Controlling the company facts you share with AI
By Jason MurphyPublished 3 min read
By Jason MurphyPublished 3 min read
The short answer
Governing company data shared with AI starts with classifying facts by who may see them, then matching each class to the terms under which a tool may receive it, with retention and an audit trail. Whether a particular tool is safe depends on its plan terms, which differ and change.
Asking whether an AI tool is safe invites a general answer to a specific problem. The same tool can be appropriate for one class of information and unacceptable for another, and its terms differ between consumer, team and enterprise plans and between providers. Terms also change. The durable question is which facts are cleared to be shared, under which terms.
A company already distinguishes, informally, between what is public, what is shared with partners and what is internal. Governance writes it down in classes tied to a rule for AI use.
For each class the allowed destinations are named, not left to individual judgment. A destination is a provider and a plan, with its documented terms on training use, retention, access by the provider, deletion and regional processing. Those terms are read from the provider's current documentation and recorded with the date they were read, since they can change underneath a standing policy.
A policy nobody can audit is a statement of intent. The control is a record of what was shared: which fact, which class, which destination, who approved it, when. For recurring flows such as a catalog export or a dealer file sent to an assistant, the record is generated by the process rather than remembered.
The same record serves the other direction. Facts that go to AI systems should be the checked ones, with sources and dates, so what an assistant reads about the company is what the company confirmed.
When an assistant is permitted to act (read a system, draft a quote, update a record), the record also states what it may do without approval. Read access to a class, drafting and approving are separate permissions. The boundary between automation and human approval is set per class of fact, and changes only by a recorded decision.
This is the governance layer described in Business Truth: the facts, their classes and their authorities, in one place.
That depends on the documented terms of that plan, which differ from team and enterprise plans and can change. Confidential data needs a reviewed arrangement.
Classes of facts, the destinations allowed for each class, the terms read and the date they were read, an approval rule and a record of what was shared.
A recorded decision per class of fact, separating read, draft and approve permissions.
A clear classification speeds routine use, because the cleared destinations are known and only the exceptions need review.
Research report
Labs · Research
Labs · Research
An answer is accurate or not only relative to a record. Build the record first, then ask.
What AI says and why3 min read
Most of a company's public facts are writable by someone else. The first step is knowing by whom.
What AI says and why3 min read
An agent comparing suppliers cannot call to ask. It needs the answer to already be data.
What AI says and why3 min read
A connector gives an assistant a front door to a catalog. It also asks which party's catalog is the answer.
What AI says and why3 min read
Search access and training consent are separate settings. Check the one that decides whether ChatGPT Search can show your pages.
What AI says and why3 min read
One email a week: what we asked the AI, what it got wrong, and how we checked.
Our research reaches 17,000+ followers across our channels. Upstream is the weekly version.
Weekly. Unsubscribe anytime. Privacy