Training: off by default, with one exception
OpenAI’s enterprise privacy page (updated January 8, 2026) covers ChatGPT Business, Enterprise, Edu, Healthcare, Teachers and the API platform. Its answer for all of them is the same: by default OpenAI does not use your business data to train its models. The one exception is when you explicitly opt in to share data, for example through OpenAI’s opt-in feedback mechanisms. Data that ChatGPT reads from connected apps is not used for training by default either.
The Business overview says it in one line: “OpenAI won’t train on your workspace’s data.” Compare that with personal plans: on Free, Go, Plus and Pro, content is used for training unless you opt out.
What OpenAI keeps, and for how long
| Question | ChatGPT Business |
|---|---|
| Who sets retention | Workspace admins |
| Deleted or unsaved conversations | Removed from OpenAI systems within 30 days |
| Exceptions to the 30 days | Legal requirement, or reasonably necessary to protect the service or third parties from harm |
| Deleted saved memories | Logs may be kept up to 30 days for safety and debugging |
| Deleted projects | Chats, project-only files and instructions removed within 30 days |
| Who inside your company can see chats | The user; admins not all private chats by default, export of member data where available |
| Who at OpenAI can see chats | Authorized staff for support, abuse and legal compliance; bound contractors for abuse review |
Sources: OpenAI’s enterprise privacy page, Memory and Projects help articles, read in November 2026. OpenAI also notes that shorter retention can degrade features like conversation history.
Security and paperwork
- Encryption: AES-256 at rest, TLS 1.2+ in transit.
- Audit: ChatGPT Business completed a SOC 2 Type 2 audit.
- Sign-in: SAML SSO is part of OpenAI’s business offering.
- DPA: OpenAI will sign a Data Processing Addendum for ChatGPT Business to support GDPR compliance, through its DPA form.
- Ownership: you keep the rights to your inputs and own the outputs, as far as the law allows.
What Business does not include: Zero Data Retention, BAAs for health data, and a choice of storage region. Those sit with contracted plans such as Enterprise. See where ChatGPT Business data is stored.
The common mistake: “no training” means “private from everyone”
No training by default means OpenAI does not use your workspace content to train its models unless you explicitly opt in. It does not mean nobody can read it. Your workspace owner may be able to export your data where that option is available, OpenAI staff can access them in the cases listed above, and automated safety classifiers may run over business data. The right mental model is a company system, not a diary.
What it means for your team
The privacy settings are workspace-wide, but knowledge is not. Private chats and personal memories are not automatically shared; shared projects can make selected context available to colleagues, but the rest stays with each person, and when that colleague leaves, the context sits in an export rather than in anyone’s working tool. If you are deciding between personal and business plans, see ChatGPT Team vs Plus.
Where Bombos fits
Bombos is a plugin that gives every AI tool your team uses one shared memory, including learned working rules and existing knowledge from mail, documents and repos, kept where you choose: in the cloud, or standalone on your own server with your own models. Every answer is traceable to what it rests on, so you can see which team knowledge went into it. One person can try it first for their own agents.
Related questions
Can my workspace admin read my chats?
Not by default. OpenAI's privacy page says Business admins can view, access, export and delete conversations, but its help center says that by default admins and owners cannot see all private member chats, and member export works only where available.
Can OpenAI staff read our conversations?
Only in limited cases: authorized employees for engineering support, abuse investigations and legal compliance, and bound contractors reviewing for abuse and misuse.
What counts as opting in to training?
OpenAI names its opt-in feedback mechanisms as an example of explicitly sharing data, after which that shared data may be used for training.
Does this also cover custom GPTs and projects?
Yes. OpenAI says the same commitments apply to GPTs in a Business workspace, and that it does not use Business workspace content, including shared projects, to train models by default.
