如何安全地用 Microsoft Copilot in Outlook 或聊天助理草擬客戶支援回覆：設定與覆核規則

A support team can use AI to draft replies in a shared mailbox safely by keeping the AI at the drafting stage only: it proposes a reply from the customer's message and your approved information, and a person checks it and sends it. The setup that matters is not the prompt but the controls around it: correct mailbox permissions, a limit on the customer data you share, approved source material, and a written review checklist. This guide shows how to put those in place with Microsoft Copilot in Outlook or a general chat assistant.

It is for team leads and administrators of small support or sales-enquiry teams who use a Microsoft 365 shared mailbox. You will finish with a configured access model, a reusable prompt template, a review checklist, and a short pilot plan. Information on Microsoft products was checked on 28 October 2026; AI features change often, so confirm what your tenant offers.

Decide what AI is allowed to do

Write down the boundary before you turn anything on. A workable policy for a small team has three levels:

Microsoft's own privacy documentation says Copilot output is not guaranteed to be factual and that users should use their judgement when reviewing it before sending it to others. Apply the same rule to any AI tool.

Step 1: Get the shared mailbox set up correctly

A shared mailbox is a good fit for support addresses, and Microsoft lists support email addresses as a typical use. Check these facts from the Microsoft 365 admin documentation before you add AI:

• A shared mailbox can store up to 50 GB without a licence. Over that limit, you can still receive mail for a while but cannot send, and eventually senders get a non-delivery receipt.
• To access a shared mailbox, a user needs a licensed Exchange Online mailbox. The shared mailbox itself needs no separate licence.
• Only people inside your organisation can use it, and the mailbox supports a maximum of 25 users. More users can cause connection failures or duplicated messages, so use a Microsoft 365 group instead if you have a larger team.
• The associated user account is not meant for direct sign-in. Keep sign-in blocked.
• Users with permissions can send as or send on behalf of the mailbox, so replies can appear to come from a name such as Contoso Support.

If the team is larger or needs a different model, see Shared Mailbox vs Microsoft 365 Group: Which to Use and When.

Review who has access. AI tools that work inside the mailbox use whatever the user can already see, so tidy permissions first. Remove people who left the team and make sure nobody has more access than their role needs.

Step 2: Choose where drafting happens

Option A: Copilot in Outlook

Copilot in Outlook is licensed per user (a Microsoft 365 Copilot licence). A June 2026 report on Microsoft's update says Copilot in Outlook now works in shared and delegated mailboxes on Outlook on the web, new Outlook, Outlook for Mac and Outlook for iOS and Android, but not in classic Outlook. The same report says it processes mailbox content using the mailbox permissions assigned to the user, that each person's Copilot conversation history is private from colleagues, and that it appeared when the shared mailbox was added to the folder pane rather than opened through "Open another mailbox." The source is a third-party blog rather than a Microsoft page, and the folder-pane point is the author's own observation, so test all of this in your own tenant before you rely on it. Microsoft's drafting support page, cited below, does not mention shared or delegated mailboxes.

Microsoft's support page describes the drafting steps. In new Outlook and Outlook on the web, select Home > New mail, select the Copilot icon and Draft, type a prompt, select Generate, then Keep it when you are satisfied. You can change tone and length from a dropdown, or retry the prompt. Selecting text in a message and then the Copilot button lets you modify that selection. On the web and new Outlook, drafting does not support plain text messages; switch to HTML under Settings > Mail > Compose.

Microsoft states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation LLMs, that Copilot only surfaces organisational data the user has at least view permission to, and that it honours Microsoft Purview sensitivity label usage rights. For the readiness steps, see Microsoft 365 Copilot Readiness: Fix Oversharing Before Rollout.

Option B: A chat assistant such as ChatGPT or Claude

If you do not use Copilot, a team can paste a customer message and approved facts into a chat assistant. That means customer data leaves Microsoft 365, so the rules matter more:

• Use an account type covered by a business agreement and read the provider's data usage and retention terms. Do not use personal consumer accounts for customer data.
• Paste only the minimum needed. Remove names, order numbers, addresses and anything sensitive unless the reply truly needs them.
• Keep a shared prompt template so each person asks the same way.

Step 3: Prepare approved source material

AI drafts are only as safe as the facts you give them. Create a short document of approved answers for the common questions: opening hours, how returns work, how to reset an account, what you can and cannot promise. Keep it in one place, such as a SharePoint page, and paste only the relevant part into the prompt, or reference it from a tool that can read it.

State what must never be promised. Examples: refunds, compensation, delivery dates, legal positions, medical or financial advice, and anything that needs a manager's approval. Put the escalation route in the document so the AI can recommend it in the draft.

Step 4: Use a prompt template

Use a fixed template so the model returns a consistent, reviewable draft. Adapt this one:

You are drafting a reply for the Contoso Support team. A person will review it before sending.

Rules:
- Use only the facts in the APPROVED INFORMATION section. If the answer is not there, say what is missing and suggest escalating to a colleague.
- Do not promise refunds, compensation, delivery dates or policy exceptions.
- Do not invent order numbers, names, links or prices.
- Match the customer's language. Keep the reply under 150 words. Friendly, plain tone.
- End with a one-line list "Check before sending:" naming any facts the reviewer must verify.

CUSTOMER MESSAGE:
<paste the relevant message only>

APPROVED INFORMATION:
<paste the relevant approved answers>

The "Check before sending" line makes the model flag what it relied on, which speeds up review. It is a prompt to the reviewer, not a guarantee, so the checklist below still applies.

Step 5: Review every draft before sending

A short checklist prevents the common failures. Print it or pin it in the team channel.

Also read the draft against the original message. Check that it answers what was asked rather than a similar question.

Handle risk cases separately

• Security or fraud reports. Do not let AI handle these; route them to the person responsible.
• Complaints with legal or regulatory angles. Draft by hand or have a manager approve.
• Prompt injection. A customer's email is untrusted input and can contain text that tells the AI to ignore your rules. Never let a draft trigger actions, such as sending, refunds or data lookups, without a person approving. Reading the draft carefully is your protection.
• Sensitive data. If a message contains payment card numbers or passwords, do not paste it into an AI tool; handle it under your normal procedure.

Run a two-week pilot

1. Pick two or three people and one category of request, such as order status or password help.
2. Record for each draft whether it was sent unchanged, edited, or rejected, and why.
3. Review the log at the end of week one and update the prompt and approved information.
4. At the end of week two, decide whether to expand, based on your own log rather than assumptions. Track edits needed per draft and any incorrect facts caught in review.

Keep the review rule even if drafts become reliable; the cost of one wrong promise to a customer is higher than the time saved by skipping a read.

Checklist

• Shared mailbox permissions reviewed, with only current team members having access.
• Policy written: AI drafts, people send.
• Tool chosen under a business agreement, with data terms read.
• Approved information document and a "never promise" list created.
• Prompt template shared; review checklist pinned.
• Pilot log started.

References

• Microsoft Learn, About shared mailboxes in Microsoft 365: https://learn.microsoft.com/en-us/microsoft-365/admin/email/about-shared-mailboxes
• Microsoft Learn, Data, privacy and security for Microsoft Copilot: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-privacy
• Microsoft Support, Draft an email message with Copilot in Outlook: https://support.microsoft.com/outlook/copilot-pages/draft-an-email-message-with-copilot-in-outlook
• Topedia, Copilot in Outlook is now available in shared and delegated mailboxes (23 June 2026): https://blog.topedia.com/2026/06/copilot-in-outlook-is-now-available-in-shared-and-delegated-mailboxes/

This guide was researched and drafted with AI assistance and checked against the vendor documentation listed in its References. Read the maintained version and related guides on EpiFive.