Enterprise AI needs company context

On 23 October 2025, OpenAI introduced company knowledge for ChatGPT Business, Enterprise and Edu. The announcement described answers drawn from connected workplace apps, including Slack, SharePoint, Google Drive and GitHub, with citations to supporting sources.

OpenAI said the feature respected existing user permissions and was powered by a version of GPT-5. Those are the vendor’s descriptions at the time of the announcement, not a claim that packaging, supported apps or model choices are unchanged today. Verify current availability before making a procurement decision.

An assistant can retrieve a document and still misunderstand its status. A draft price list, superseded policy and approved customer agreement may all contain plausible answers. Source freshness, ownership and clear document status are therefore as important as the connection itself.

Permission-respecting retrieval is also not a guarantee of factual accuracy. A citation helps a reviewer inspect the basis of an answer; it does not automatically prove that the answer interpreted the source correctly. The model race analysis explains why buyers need evaluations on their own tasks rather than relying only on model rankings.

Start with one recurring question and the sources needed to answer it. Name the person responsible for each source, identify sensitive fields, and decide which employees should be able to see the information. Clean up access problems before making them easier to search.

Build a small set of test questions that includes an outdated document, an inaccessible document and an answer that should not exist. Review whether the assistant cites the correct evidence, admits uncertainty and avoids drawing on material the user cannot access.

Decide what happens after retrieval. Finding the answer to an invoice question is different from changing an invoice or sending a reply to a customer. Keep those actions separate until the team has evidence that the extra permissions are justified.

If a task follows stable rules and has structured inputs, conventional automation may be simpler to test and operate. AI becomes more useful when language, ambiguity or a wide range of sources makes rigid rules insufficient. More autonomy should not be the default response to every process problem.

Compare the total work, including source maintenance and human review, with the existing method. A short demonstration is not a business case. The Stena Line case illustrates the value of changing the process; enterprise agents discusses the additional decision about allowing software to act.

Assign someone to keep the sources current, someone to review failures and someone to decide when access should change. Without those roles, a promising pilot can slowly become a confident interface to outdated information.

Leadership teams can use AI advisory to prioritise suitable workflows and assess the evidence needed for expansion. For examples of products with different purposes, see My Brands; their presence in the portfolio does not establish that they use the OpenAI feature described here.