AI developers now buy records that show how work actually gets done: internal chat and email, source code with its history, support and service tickets, call recordings, procedures and project files. Public reports put closed-startup workspace archives at roughly $10,000 to $100,000 per company, a single startup code base at about $25,000, and a bankrupt airline's emails and chats at a $10 million bid. Value rises with depth, linkage between records and clear rights. Legal risk rises with personal information, which is why chat, email and call audio need the most preparation.
- Buyers want records of real work with the reasoning left in: problems, decisions, disagreements and outcomes.
- Linked records are worth more than isolated ones. A ticket tied to the code change that fixed it beats either alone.
- The record types with the most demand, chat, email and calls, also carry the most personal information.
Why AI companies want ordinary company records now
Large language models were trained mostly on the public web. The next step is agents: software that schedules, debugs, answers customers and completes multi-step work. Training and testing an agent needs examples of real workplaces, and those exist mainly inside companies.
Buyers building simulated workplaces for agents look for varied, deep records, including disagreements and judgment calls, MIT Technology Review reported in October 2026. One company's Slack history can be worth more than another company's Slack, email and calendars combined.
Company data types compared
Prices are public reports, not quotes. Most sales are private, so a blank price column means no reliable public figure, not that the category has no buyers.
| Data type | What buyers use it for | Public price evidence | Main legal risk | Preparation effort | Guide |
|---|---|---|---|---|---|
| Slack, Teams and email archives | Simulated workplaces and agent training | $10K to $100K per closed startup; $10M bid for an airline's emails and chats | Employee and customer personal information | High | Chat and email archives |
| Source code and repository history | Coding agents and software engineering tasks | About $25K for one startup code base | Ownership, open-source licenses, secrets in history | Medium | Source code |
| Customer call recordings | Speech recognition and voice agents | $0.15 to $0.30 per minute paid to consumers by one app | Recording consent and voice biometrics | High | Call recordings |
| Support tickets and resolutions | Customer service agents | No reliable public figure | Customer personal information in messages | Medium | Support tickets |
| Field service and maintenance records | Diagnosis, scheduling and dispatch agents | No reliable public figure | Customer addresses, OEM manuals | Medium | Service records |
| Software documentation and runbooks | Technical agents and retrieval | No reliable public figure | Third-party content inside docs | Low | Documentation |
A five-question scorecard for any record type
Score each question 0, 1 or 2. A total of 7 or more suggests a record category is worth describing to a buyer. Below 4, the preparation cost will usually exceed what a buyer pays. This is a screening aid, not a valuation.
| Question | 0 | 1 | 2 |
|---|---|---|---|
| Depth: does it show reasoning, not just outcomes? | Final results only | Some notes or comments | Problem, discussion, decision and outcome |
| Linkage: can records be connected? | Isolated files | Partly linked | Consistent IDs across systems |
| History: how many continuous years? | Under 2 | 2 to 5 | More than 5 |
| Rights: does the company clearly control it? | Unknown or disputed | Mostly, with exceptions | Documented and clear |
| Personal information: how much must be removed? | Most of it is personal | Mixed | Little or none |
What lowers value or stops a deal
- Records the company does not own: client deliverables, licensed third-party content, manuals from equipment makers.
- Promises that block a new use: privacy notices, customer contracts and employee agreements written before AI licensing existed.
- Personal information that cannot be removed without destroying the useful part, such as voice audio or long email threads between named people.
- Gaps: missing years, deleted channels or tickets without outcomes.
Where to start
Pick the one category with the highest score and describe it at company level: the system it lives in, the years covered, approximate volume and who owns it. No export is needed for a first conversation. The data inventory template gives you the columns, and how to license company data covers the full sequence.
Origin Data Partners is not a law firm. This guide summarizes public sources so you can frame questions for your own counsel. It is not legal advice.
Common questions
What kind of company data do AI companies buy?
Records of real work: internal chat and email, source code with commit history, support and service tickets, call recordings, procedures and project files. Buyers value records that show reasoning and outcomes, not just final results.
How much can a company get for its data?
Public reports range from about $10,000 to $100,000 for a closed startup's workspace archive, about $25,000 for one startup code base, up to a $10 million bid for a large airline's emails and chats. Most sales are private and depend on depth, rights and preparation.
Which company data is hardest to sell?
Data heavy in personal information, such as call audio and email between named people, and data the company does not fully own, such as client deliverables or licensed content.
Sources and further reading
Origin Data Partners is independent and not affiliated with any buyer named on this page.