Why AI data deals are moving from publishers to operating companies

What did the first wave of AI data deals license?#

The first wave of widely reported AI data deals licensed content that was already published: news archives, forum and question-and-answer posts, stock images and video, and code. Those deals drew attention because the rights holders were well known and the content was easy to describe.

Published content suited early buyers for practical reasons. Rights sat with a small number of owners, the material was already organized and searchable, and it covered the broad range of topics that general-purpose language models needed. Many terms were never disclosed, so the public record shows that deals happened more clearly than what was in them.

For an operating company, the lesson from that wave is limited. A distributor or an HVAC contractor does not publish its order exceptions or job histories, so coverage of publisher deals says little about whether those records have a market.

What changed in what AI developers need?#

AI developers now need examples of work being done, not only examples of writing. Models are being built as agents that open tickets, schedule jobs, reconcile orders and draft responses, and training or evaluating them takes records that show a request, the decisions made along the way and the outcome.

Public text rarely contains that chain. A news archive shows how events were reported; it does not show how a dispatcher reassigned a job after a technician called in sick, how a support engineer traced a defect to a specific release, or how a quality manager closed a nonconformance report. Those sequences live in ServiceTitan, Zendesk, Jira, NetSuite and QMS records inside private companies.

That is the core of the shift. When the useful record moves from the public web into company systems, the counterparty moves from publishers to operators.

Deal categories and what drove each#

Each category of AI data deal followed a capability that developers were trying to build. The table summarizes the pattern without dates, because timing and terms varied by deal and many were never made public.

The last row differs from the others in one important way. Workflow records were never meant to be public, so the rights and privacy work falls on the company that holds them, not on a platform's terms of service or a publisher's contributor agreements.

A dated timeline of the shift, 2023 to 2026#

A dated timeline of public reports shows the shift in stages: publishers and forums licensed archives first, open platforms then tightened access, and by 2026 reporting turned to private workplace records. The entries below are third-party deals and reports, not SourceX transactions, and most financial terms were never disclosed.

Documentation expectations moved in parallel. Datasheets for Datasets was published in Communications of the ACM in December 2021, MLCommons published the Croissant 1.0 dataset metadata format in March 2024, and the Data & Trust Alliance released Data Provenance Standards in July 2024 defining Source, Provenance and Use metadata for AI training datasets. None is a law, but together they show buyers expecting provenance, rights and permitted use to be recorded for each dataset.

• 1April 2023 (forums): Reddit announces paid premium access to its Data API for large-scale commercial use, a change its CEO linked to companies training AI models on Reddit data.
• 2July 2023 (news): The Associated Press and OpenAI announce a deal under which OpenAI licenses part of AP's text archive; financial terms were not disclosed.
• 3December 2023 (news): Axel Springer and OpenAI announce a partnership covering news content from brands including Politico and Business Insider.
• 4February 2024 (forums): Reuters reports that Reddit struck a deal to make its content available for training Google's AI models.
• 5June 2025 (platforms close off): X updates its Developer Agreement to bar using its API or content to fine-tune or train a foundation or frontier model, as TechCrunch reported.
• 6October 2025 (workplace expertise): TechCrunch reports Mercor's CEO describing AI labs hiring former senior staff of banks, consulting firms and law firms because the companies themselves do not want to hand over data.
• 7January 2026 (marketplaces): Cloudflare announces it has acquired Human Native, an AI data marketplace where content owners license material for AI training.
• 8April 2026 (operating records): Forbes reports that, after cielo24 was wound down, its remaining internal chat, project-tracking tickets and emails became items for sale to AI developers.

Why operating companies hold what developers now want#

Operating companies hold records that are private, written by people doing the work and tied to outcomes, which is the combination public sources lack. In the SourceX Enterprise Data Value Framework those traits map to uniqueness, domain expertise and human-generated signal, which increase value, while reproducibility, the ease of recreating records from public sources, reduces it.

A typical fit is a US company with 50+ full-time employees at peak and several years of operating history, running connected systems such as a CRM, a help desk, a field service platform or an ERP. Software companies, engineering and consulting firms, trades, logistics businesses and manufacturers qualify on their records, not on being in technology.

Illustrative: a property software company checks whether it fits#

Illustrative: the CEO of a fictional property management software company reads coverage of publisher licensing and assumes it has nothing to do with a business that sells to landlords. Its records tell a different story: years of Zendesk tickets linked to Jira issues, GitHub pull requests and release notes, so a customer complaint can be traced to the fix that resolved it.

The company runs a metadata-only fit check, describing systems, years of history and record types without sharing files. The review flags the support-to-fix chain as the strongest package and recommends excluding documents that tenants uploaded to customer portals, which customers control. The CEO approves a rights review of that narrower scope before any export is prepared.

How SourceX approaches the operating-company side#

SourceX works on the operating-company side of this shift, helping established US companies license support conversations, CRM histories, job and dispatch records, orders and exceptions, and quality and maintenance records. Data is licensed, not sold outright, and the company keeps ownership.

Each transaction follows the SourceX five-step transaction of Supply, Rights, Preparation, Approval and Delivery. Each package carries a SourceX Evidence Packet with provenance, licensing rights, permitted use, the privacy record and release authorization, the same categories the documentation standards above ask dataset builders to record.