How to set up a Microsoft 365 help desk bot in Grok Bot, safely

Low-poly faceted mountain valley at dusk with a small glowing lighthouse on a distant ridge

Most small businesses I work with have the same help desk problem. The same dozen questions arrive every week: Outlook will not sync, Teams will not sign in, a shared mailbox has gone missing, where do I find this setting. A person has to read each one, find the answer, and reply. A Microsoft 365 help desk bot will not replace that person. It can take the first pass, so the person spends time on the problems that need judgement.

I run a Help desk bot in Grok Bot that can start remote support through AnyDesk from the shared Linux computer. It sits alongside a PowerShell bot post that runs my scripts, a Teams bot, and read-only security monitor bots for individual tenants. This guide is how I would set one up from scratch, with the safety rules first. I have checked the Grok Bot steps against Cursor’s published Grok Bot documentation and the Microsoft claims against Microsoft Learn. Where I could not confirm a screen or a setting, I describe it generally rather than guess.

1. Decide what the bot is for

Write the scope down before you create anything. A help desk bot that tries to do everything ends up with access to everything.

A sensible first scope is this:

  • Answer “how do I” questions from official Microsoft documentation, with a link.
  • Triage a request: ask the questions you would ask, and summarise the likely cause.
  • Run read-only checks, such as service health or a user’s licence and sign-in status, and report what it found.
  • Draft replies and ticket notes for a person to send.
  • Hand over to a human, with the details collected, when the fix needs a change.

And what it does not do at the start: reset passwords, change licences, edit mail flow, delete anything, or touch a production setting. You can move a task across that line later, one task at a time, once you trust the result.

2. Create the bot and give it a job

In the Grok Bot app, choose New Bot (the + button at the top of the sidebar) or press Cmd/Ctrl+N, then type a name and choose Create. Cursor’s documentation says to use your first message to give the Bot its job, its sources and its approval boundaries, and to ask it to remember them. I follow that. A first message along these lines works, using a fictional business:

You are the Help desk bot for Contoso Pty Ltd, a small business using Microsoft 365. Your job is first-line triage for Outlook, Teams, OneDrive, SharePoint and sign-in problems. Use Microsoft Learn first, and link the page you used. Stay read-only. Draft replies for me and never send anything to a user without my approval. Never ask a user for a password or a one-time code. If a fix needs a change to the tenant, stop, summarise the request, and ask me. Remember these rules.

Cursor’s documentation says a personal Bot keeps lasting rules in its memory, and that a Team Bot still has a description its owner or a Manager edits under About in the Bot’s details. The app’s own settings map also lists description among the per-agent settings, reached from the Bot’s name in the chat header. So I tell the bot in chat, ask it to remember, and check where the description lives in my own version of the app.

Personal bot or Team Bot?

If more than one person will use the help desk bot, a Team Bot is the better fit. Cursor says Team Bots need a Team or Enterprise plan and at least one other person on your team. You create one from the + menu, tell it its job, and follow its setup, which asks for plugins, secrets, skills and files. Until you publish it, it is a Draft that only you can see. Teammates then find it under Browse Team Bots, or you share it with the people you choose. The owner and Managers change the Bot; each person’s chat with it stays private.

There is one catch. A Team Bot uses the apps and secrets its owner added, and Cursor warns to add only what everyone you share it with may use. A shared connector lets anyone who talks to the bot act through that one account, changes included, and Cursor says to share only an account that holds nothing sensitive. That sentence matters more than any other in this guide.

3. Connect the right connectors, with the least access

Open Connect apps from the sidebar, find the app, add it, and complete the sign-in when asked. Check it appears under Installed. Cursor says a connected app can do only what the account you authorise can already do in that service, and the connection is account-wide, so every Bot you run can use it. For a help desk, the useful set is small. I use Microsoft Learn for documentation search (I wrote up connecting Microsoft Learn), plus Teams, Outlook, OneDrive and SharePoint where the help desk genuinely needs them, and a Microsoft Graph connector for read-only tenant checks.

A Graph admin connector is not in every catalog. Grok Bot lets you add a custom connector (a remote HTTPS server URL, optional headers and an OAuth client ID, or a command), and your team admin’s connector policy can block it. If your Microsoft 365 account sees a message that your organisation’s admin needs to approve the app, that is the Microsoft 365 admin approving Grok Bot, and retrying the sign-in does not help. Send them the approval link the page shows.

Entra app registration: delegated first, read-only first

If your connector uses an app registration in Microsoft Entra, create it deliberately. Microsoft Learn’s steps are to sign in to the Microsoft Entra admin center, go to App registrations, choose New registration, give it a name, choose the supported account types, register it, and record the Application (client) ID. After that the decisions that matter are the permissions.

Microsoft’s guidance on Graph permissions is to apply least privilege and request the fewest permissions the app needs. It describes two access scenarios. With delegated access the app calls Graph on behalf of a signed-in user, so it is limited to that user’s scope. With application access the app calls Graph with its own identity and no signed-in user, and Microsoft says application permissions carry the most privacy risk. So I start with delegated, read-only permissions. I move to application permissions only when unattended work needs it, and then only the read permissions that the check needs.

Never give the bot Global Administrator

Microsoft recommends fewer than five Global Administrators, with all of them protected by multifactor authentication, and two cloud-only emergency access accounts. A help desk bot is not one of the five. Be careful with Helpdesk Administrator too. Microsoft marks it a privileged role: it can change passwords and invalidate refresh tokens, and Microsoft warns that changing a user’s password can mean assuming that user’s identity, including owners of app registrations. That is exactly the sort of action I want a human to approve, so I do not give the bot that role.

If administrators have to sign in anywhere in this chain, a Conditional Access policy that requires multifactor authentication for administrators is the baseline. Microsoft’s steps say to set the policy to Report-only first, review the results, then turn it on. Note Microsoft’s warning that Conditional Access policies are not enforced for administrative-unit-scoped or custom roles, so check which roles your accounts use.

4. Give it knowledge it can cite

Microsoft Learn is the main source. I tell the bot to answer from Learn first and link the page, because a link lets the technician check the answer in seconds. For your own procedures, put the documents where the bot can read them, and for scripts point it at a repository. My scripts live at https://github.com/directorcia/Office365, and the PowerShell bot post explains how I keep script work in one bot, read-only unless I say otherwise.

Do not treat the bot’s memory as the source of truth. Cursor says to keep changing facts in the source system and ask the Bot to cite or reopen current data for consequential decisions. For a help desk, that means licence counts, policy settings and service status are looked up each time, not remembered.

5. Turn repeat fixes into skills

Cursor’s guidance is to start with a one-time task, make it reliable, save the method as a skill, and only then automate it. A good skill states when to use it, the inputs and access it needs, the sequence of work, how to validate the result, what to return, and what needs approval. Ask for it in chat. For example:

Save what we just did as a skill called “Outlook not syncing triage.” Include the questions to ask the user, the read-only checks to run, the Microsoft Learn pages to cite, the format of the summary, and the rule that any change to a mailbox needs my approval.

The ones I would build first are an Outlook sync triage, a Teams sign-in triage, and a “password reset request” skill that does not reset anything. It collects who is asking, confirms the request through a second channel and hands it to a person. Private skills are one library shared across your Bots, and on a team you can publish a skill so teammates can use it.

6. Remote support needs a human in the loop

Because your Bots share a cloud computer with a browser, filesystem and terminal, a bot can start a remote support tool there. Mine can start AnyDesk from the shared Linux computer. I would treat that as a step that needs a person’s approval every time, not something to automate. Cursor’s documentation says passwords and one-time codes should never go in ordinary chat. A bot should use a login saved in Vault, or a Secure Form, and hand you the computer for steps it cannot do itself, such as single sign-on or an approval on another device.

7. Routines for daily checks

A routine runs a saved instruction on a schedule or when something happens. Ask the bot that should own the job, in plain words, and confirm the owning bot, the schedule and time zone, the input source, the expected result, the approval boundary and what to do when the source is missing. Schedules use the timezone in Settings, under Bot. A new routine waits for its next time, and you can use Test to try it, remembering that a test does real work.

For a help desk I keep it modest:

  • A weekday morning check at a sensible hour that reads service health and open requests, and posts a short summary to me. It does not contact anyone.
  • A routine that starts when mail arrives at the bot’s own address, limited to the senders you name, that reads the request and drafts a triage note.

Each run spends usage, so a broad trigger on a busy mailbox is expensive. Cursor says to define a narrow matching rule and avoid listeners such as every new message. Email from a routine or an incoming message also waits as a draft until you press Send, which is the behaviour I want.

8. How people reach it

Microsoft Teams

This takes two administrators. Cursor’s Microsoft Teams documentation says a Cursor team admin who is also a Microsoft Teams Administrator or Global Administrator turns Teams on, once, for the organisation. The organisation has to allow custom apps in Teams. The bot’s owner or a Manager then connects it, and a Teams admin reviews and publishes it. Setup can take minutes and up to about an hour, and the app shows when the bot is ready. People should wait for ready before adding it to a 1:1 chat. In a channel the bot only receives messages that really @mention it, picked from Teams’ mention list, and it answers in the thread. Everyone who chats with it needs a Teams licence in that organisation, and guest accounts cannot add it.

The approval grants delegated Microsoft Graph permissions only. Cursor says the bot holds no application permissions in the organisation, and asks for no organisation-wide access to mail, calendars, files or Teams messages.

Email

A bot can also take requests by email. You can claim one address at mail.grokbot.com, such as contoso-desk@mail.grokbot.com (a made-up example). I covered the limits in claiming a Grok Bot email address. It is one address per account, you cannot rename or delete it, and all your Bots can use it in your own chats and routines. Plus aliases such as contoso-desk+vip@mail.grokbot.com land in the same inbox. Anyone can write to an alias, so keep a sender filter on any routine that should act only for certain senders.

9. Approvals and safety

Cursor’s Auto Review checks each action your Bot is about to take, and when one needs your OK you see an approval card showing the operation and its inputs. You can Allow once, Always allow, or Deny. Be careful with Always allow on anything that changes a tenant. Custom Rules let you add narrow Ask first rules, such as “ask first before sending any external email”. If both an Allow and an Ask first rule match, Ask first wins. Auto Review is on by default, and the Auto-review switch is in Settings, General, under Bot.

Two points are easy to miss. Auto Review treats text from an email or a web page as not being your request, which is a defence against someone emailing your help desk bot instructions. That reduces the risk but Cursor says it should complement least privilege, not replace it. And cards raised by a routine or other background work expire after 30 minutes, so an unattended routine cannot sit there with a standing approval.

My short list of rules for a help desk bot:

  • Draft first. A person sends anything that goes to a user.
  • Read-only by default. Changes need approval in the bot’s own chat, never relayed from another bot.
  • No passwords or one-time codes in chat, ever.
  • No Global Administrator, and no standing Helpdesk Administrator.
  • Pause or delete routines and revoke connectors when a project ends.

10. If you are an MSP with many tenants

This is where I am most cautious. My tenant monitor bots are read-only security monitors, one per tenant, built from the Microsoft 365 security monitor template. It is tempting to treat one bot per customer as a security boundary. Cursor says it is not: all your Bots share one cloud computer, so files, browser sessions and command-line credentials on it are available across your whole roster, and Cursor says not to use separate Bots as a security boundary. Connected apps are also account-wide. I covered the wider picture in Grok Bot security and customer Microsoft 365 data.

So the separation has to come from the Microsoft side. Use a separate, least-privilege identity per tenant, so a bot that is compromised or confused can only see the one customer it was given. For partner access, Microsoft says granular delegated admin privileges (GDAP) give partners least-privileged, time-bound access that customers explicitly grant, in place of the Global Administrator role. Use that model for any human or tool that works across customers. I would not give one shared account across tenants to a Team Bot that your whole team can talk to.

What to be aware of

  • I have not verified every Grok Bot screen. Where I name a menu, it comes from Cursor’s documentation or the UI map; anything else I have described generally.
  • A bot’s answer is only as good as its source. Make it cite the Learn page, and check anything that changes a tenant yourself.
  • Routines spend usage. Review them regularly.
  • Microsoft changes roles and consent screens often. Check the linked Learn pages before you rely on a permission name.
  • Voice and phone calling are separate topics. See my voice setup post if the help desk will take calls.

Start small

Create the bot, give it Microsoft Learn and nothing else, and let it draft replies for a week. Add one read-only check. Save the first skill. Only then add a routine, and only then think about Teams. If a help desk bot never needs more access than that, it is still saving you time, and nobody has handed an AI a master key.

Sources

One thought on “How to set up a Microsoft 365 help desk bot in Grok Bot, safely”

Leave a Reply