The business wants to try it tomorrow. IT needs to decide today.

Imagine this message arriving from sales:

Can we use this AI for meeting notes? We’d like to try it at tomorrow’s meeting.

It sounds useful. You understand how much time meeting notes take. But that one sentence is a difficult basis for saying yes.

Will customer information go into it? Who will read the notes? Could someone connect a shared drive along the way? The questions quickly multiply.

A long questionnaire as your first reply could also bring the team’s work to a halt.

Start with three questions: What goes in? How will the answer be used? What can the AI do?

1. “What information will you put into it?”

A recording of a sales meeting. It includes customer names and a project that hasn’t been announced yet.

Now there is something concrete to review. The product name alone did not tell you what information the team intended to enter.

You do not need the actual data at this stage. A description such as “sales meeting recordings containing customer names” is enough to begin.

Check data handling for the actual subscription plan: training use, retention, deletion, who can access the information, and where it is sent. Compare the provider’s explanation and administrative settings with your internal information-handling rules.

“Not used for training” does not, by itself, explain retention or access. Record unknowns as questions for the provider.

2. “What happens to the answer afterward?”

Ideally, we’d summarize the meeting and send the notes straight to the customer.

That reveals another review point: AI-generated text will reach someone outside the company.

For meeting notes, agree on a process for checking prices, deadlines, and decisions against the recording or original notes. A plausible piece of writing still needs to be checked for an accurate account of the meeting.

An initial trial could limit use to drafting, with a person reviewing before anything is sent. Specify who checks it and what they compare it with, so the team can picture the work involved.

The same AI service can need different checks for internal brainstorming and customer-facing documents. Ask applicants to describe how they will use the output.

3. “What can the AI access or operate?”

Apparently, we can connect our shared drive too. It could find related documents for us.

Useful, but worth looking at the settings together.

Which folders can it read? Could it access documents the user should not see? Is access read-only, or can it edit and send information externally? Review each connection together with the actions allowed there.

Instructions embedded in external documents can also steer an AI’s behavior. This problem is called prompt injection. OWASP’s explanation covers manipulation through external content.

OWASP’s Excessive Agency guidance discusses risks from excessive functionality, permissions, and autonomy. Its recommendations include limiting permissions to necessary operations and requiring human approval for high-impact actions.

If the trial does not need a shared-drive connection, starting without it is an option. Review the access scope when a connection is added later.

Record what “yes, for this trial” means

Before approving, also check the relevant internal policies, contract terms, and information classification.

Once those checks are complete, record the approved use, rather than simply marking the tool “allowed.” For a meeting-notes trial, you could write:

Trial: Draft meeting notes; limited to the sales team’s trial participants.
Input: Internal test recordings approved for this evaluation. No customer information.
Output: The owner checks it against the original. It is not used for sending notes to customers.
Connections: No shared-drive connection or automatic sending.
Management: Record the subscription plan, settings, dates of reviewed documents, and usage owner.
Review: The trial ends next Friday. Review again before using real sales meetings or expanding connections or purposes. Stop use and contact IT if a problem occurs.

To start from a blank note, copy these six lines:

Purpose and users:
Permitted input:
Output use and reviewer:
Connections and permitted actions:
Subscription plan, reviewed documents, and usage owner:
End date, review triggers, and incident contact:

The note records what was allowed and provides a basis for revisiting the decision when the use changes.

IT does not have to carry every request alone

Working through one request may reveal a wider issue.

Perhaps you do not know which departments use which tools. Perhaps a guideline exists, but nobody has agreed how to answer requests. Or every new use brings the same checks again. Those are reasons to consider an inventory and rules that fit actual workflows.

U-Rec’s AI Security Consulting offers an AI Security Assessment covering an inventory of internal AI tools and risk prioritization, as well as the creation and revision of AI usage guidelines. The service page describes the support available.

Start with one request you are unsure how to answer. What goes in? How will the answer be used? What can the AI do? Answering those three questions can make it clearer whom to ask and what to check next.

References

Accessed October 23, 2026.