AI email triage: how to sort a shared inbox with AI

Give the AI a narrow first job: read each new email, label it with a short list of categories, mark urgency and suggest an owner. People still decide and reply. Test it on past mail, then run it as suggestions only on new mail. Compare its labels with your team’s for two weeks before it drafts routine replies.

This guide covers AI email triage for a team inbox such as support@ or accounts@. It also says when a filter or built-in assistant is enough, so you do not add an agent you do not need.

What does AI triage do in a shared inbox?

Triage means deciding, for each new email, what it is, how urgent it is and who handles it. An AI model can do that reading and sorting, so people spend their time on the replies.

In a shared inbox, a good setup does these jobs on each message:

• Reads the email and picks one category from your list.
• Marks urgency using rules you wrote.
• Pulls out fields such as an order number or an invoice number.
• Suggests an owner and writes a one line summary.
• Drafts a reply for routine requests.

It should not decide refunds, change a contract or answer a complaint on its own. Those stay with a person. This guide treats AI email management in that narrow sense: sorting, routing and drafting, with people making the decisions.

What should you set up before you turn AI on?

Write down your categories, an owner for each and your urgency rules before any AI touches the inbox. An AI model sorts against the list you give it, so a vague list gives vague labels.

• Keep 6 to 10 categories. If two of them overlap, merge them.
• Name one owner and one backup for each category.
• Write the urgency rules down, for example “a payment failure is urgent” or “a request with a stated deadline this week is urgent”.
• Decide which emails need no action, such as automatic notices.

Then decide where the work happens. In Google Workspace, a group owner can turn a Google Group into a Collaborative Inbox. As of October 2026, Google’s help page says you can assign responsibility for a conversation to yourself or another group member. It also says you can mark a conversation as complete, needing no further action, or a duplicate. Microsoft 365 has shared mailboxes, which Microsoft describes as playing a key role in team collaboration.

How do you set up AI triage step by step?

Start with one inbox and test on past mail. Let the AI act only after it clears a bar you set in advance.

1. Pick one inbox with steady volume and clear categories.
2. Export a sample of past mail, a few hundred messages if you have them. Remove or mask personal and financial details you are not allowed to share with the AI tool.
3. Write the category list, owners and urgency rules into the instructions, in plain sentences.
4. Set a bar before going live, for example no missed urgent emails and agreement on most categories.
5. Run the AI on the past sample. Compare its labels with how your team filed those emails.
6. Run it as suggestions only on new mail. The AI labels and suggests an owner, and people carry on as usual.
7. After two weeks, or once you have a few hundred emails, compare its labels and owners with what people did. Read every urgent email it missed.
8. Fix the instructions where it disagreed with your team. Turn on labelling and assignment only when it clears your bar.
9. Turn on drafts for routine request types only. A person reviews each draft until the team trusts a type.

Can AI answer customer emails?

AI can draft replies to routine questions, and a person should approve anything about money, contracts, complaints or legal matters. Drafts work best when they draw on your own documents, such as price lists and policies.

In Australia, the Office of the Australian Information Commissioner (OAIC) publishes guidance on commercially available AI products, updated 17 January 2025. It says privacy obligations apply to any personal information put into an AI system. They also apply to the output where it contains personal information. Customer emails hold names and details, so check with whoever handles privacy in your organisation how this applies to you.

The same guidance says public-facing AI tools, such as chatbots, should be clearly identified as such to external users. If customers talk to an AI directly, tell them. It also says businesses should update their privacy policies and notices to explain how they use AI.

What to automate and what to keep with people

Automate the sorting and the routine drafts, and keep judgement calls with a person. Use this table as a starting point and adjust it to your own risk.

Is a rule, a built-in assistant or an agent the right tool?

A rule is enough when the sorting follows a fixed pattern. A built-in assistant can cover reading and drafting inside one mailbox. Choose an agent when the work spans several systems.

Rules and filters

If the sorting depends on a sender, a subject line or a phrase, a rule is enough. As of October 2026, Gmail filters can send email to a label, or archive, delete, star or automatically forward it. Outlook has rules for managing messages as well. Rules fail when the wording varies, because a rule cannot read for meaning.

Built-in assistants

As of October 2026, Microsoft says you can use the Microsoft Copilot app to read emails, draft responses and summarise emails in a shared mailbox. You name the mailbox in a chat prompt. You need a Microsoft Copilot licence. Microsoft says Copilot currently cannot help if you only have shared folder permissions in a delegate mailbox. If your team works in one mailbox and only needs summaries and drafts, this may be enough.

An agent

Choose an agent when the work leaves the mailbox. Examples are looking up an order in another system or updating a customer record. An agent plans those steps and uses tools to carry them out. It takes more care to set up and review, so use one only when a rule or an assistant falls short. Our post on governing agents in production covers the controls.

Where does TobyBot fit?

TobyBot is an agentic bot by Xinference. It has an inbox manager template, and an agent can start when a new Gmail message arrives. Native connectors cover Gmail and Outlook, and other systems, including internal ones, connect through MCP servers. Our guide to building a first agent from a template shows the steps.

You set a goal. TobyBot plans the steps, shown as one card for each, then runs them with your tools. Every tool call and result is visible while it runs, and you can pause it. Every tool call and result is visible while a task runs, with the token count, and you can pause it. For draft replies, upload your documents to the knowledge base. Agents search it with keyword and semantic search together.

The review steps above are a process your team sets up and owns. You can save a goal as a bot and schedule it. Our weekly report agent post uses the same method on a different job. To test it on your own inbox, book a demo.

Try TobyBot. See all use cases, or book a demo on your own workflow.