Skip to Content

Your AI Vendor Receives What Someone Connected

Not What Anyone Decided
November 19, 2026 by
Luis Roberto Aguirre Salazar

In most mid-sized companies, the AI in use was not built in-house. It was bought, subscribed to, or switched on inside software the company already had. That is usually a sensible choice. It also means that every one of those uses involves handing company information to someone else.

The decision about what to hand over tends to happen somewhere nobody thinks of as a decision. Someone sets up the tool, a screen asks for access to a mailbox, a shared drive or the customer database, and they click to continue so the tool can do its job. From that moment, the vendor receives whatever that access allows, which is often far more than the use needs.

A policy defining what information must stay inside the company is a good start, but it is written once, in general terms. Here is how to apply it use by use.

This is the fourth part of our series on AI accountability for mid-sized companies. The first three covered classifying each use by what it decides, designing the human review, and giving affected people a way to ask. This one moves to the boundary of the company: what leaves it, to whom, and on what terms. It adds three columns to the register: the vendor, what information it receives, and whether its conditions were reviewed, with the date.

The vendor is not always the one you think.

The first column looks trivial: write down who the vendor is. In practice, it is easy to fill in wrongly. When an AI feature is switched on inside an existing product, the information may go to the company that sells the product, to a separate provider whose model powers the feature, or to both. A meeting assistant, a writing aid in the email client, a summary button in the CRM: each may send content somewhere the original contract says little about.

The useful question is not "which tool is this?" but "who ends up holding our information when we use it?". If the vendor’s documentation does not make that clear, the gap itself is worth recording.

A row that says "not confirmed" is more useful than a row that names the wrong company.

What it receives, not what it was meant to receive

The second column asks what information the vendor receives. The temptation is to write down what the use needs: "customer emails, for drafting replies". The column should record what the access allows: "the whole shared mailbox, attachments included".

The gap between the two is where most surprises sit. Access is usually granted at the level of a folder, a system or an account, because that is how set-up screens offer it. Few tools ask for exactly the documents a task requires. Over time, more is added to the folder, the mailbox grows, a new field appears in the system, and the vendor’s reach grows with it, without anyone granting anything new.

My recommendation is to fill in this column by looking at the access settings, not by asking the person who uses the tool what they upload. And where the access is broader than the use, to narrow it before doing anything else in the register: a dedicated folder, a restricted account, a field left out of the connection. It may be the cheapest control the company ever puts in place.

Conditions reviewed, and when

The third column is a yes or no, with a date. Has anyone read what the vendor says it will do with the information?

The questions worth asking are practical, and they do not depend on any country. Does the vendor keep the information, and for how long? Does it use it to improve its own models or services? Who else processes it on the vendor’s behalf? Can it be removed when the contract ends, and how would the company know? What happens to all of this if the terms change?

If a vendor cannot answer these questions, or answers them only in general marketing language, that is also an answer. It does not have to end the relationship, but it should shape what information that use is allowed to receive.

Whether the answers are acceptable is the company’s decision, and it will vary by use: the same answer can be fine for marketing copy and unacceptable for payroll data. The column does not make that decision. It records that someone looked, and when.

The date matters because terms change. Vendors update their conditions, add features and change the providers behind their products. A review done long ago may describe a service that no longer works that way. In my view, the review should be repeated when the vendor announces a change to its terms, when a new AI feature is switched on, and in any case on a fixed cycle the company chooses for itself.

Where this leaves the register

With these three columns, each row of the register now says what the use decides, who reviews it, how an affected person can ask, and where the information goes. That last part is often the one a company can act on fastest: narrowing an access or reading a set of terms takes far less than redesigning a process.

A practical way to start, take the uses already in the register, open the access settings for each one, and fill in the second column first. It is the column most likely to change what the company does next.

The next and final part of the series covers what happens when something goes wrong: who manages each use, who approved it, and where incidents are recorded. It closes with the complete register as a download.

Rhisco designs AI governance architecture to ensure positive outcomes when deploying AI. If your company has AI tools in use and no clear picture of what information they receive, we can help you build it: www.rhisco.com/services

This article was co-created with the assistance of artificial intelligence under strict supervision, editing, and verification of our team.


The AI Grounding a Board Needs Is Smaller Than It Fears