OpenAI's company knowledge makes the retrieval process a visible frontend surface

The toggle is the tell

Company knowledge is not always on. To use it, you "tap 'Company knowledge' under the message composer," and OpenAI is explicit that this is a manual step you take when you start a new conversation.

That single design choice carries weight. OpenAI acknowledges that when the toggle is off, "ChatGPT may still use your connected apps to help answer questions, but its responses won't include the same depth or detailed citations." So the same connected data produces two different experiences depending on a button state—one shallow, one deep with sources.

For a frontend, that's a hard thing to explain to users. The company says it is "working to bring these experiences together in the coming months," which is a fair admission that the current state asks people to remember a mode switch to get the good version.

A sidebar that narrates the search

The most concrete UX decision here is making retrieval observable. As company knowledge runs, "you can follow along in the sidebar to see what it's looking at and how it's using the information."

When it finishes, the interface shows "exactly what sources it used to inform its response, along with the specific snippets it drew from," and any citation can be clicked to "open the original source for more details."

This is the part that separates a trustworthy internal tool from a plausible-sounding one. Surfacing the snippet, not just the source name, lets a user verify the claim without leaving the answer. It treats citation as an inspectable artifact rather than a footnote.

Mode exclusivity is the current cost

The toggle also gates capabilities against each other. OpenAI states plainly: "When company knowledge is on, ChatGPT can't search the web or create charts and images."

The escape hatch is designed into the conversation: "You can turn it off and continue working in the same conversation to use these capabilities, while keeping your existing context." Context persists across the mode flip, which softens the friction, but the user still has to know to do it.

So a request like "report on customer feedback from the mobile launch and chart the sentiment" can't happen in one pass today. The interface asks the user to sequence the work manually across two states.

Permissions render invisibly, and that's the point

One thing the frontend deliberately does not show is data the user isn't allowed to see. "Company knowledge respects your existing company permissions, so ChatGPT only has access to what each user is already authorized to view."

This means two people in the same company can ask "Where did we land on company goals for next year?" and get answers drawn from different underlying sources. The interface never signals what was filtered out—by design—because surfacing the existence of restricted content would itself leak information.

OpenAI pairs this with admin controls: custom roles, group-level permissions, SSO, SCIM, IP allowlisting, and conversation logs via the Enterprise Compliance API. The user-facing surface stays clean precisely because the access boundary is enforced upstream.

What the visible sidebar commits OpenAI to

By showing the search as it runs and exposing the exact snippets behind each answer, OpenAI has set a bar for its own product. A feature that resolves "conflicting details" and "highlights differing viewpoints" is only credible if users can trace each viewpoint to its source—which is why the citation UI matters as much as the GPT-5 variant behind it.

The near-term work is unifying the toggle so the deep, cited experience stops being a separate mode the user has to opt into. Until then, the honest read is that OpenAI built the trustworthy version first and is now working to make it the default rather than the exception.