01Introduce AI agents
What Counts as an AI Agent? A 30-Second Test for Small Businesses
A 30-second test to tell whether an AI tool is an agent: can it act, start by itself, act without approval, reach your accounts?
- Topic
- Introduce AI agents
- Written by
- AI Ground Rules
- Last reviewed
- Sources
- 10 checked sources
An AI tool counts as an agent when it can take an action in your accounts. Sending an email, moving a file, changing a record or booking a meeting are actions. It still counts if the tool waits for you to click “approve” first, because the tool does the action, not you. Writing a draft that you then copy and send yourself is not an action. If any tool in your business can act, check 07 “AI agent permissions” of the AI Use Checklist applies to you. If none can, you can mark it N/A. The test below takes about 30 seconds per tool.
The 30-second test
Ask four questions about each AI tool your business uses. Answer for the way it is set up in your business today, not for what the product could do.
1. Can it take actions? Can the tool itself do any of these, even if it asks you to approve first:
- send an email, message or calendar invite
- move, rename, share or delete files
- create or change a record, such as a client, invoice or task
- book, buy, refund or spend anything
- post publicly, such as on social media or your website
2. Can it start by itself? Does it run on a schedule or a trigger, for example every morning, when an email arrives or when a file is added to a folder?
3. Can it act with no one approving? Can it send, change, delete or spend without a person clicking approve first?
4. Is it connected to your accounts? Has anyone given it access to your email, files, calendar, client system or another app?
How to read the answers:
- Yes to question 1: it is an agent. Check 07 applies. This includes tools that only act after someone clicks approve. Check 07 is where you test that approval step.
- Yes to question 3: it is an agent that can act with no one watching. This is the higher-risk kind. Put it first when you work through check 07.
- Yes to question 2, but every action waits for approval: it is an agent that starts by itself. Check that the approval request reaches a person, and that nothing happens if nobody answers.
- Yes to questions 2 and 4, no to question 1: treat it as an agent. A tool that starts on its own inside your accounts is acting on your behalf, even if it only writes drafts today. One changed setting can let it do more.
- Yes to question 4 only: it reads your data. It is not an agent, but it still belongs in your AI inventory, because a tool that only reads can still expose client information.
- No to all four: it just answers. Check 07 does not apply to it.
If you cannot answer a question, find out before you mark N/A. Ask whoever set the tool up, or your IT provider. “I don’t think so” is not the same as “no”.
Common tools, sorted
Most small firms have more in the middle column than they expect. The same product can sit in different columns depending on how it is set up, so check your own settings.
| Tool | Just answers | Reads your data | Takes actions |
|---|---|---|---|
| Chat assistant (ChatGPT, Copilot Chat, Gemini, Claude) with no connected apps | Yes. You paste text in and copy answers out | ||
| Meeting note-taker | Yes. Joins calls, records or transcribes them, and saves notes | Sometimes. Some send summaries or follow-up emails on their own. Check the settings | |
| Copilot or Gemini inside Outlook, Word, Gmail or Docs, used to draft and summarise | Yes. Drafts and summaries draw on your mail and files. You still click send | Some modes can. For example, Microsoft says Copilot Cowork “carries out tasks across your Microsoft 365 environment on your behalf.” See the Microsoft 365 Copilot checklist | |
| ChatGPT with connected apps | Yes, when connected apps can only read | Yes, when they are allowed to write. In ChatGPT Business, admins can set plugin permissions to “Always ask” or “Allow read actions” | |
| ChatGPT Work (OpenAI now points users here for the multi-step tasks agent mode used to do. Check OpenAI’s ChatGPT agent help page for the current status) | Often. It runs longer tasks across connected apps. Check what your workspace lets it do before you mark N/A | ||
| Microsoft 365 Copilot agents and Copilot Studio agents | Yes, if built only to answer questions from your files | Yes, when given actions or agent flows. With an event trigger, an agent acts “in response to something happening,” such as a new file or a set schedule | |
| Google Workspace Studio flows | Yes. Flows start from an event, a schedule or by hand. Google warns that test runs “will send messages, update files, and set up meetings” | ||
| Zapier or Make with AI steps or agents | Yes, for search-only actions | Yes. Zapier Agents can use actions that “can change data” across thousands of apps. Zaps run on triggers | |
| Browser agents (AI that clicks and types in a web browser for you) | Yes. They can use the sites and accounts the agent’s browser is signed in to and allowed to use | ||
| Coding agents (for example Claude Code or Codex) | Yes. They edit files and run commands. Claude Code has permission modes. Its Manual mode asks before most file edits and commands. Other modes approve actions automatically, and a session may start in one of them. Check which mode is active |
(Sources: OpenAI Help Center, plugins and apps admin controls; OpenAI Help Center, ChatGPT agent; Microsoft Learn, Copilot Studio event triggers; Google Workspace Studio Help; Zapier Help; Claude Code Docs, permission modes. All checked 2026-11-14.) Products change often. Treat this table as a starting point and check what your own setup allows.
Fictional examples
| Firm and tool | Test result | Check 07 |
|---|---|---|
| Northgate Advisory uses ChatGPT Business. Staff paste in text. No connected apps are turned on | Just answers | N/A for this tool |
| Harbor Accounts uses Copilot in Outlook to draft replies. Staff read and send each one | Reads your data | N/A, but list it in the inventory |
| Lantern Talent connected a note-taker to the recruiters’ calendars. It joins interviews and emails a summary to every attendee | Takes actions. It sends email on its own | Applies |
| Ashby Lettings connected an AI assistant to the shared mailbox. It drafts replies and sends each one after a staff member clicks Approve | Takes actions, after approval | Applies. Check 07 tests that the approval step really works |
| Fernway Plumbing has a Zap that reads new web enquiries and uses an AI step to draft a quote in the job system | Runs on a trigger, connected to accounts, changes a record | Applies |
| Calder & Rowe Solicitors built a Copilot Studio agent that answers staff questions from the firm’s precedents library. It has no actions or triggers | Reads your data | N/A, but list it in the inventory |
Lantern Talent is the common surprise. Nobody thought of the note-taker as an agent. But it sends email to people outside the firm, every time, with no one checking first.
If you have one
You do not need to switch it off. You need five things in place.
- An owner. One named person is responsible for the agent, plus a backup. If nobody will own it, pause it.
- Least access. It reaches only the mailboxes, folders and systems its task needs. Turn off any write access it does not use. OWASP’s guidance on excessive agency advises giving an agent only the tools and permissions it needs. (Source: OWASP LLM06:2025 Excessive Agency, checked 2026-11-14.)
- Approval for risky actions. Anything that leaves the business, changes what others rely on, spends money, grants access or is hard to undo waits for a person. The human approval checklist walks through which actions, who approves and what they see.
- A tested approval step. Try it with fictional data. Check that the agent really waits for approval. Then leave a request unanswered and check that nothing happens. An agent that carries on when nobody responds has no approval step.
- A way to pause it. Know who can stop it, where the switch is, and test it once. In Microsoft 365, admins can block agents in the admin center. (Source: Microsoft Learn, Manage agents in the Microsoft 365 admin center, checked 2026-11-14.)
Record all five in your AI inventory, with the “Reads only / Can act” column set to “Can act”.
Why “ask me first” in the instructions is not a control
Many agents have a box for instructions. It is tempting to write “always ask before sending” there and call it done. That is not a control, for three reasons.
- Instructions can be misread. The agent decides how to apply them, and it can get that wrong.
- Content can override them. An agent reads emails, documents and web pages. Any of these can contain text telling it to do something else.
- Trained caution is not a setting. In 2025, OpenAI said its ChatGPT agent was “trained” to ask for confirmation before actions such as completing a purchase or sending an email. (Source: ChatGPT Agent System Card, July 2025, checked 2026-11-14.) That helps. But it is the agent’s behaviour, not a switch you set and can test.
Microsoft’s documentation makes a related point. Copilot Studio event triggers use the agent maker’s own sign-in, so other users of a published agent may reach data and actions through the maker’s access. (Source: Microsoft Learn, Copilot Studio event triggers, checked 2026-11-14.) What an agent can do depends on the accounts and permissions behind it, not on what it was told.
A real control is something the agent cannot get around: a permission it does not have, or an approval step built into the tool. If the agent’s mailbox connection has no send permission, it cannot send, whatever an email tells it.
Questions
We only use ChatGPT or Copilot Chat to write drafts. Can we mark check 07 N/A? Yes, if no connected apps can take actions (even after someone approves), no agents or multi-step task modes (such as ChatGPT Work) are in use, and staff copy and send everything themselves. Check the connected apps settings first. In ChatGPT Business, connected apps are on by default.
A staff member might be using an agent we don’t know about. What then? Run the 20-minute check in the AI inventory template. It finds AI apps already connected to your email, files and calendar.
It only drafts. Is that really not an agent? Drafting alone is not acting. But if it drafts on a schedule or trigger inside your accounts, treat it as an agent and record it. The next update or setting change may let it send.
Is this the same as the human approval checklist? No. This guide answers “do we have one?” The human approval checklist is the next step: which actions need a person, and how to make that approval real.
02Evidence
Sources
- Microsoft Learn, Event triggers overview (Copilot Studio)Checked
- Microsoft Learn, Manage agents in the Microsoft 365 admin centerChecked
- Google Workspace Studio Help, Get started with Google Workspace StudioChecked
- OpenAI Help Center, ChatGPT agentChecked
- OpenAI, ChatGPT Agent System Card (17 July 2025)Checked
- OpenAI Help Center, Admin controls, security, and compliance for plugins and appsChecked
- Zapier Help, Use actions on Zapier AgentsChecked
- Claude Code Docs, Choose a permission modeChecked
- Microsoft Learn, What is Microsoft Copilot? (license comparison)Checked
- OWASP GenAI Security Project, LLM06:2025 Excessive AgencyChecked