OpenAI embeds company knowledge

ChatGPT is turning into more of a work operating layer than a blank chat box. That’s the real story here. OpenAI now has a formal “company knowledge” feature for ChatGPT Business, Enterprise, and Edu that lets people pull context from connected internal apps directly into a conversation, instead of copying links, pasting notes, and stitching answers together by hand. It’s live in OpenAI’s product docs and product pages, and the pitch is simple: ask one question, get an answer grounded in your company’s own tools. ### What is “company knowledge” actually? Basically, it’s a mode inside ChatGPT that searches across the work apps your organization has connected and uses that material to answer questions with citations back to the source. OpenAI describes it as a way to bring organization-specific context into ChatGPT so answers reflect your company, your projects, and your documents rather than just the model’s general knowledge. (help.openai.com) ### What kinds of apps are involved? OpenAI’s examples include tools like Slack, SharePoint, Google Drive, and GitHub. The help docs also make clear that this is tied to the broader “apps” system — what OpenAI used to call connectors — and that custom apps built with MCP can also feed in company-specific data if they support the right actions. ### How do people use it in practice? Users turn it on inside a specific chat. (help.openai.com) On web, there’s a “company knowledge” option under the composer for new conversations, or in the tools menu for existing ones. Then they pick which connected apps to include. The model searches those sources, pulls relevant context, and returns an answer with citations and source snippets so the user can inspect where each claim came from. (openai.com) ### Why does this matter more than a normal file upload? Because file upload is one-shot context. This is cross-system retrieval. If someone is preparing for a customer call, ChatGPT can combine Slack messages, docs, email context, and support tickets into one briefing instead of forcing the person to hunt through four systems first. That’s a much bigger workflow claim — less “summarize this PDF,” more “understand what my company knows about this situation.” (help.openai.com) ### What’s the permissions story? OpenAI is leaning hard on existing access controls. The company says company knowledge respects the permissions a user already has, and admins can manage app access at the workspace level. Enterprise and Edu admins can control access with RBAC, while Business has apps enabled by default. Users may still need to authenticate each app the first time they use it. (openai.com) ### Where are the limits? The biggest one is platform support. Right now, OpenAI says company knowledge is available on ChatGPT web, but not on the Windows or macOS desktop apps or on Android and iOS. There are also app-level requirements — an app needs file search plus search and fetch actions to qualify. So this is not “everything in your stack instantly works.” It depends on the connector layer being set up correctly. (help.openai.com) ### How does this fit the bigger product push? It lines up with OpenAI’s broader move toward workplace tooling: apps, shared projects, spreadsheet integrations, admin analytics, and workspace agents that can act across connected systems. Company knowledge is the read layer in that stack — the part that helps ChatGPT see across the company before agents start doing more inside it. ### So what’s the catch? (help.openai.com) Freshness and governance. A cited answer feels authoritative, but only if the right sources are connected, permissions are set correctly, and the system is pulling from the current source of truth. If a team’s decisions live half in Slack, half in docs, and half in somebody’s head, ChatGPT can surface the mess more efficiently — but it can’t magically clean it up. (help.openai.com) ### Bottom line? This is OpenAI trying to make ChatGPT useful at the exact moment work gets messy — when the answer exists somewhere inside the company, but nobody wants to go find it. That’s powerful. But the value won’t come from the model alone. It will come from whether a company’s apps, permissions, and internal knowledge are organized well enough for the model to trust. (help.openai.com)