When AI uses another program: reading or making changes?

AI can suggest an action, but the service carries it out

You may have asked AI to check whether information is up to date or whether something is available in another service. In that case, it is no longer answering only from what it has learned: it is using a tool it was given to retrieve information. An application can give AI a limited tool such as search, a calendar, or a database query. AI does not grant itself access; the system owner defines it.

Try your first integration with limited permissions

1. 1

   Choose a read-only task

   Try a task that only retrieves or summarises information, such as checking whether a book is available at the library or summarising a made-up table. Give AI a clear request and check whether its answer is accurate and up to date.

2. 2

   Use sample data

   Create a small fictional calendar or table and give it to AI to work with. Use made-up information, not your work, customer, or bank account, so you can see directly how it reads and interprets the content.

3. 3

   Grant only the permissions needed

   Ask AI what permissions this particular task actually requires, such as read-only access without write access. Compare the answer with the task: a good suggestion asks only for the permissions the task requires.

4. 4

   Approve consequential actions yourself

   Read the proposed change or message all the way through. Is the time, recipient, or content exactly what you intended? When it is correct, approve the action yourself before it is carried out.

Design a tool for me that checks whether a particular book is currently available to borrow from my local library. The tool may only search the library’s public catalogue by book title. It must not place a hold or change anything. Explain how it would work in practice: what goes in, what it returns, what permissions it uses, and when the final decision to place a hold remains mine rather than the AI’s.

I am considering giving AI access to my calendar so it can suggest suitable meeting times. Before I do, ask me clarifying questions to help me decide what AI may read in my calendar, what it must never change on its own, and when it should always ask for my permission first.

How can you spot a higher-risk action?

• Reading information: for example, checking whether a book is available at the library. Still check the source and when it was updated.
• Creating a draft: for example, suggesting a calendar entry. A person can still edit and approve it.
• Making a change: for example, sending a message, placing a hold, deleting something, or making a purchase. This needs clear approval and limited permissions.

Use a fictional calendar only. Find two possible times next week and show them as a list. Do not book anything or send an invitation.

Once you know the basics

The next step is to describe the workflow before building it: what information is read, what is suggested, what happens only after a person approves it, and how to recover from an error. The guide “How an AI project takes shape — from idea to finished product” walks through this planning stage. Treat every new permission as a separate decision.