# What an AI agent can and can't do for you

Agents work with your permissions and only in your workspace. What they can do, what they never do, when a person must approve, and what they remember.

An AI agent is a capable assistant, not a free agent. It works inside the same limits you do, and some things always stay with a person.

## It acts as you

When you chat with an agent, it has **exactly your permissions**: what your roles allow you to see and change, and nothing more.

- If you can't approve invoices, neither can the agent when you ask it to.
- A step it isn't allowed to take fails with **Permission denied**, and it tells you.
- Ask it "what can you do?" to hear what it can do for you. The **Permissions** tab in the chat explains the same.

## What it can do

- **Find and explain**: look up records, answer questions about them, summarise what's going on.
- **Do the module's work**: create and update records, fill in forms, run the module's actions, as far as your role allows.
- **Read what you give it**: photos, PDFs, spreadsheets, emails and screenshots you attach.
- **Read the knowledge base and the module's help** to answer how-to questions.

## What it never does

- **Reach another company's data.** An agent works in your workspace only, for the whole conversation. It can't be pointed at another workspace, even if asked.
- **Go beyond your role.** It has no permissions of its own when it works for you.
- **Type passwords, card numbers or one-time codes** on your screen. Those are always yours to enter.
- **Press Save, Submit, Send or Delete on your screen without you.** When it fills in a form for you, you check it and click **Allow**. See [Shotgun Assistant: fill in a form](article:e1c1bbeb-58ac-43b7-8b3d-6914203b8540).

## When a person must approve

Some actions always wait for a person, such as actions that change a computer at your site through the Bridge. The agent pauses and tells you it's waiting. The paused action shows in a notice at the top right of the screen, where you click **Resume** to go ahead or **Abort** to cancel it. An approval covers that one action only.

## Agents that run by themselves

Agents can also run on a schedule or when something happens, such as every morning or when a new order arrives. Then the agent works under **its own identity**, with roles your Workspace Admin gives it. Without a role it isn't allowed to run on its own. See [Schedules](article:2233fce3-ea1d-491c-9b70-da85d84c4a87).

## What it remembers

- **Your recent conversation** with each agent, for 24 hours.
- **Notes it saves about how you work**, such as a preference or a correction you gave it. They are yours alone and carry over to later conversations, so you don't repeat yourself.

## Check its work

AI can be wrong. Before you rely on an answer that matters, click **▶** on the step cards to see what it looked up, and check anything it created. Tell it when it's wrong: it uses the correction.


Source: https://datavisions.ai/docs/everyday-work/what-an-ai-agent-can-and-can-t-do-for-you
