01Introduce AI agents
AI Inventory Template for Small Businesses: Tools, Connected Apps and Agents
Find the AI already connected to your email, files and calendar, then list it: a 20-minute check, a short and full inventory table, and a worked example.
- Topic
- Introduce AI agents
- Written by
- AI Ground Rules
- Last reviewed
- Sources
- 7 checked sources
An AI inventory is a list of every AI tool, connected app and agent that can reach your business’s information. One row per item. It records who owns it, who uses it, what it can reach, whether it only reads or can also act, and when it will next be checked. Start with the 20-minute check below. It finds the AI that is already connected to your email, files and calendar.
The 20-minute check: what AI is already connected to your email, files and calendar
Do this with whoever holds admin access. Write down every AI app you find, even if you think it is harmless.
- Google Workspace. In the Admin console, go to Security > Access and data control > API controls > Manage App Access. Under Accessed apps, select View list. Google shows how many users use each app and which Google services it uses, such as Gmail, Drive or Calendar. You can set each app to Trusted, Limited, Specific Google data or Blocked.
- Microsoft 365. In the Microsoft Entra admin center, go to Enterprise apps > All applications. Select an app, then Permissions. The User consent tab shows access that individual staff granted. Microsoft notes you cannot revoke user consent in the portal; that needs PowerShell or Microsoft Graph, so ask your IT provider. To stop new grants, go to Enterprise apps > Consent and permissions > User consent settings. Microsoft recommends allowing user consent only for apps from verified publishers. This only affects future consents.
- AI workspaces. In ChatGPT Business, connected apps are turned on by default. A workspace admin goes to Admin Console, selects the workspace and opens Plugins to see which apps are enabled, and turns off any that nobody has approved. In Workspace settings > General > Plugin permissions, choose Always ask or Allow read actions so ChatGPT asks before it changes anything in a connected app. For apps that offer it, Manage domains limits connections to accounts on your own email domain. Other AI workspaces have similar settings; ask your provider where they are.
- Calendars and meetings. Look at the next week of meetings for extra attendees such as “Notetaker” or an app name. These are often recording or transcription bots that a staff member connected to their calendar. Recording follows your firm’s recording procedure, including notice and any participant consent the law requires.
- Browsers. Ask staff to open their browser’s extensions page and list any AI writing, summarising or meeting tools. If your Chrome browsers are enrolled and reporting is on, the Google Admin console shows this at Devices > Chrome > Reports > Apps & extensions usage.
- Microsoft agents, if you have Microsoft 365 Copilot licences. In the Microsoft 365 admin center, the Agents page lists shared agents in the Agent Registry, with name, creator and creation date. Admins can block agents there.
Anything you find that is not in your approved tools register goes on the list as “not approved” until the data decision owner decides. Connected AI apps need approval like any other tool.
(Sources: Google Workspace Admin Help; Chrome Enterprise Help; Microsoft Learn, Entra and Microsoft 365 admin center pages; all checked 2026-11-14.)
Ask your staff
The admin check finds what is connected. It misses tools staff use in their own browser or on their phone, often on personal accounts. The only way to find those is to ask. Make it safe to answer honestly, or people will leave things out.
Send something like this:
We’re putting together a list of the AI tools people here use for work. Nobody is in trouble. We want to know what helps you, so we can keep it and set it up safely. Please answer the questions below by Friday. “None” is a fine answer. One exception: if you think something may have gone wrong, such as client information in a personal AI account, or an AI tool that sent, changed or deleted something, tell [data decision owner] now. Do not wait for this survey. Here is what to do.
Ask these questions:
- Which AI tools do you use for work, even now and then?
- For each one, which account do you use: work or personal? Free or paid?
- What tasks do you use it for?
- What information goes into it? For example, client names, emails, documents or figures.
- Do you have any AI browser add-ons, such as writing or summarising tools?
- Do you use a meeting note-taker, or AI features inside tools we already use, such as email, documents or accounting software?
- Have you connected any AI app to your work email, files or calendar?
- Can anything you use take actions, such as sending messages, moving files or booking meetings? Count it even if it waits for you to approve first.
No time for a survey? Use 10 minutes of a team meeting. Explain why you are asking, read the questions out, and note answers on a shared screen. Note tools and accounts only; do not show client names or details on the shared screen. Follow up one to one with anyone who mentions client or sensitive information.
Add the answers to the inventory. Each tool a person names becomes a row in the inventory template, or adds a user to a row the admin check already found. Put the person and account (work or personal) in Who uses it, which account. In What it can reach, record two things: what staff put in (question 4), and what the app’s connection can actually reach, checked in the admin check or the app’s settings. They are often different: an app connected to an inbox can read all of it, not just what someone pasted. If you have not checked the connection yet, write “access: to check”. Use question 8 for Reads only / Can act. In the full version, set How it was found to “staff survey”. Anything not in the approved tools register goes on as “not approved” until the data decision owner decides. Thank people for answering. If one answer leads to a tool being stopped, offer an approved alternative where you can.
What goes on the list
List three kinds of AI:
- AI tools staff use directly, such as a chat assistant.
- Connected apps that were given access to email, files or calendars. Note-takers, mailbox and file summarisers, browser extensions and any app granted access through a “Sign in with Google” or “Sign in with Microsoft” permission screen all count.
- Agents, a subset of the above that can take actions on their own, such as sending email, moving files or booking meetings.
A tool that only reads can still leak information. A note-taker that transcribes a client call, or a summariser that reads every email, holds your data even if it never sends anything. That is why the inventory has a Reads only / Can act column, not just a list of agents.
If you are unsure whether something counts, list it. It is easier to remove a row than to find out later about an app nobody knew was connected.
Inventory and register: two different jobs
- The inventory finds what is in use, approved or not.
- The approved tools register records approvals: the exact product, plan and work account, and the tasks and data types the data decision owner has approved.
Each inventory row says whether the item is in the register. If it is not, it needs a decision.
The inventory template
Copy this into a spreadsheet. Start with the short version. Add the full-version fields for anything that can act, or that reaches client or staff information.
| Field | What to record | Short version |
|---|---|---|
| Name | The name staff use | Yes |
| Type | AI tool, connected app or agent | Yes |
| Owner | The person responsible for it, plus a backup | Yes |
| Who uses it, which account | Staff or team, and whether it is a work account or personal account | Yes |
| What it can reach | Mailboxes, folders, calendars or systems, as narrowly as you can | Yes |
| Reads only / Can act | “Reads only” or “Can act”, and which actions | Yes |
| In approved tools register? | Yes, no, or waiting for a decision | Yes |
| Next review | A date, every three months by default | Yes |
| Vendor and plan | Who supplies it and which plan | Full |
| How it was found | Admin console, staff survey, calendar, browser | Full |
| What it can change | Records, files, drafts or systems it can write to, or “none” | Full |
| Approval points | Which actions wait for a person (see the human approval checklist) | Full |
| Must never do, and how that is blocked | The forbidden actions, and the setting or permission that blocks each one | Full |
| How to switch it off | Who can pause or disconnect it, and where | Full |
| Status | Found, waiting for decision, approved, paused, rejected or removed | Full |
| Date added | When it first appeared or went live | Full |
Short version, row to copy:
| Name | Type | Owner | Who uses it, which account | What it can reach | Reads only / Can act | In approved tools register? | Next review |
|---|---|---|---|---|---|---|---|
Download the short version (CSV): ai-inventory-template-short.csv
Full version, extra fields to add to the right of the short version:
| Vendor and plan | How it was found | What it can change | Approval points | Must never do, and how that is blocked | How to switch it off | Status | Date added |
|---|---|---|---|---|---|---|---|
Download the full version (CSV): ai-inventory-template-full.csv
A fictional worked example
Harbor Accounts is a fictional 40-person accounting firm. Its 20-minute check found three items. Short version:
| Name | Type | Owner | Who uses it, which account | What it can reach | Reads only / Can act | In approved tools register? | Next review |
|---|---|---|---|---|---|---|---|
| MeetNote (fictional) | Connected app | Tom Reyes | Two staff, personal accounts | Their calendars and meeting audio | Reads only | No. Disconnected pending decision | Not set |
| SumIt extension (fictional) | Connected app | Tom Reyes | One staff member, work browser | Every web page and webmail opened in that browser | Reads only | No. Removed | Not set |
| Intake Helper | Agent | Priya Shah | Operations team, work account | Screened staging folder; client folders in FileWell | Can act: moves files, creates drafts | Yes | 3 December 2026 |
The full entry for the agent:
| Field | Entry |
|---|---|
| Name | Intake Helper |
| Type | Agent |
| Owner | Priya Shah, operations manager. Backup: Tom Reyes, office lead |
| Who uses it, which account | Operations team, through the firm’s work account in FileWell |
| What it can reach | Only messages a team member has screened and moved into an approved staging folder in FileWell (it cannot read the shared intake mailbox directly); client folders in FileWell. Payroll, health details and passwords or other credentials must not be routed to the staging folder, and client restrictions still apply |
| Reads only / Can act | Can act: reads messages, moves attachments into client folders, creates reply drafts |
| In approved tools register? | Yes |
| Next review | 3 December 2026 (quarterly; last reviewed 3 September 2026) |
| Vendor and plan | Agent built in the firm’s document management software, fictional “FileWell”, business plan |
| How it was found | Already known; added when it went live |
| What it can change | Moves attachments into client folders, after approval. Creates reply drafts in the intake mailbox. Cannot edit or delete files |
| Approval points | A team member reviews and sends every draft. Every move into a client folder waits for a team member’s approval, because it changes a shared file. Filing for a client not on the active list waits for Priya |
| Must never do, and how that is blocked | Send email: its mailbox connection has no send permission, set by Northside IT (fictional IT provider). Delete files or share folders: those permissions are not granted in FileWell. Change client records: it has no access to the client database |
| How to switch it off | Priya or Tom can pause it in FileWell settings. Northside IT can remove its mailbox access |
| Status | Approved, live |
| Date added | 3 March 2026 |
Notice what the entry does. Each forbidden action is blocked by a permission the agent does not have, not by an instruction asking it to behave. The drafts it creates are listed as something it can change.
How to keep it current
- Find an owner for every row. If nobody will own an item, pause or disconnect it.
- Describe access narrowly. “Reads the intake mailbox” is better than “has email access.”
- Review every three months. Review sooner when an item gets a new connection, new data or a new action.
- Add before go-live. Add a row before an agent or connected app goes live, not after.
- Record removals. When something is removed, mark it removed and note that its access was revoked.
- Limit who sees the details. Store the inventory where the owner and IT can update it. Do not record passwords, keys or sign-in links in it.
If the check finds an app that has already handled client or sensitive information without approval, stop using it and disconnect it if you can straight away. Keep a record of what it could reach. Tell the data decision owner (or backup) which app and account, since when, and what information was involved. Do not copy the sensitive content into the report.
Standards for AI agents are still developing. In 2026 NIST launched an AI Agent Standards Initiative focused on “agents capable of autonomous actions.” (Source: NIST, checked 2026-11-14.) Expect good practice to change, and review this template as it does.
What an inventory does not do
A list does not limit what any app can reach. Permissions, approval steps and switch-off controls still need to be set and tested in the actual system. The inventory records those decisions so a person can check them.
02Evidence
Sources
- OpenAI Help Center, Admin controls, security, and compliance for plugins and appsChecked
- Google Workspace Admin Help, Control which third-party and internal apps access Google Workspace dataChecked
- Chrome Enterprise and Education Help, View app and extension usage detailsChecked
- Microsoft Learn, Review permissions granted to enterprise applicationsChecked
- Microsoft Learn, Configure how users consent to applicationsChecked
- Microsoft Learn, Manage agents in the Microsoft 365 admin centerChecked
- NIST, AI Agent Standards InitiativeChecked