AI serves multiple companies—how do we keep their data separate? The first line of defense for enterprise AI (reprinted from an article on our company’s official WeChat account)
A screenshot raised a troubling question: When an AI assistant for one company answered a question from Company B’s customer about renewing a contract, it casually mentioned the contract amount for Company A. The trust of both customers nearly went up in smoke. The problem wasn’t with the model itself, but with the isolation boundary further upstream.
Enterprise AI isolation usually involves three dimensions: companies—facts, files, leads, and deliverables must be kept separate and organized independently for each company; workspaces—different business lines and projects within the same company must each have their own work data and context; and accounts—the same person often manages multiple companies, so when switching between them, the system must ensure that “the person stays within the boundary,” rather than dumping all the data into one big pool. Capabilities can be shared, but facts must stay with their owners.
There are already precedents in the industry: In March 2023, OpenAI disclosed a ChatGPT incident in which a flaw in a client library briefly caused some users’ chat titles and even certain payment information to appear in other users’ interfaces. In April 2023, multiple media outlets reported that Samsung tightened its use of external AI tools after employees pasted internal code into them. If tenant boundaries are added only after the fact, the risk never goes away.
Before evaluating a system, ask one question: Along what dimensions does it isolate data? If it isolates data only by account, the boundary starts to blur as soon as one person manages multiple companies.
This does not constitute investment advice.
A screenshot raised a troubling question: When an AI assistant for one company answered a question from Company B’s customer about renewing a contract, it casually mentioned the contract amount for Company A. The trust of both customers nearly went up in smoke. The problem wasn’t with the model itself, but with the isolation boundary further upstream.
Enterprise AI isolation usually involves three dimensions: companies—facts, files, leads, and deliverables must be kept separate and organized independently for each company; workspaces—different business lines and projects within the same company must each have their own work data and context; and accounts—the same person often manages multiple companies, so when switching between them, the system must ensure that “the person stays within the boundary,” rather than dumping all the data into one big pool. Capabilities can be shared, but facts must stay with their owners.
There are already precedents in the industry: In March 2023, OpenAI disclosed a ChatGPT incident in which a flaw in a client library briefly caused some users’ chat titles and even certain payment information to appear in other users’ interfaces. In April 2023, multiple media outlets reported that Samsung tightened its use of external AI tools after employees pasted internal code into them. If tenant boundaries are added only after the fact, the risk never goes away.
Before evaluating a system, ask one question: Along what dimensions does it isolate data? If it isolates data only by account, the boundary starts to blur as soon as one person manages multiple companies.
This does not constitute investment advice.