an AI lab? See what we supply
