Key takeaways
- Coding benchmarks such as SWE-bench and SWE-bench Verified are built from real GitHub issues and the pull requests that solved them.
- The most useful records link up: a support ticket, the issue it opened, the pull request that fixed it and the incident review.
- Data customers store in your product is data you hold on their behalf, so it is out of scope.
What does each ticket, pull request and postmortem show?
| Record | What it shows | What comes out first |
|---|---|---|
| Issue tracker history | How bugs and features are reported, sized, assigned, argued over and closed, with status changes and links to code | Customer names and account IDs in issues, and logs pasted in with user data |
| Pull request discussions and code reviews | Why a change was made, what reviewers questioned, and what was merged, reworked or rejected | API keys, tokens and passwords pasted into comments or diffs |
| Incident reports and postmortems | How a production failure was spotted from alerts and logs, diagnosed, fixed and prevented next time | Affected customers, IP addresses and infrastructure details that would help an attacker |
| Product specs and design docs | How a customer problem becomes a scoped feature, with the trade-offs the team weighed | Named customers, and partner terms or roadmaps shared under NDA |
| Support escalations to engineering | The path from a user's report to a reproduced bug and a shipped fix | User names, emails, account IDs and anything copied from customer workspaces |
| Sprint and release history | What was planned against what shipped, estimates against actuals, and what slipped and why | Performance comments tied to named engineers |
Why AI companies test coding agents on real GitHub issues
SWE-bench, published by Princeton researchers in 2023, tests models on 2,294 problems drawn from real GitHub issues and the pull requests that solved them. In August 2024 OpenAI and the SWE-bench authors released SWE-bench Verified, a subset of 500 problems screened by 93 professional software developers. OpenAI's GDPval lists software developers, computer and information systems managers and project management specialists among its 44 occupations.
Public benchmarks draw on open-source projects. Private trackers hold the same kind of work on products the public never sees, which is why OpenAI's Data Partnerships programme asks for data "not already easily accessible online".
Your history is split across Jira, GitHub and the support desk
Software records are spread across the development and support stack, and much of the value sits in the links between tools.
- Jira
- Linear
- GitHub
- GitLab
- Confluence
- Notion
- Slack
- PagerDuty
- Sentry
- Zendesk
- Intercom
Is customer data in our product off limits?
For a SaaS company the main line runs between your own working records and your customers' data. Under the UK and EU GDPR, a provider hosting personal data for customers usually acts as a processor and may only use it on their documented instructions (Article 28), and enterprise contracts often add tighter limits.
- Customer data held in your product: workspaces, uploaded files, database rows and usage analytics tied to users. Data you hold on behalf of customers is out of scope
- Secrets in repositories: API keys, tokens and passwords, including ones removed from current code but still in Git history
- Anything a customer contract or data processing agreement restricts, even with names taken out
- Third-party code and documents you cannot pass on, such as vendored libraries or partner API docs under NDA
- Security reports, penetration test findings and infrastructure details
- The source code itself: codebases are licensed separately through Pangea.ai code licensing, so this page covers the records around the code
Linked history beats a pile of tickets
When we review a software company's records, connected history is worth most. A ticket that links to the issue, the Slack thread where engineers argued over the cause, the pull request, the review and the postmortem is a complete record of a problem being solved, with its own result: merged or reverted, fixed or reopened. Years matter too, because a long history shows a product changing through rewrites and migrations. The guide to what company data is worth covers what sets the price. When you're ready to sell company data to AI companies, the first step is telling us which of those tools you use.
Frequently asked questions
Can a SaaS company sell its customers' data to AI companies?
Usually not. Data customers put into your product is normally processed on their behalf, and under the UK and EU GDPR a processor may only use personal data on the customer's documented instructions. Enterprise contracts often go further. What a SaaS company holds in its own right is the record of its own work: issues, reviews, specs, incidents and internal support notes, with customer details taken out.
Is our codebase part of company data licensing?
No. Source code and its commit history are licensed separately through Pangea.ai code licensing. Company data licensing covers the records around the code: issue trackers, pull request discussions, specs, postmortems and support escalations. Many software companies hold both. Support queues on their own are covered on the customer support page.
Do AI companies want data from a small software company?
Depth counts for more than headcount. Benchmarks like SWE-bench are built from individual issues and the fixes that closed them, so a small team with years of well-linked tickets, reviews and incident write-ups can hold useful records. When we review a company's data, we look at years of history, outcomes attached and reasoning written down. The free check tells you whether yours could qualify.
Sources6
- Jimenez et al., SWE-bench: Can Language Models Resolve Real-World GitHub Issues?, arXiv, October 2023
- OpenAI, Introducing SWE-bench Verified, 13 August 2024
- OpenAI, OpenAI Data Partnerships, 9 November 2023
- OpenAI, Measuring the performance of our models on real-world tasks (GDPval occupations list), 25 September 2025
- GDPR Article 28, Processor
- GitHub Docs, About secret scanning (scans entire Git history for credentials), checked 20 October 2026


