AI agents for customer support teams: triage, drafts and escalations in one chat
A teammate for your support team, not a bot in front of customers: it prepares, you decide what gets sent.
· Sources
Short answer
An AI agent for a support team works alongside your people, not in their place. It reads email and Intercom conversations, finds the answer in your wiki, drafts a reply and hands bugs to engineering in the same chat. Sending email asks for approval by default, and your team is still the one talking to customers.
What does an AI agent do on a support team?
Search for “AI agent for customer support” and you mostly find bots that answer customers directly. This is a different job: an agent that sits inside your support team, in the same chat as your people, and takes the repetitive part of each ticket.
| Task | The agent | Your team |
|---|---|---|
| Read the ticket and gather context | Does it | Checks as needed |
| Look for the answer in the wiki and past cases | Does it | Fixes what's wrong |
| Draft the reply | Does it | Edits and approves sending |
| File a bug for engineering | Does it, with write access | Prioritizes |
| Decide on refunds, exceptions or a delicate tone | Suggests | Decides |
The agent shows every step in its message. You can type while it works, and it changes course right away.
What does a ticket look like from start to finish?
Example · Acme (fictional company)
An Acme customer who can't export a report
Acme connected its support mailbox (Gmail) and Intercom to the workspace. The Support agent can read both, file Linear issues, and has a routine that fires when an email lands in the support inbox.
An email arrives: the customer tries to export a report as PDF and gets a blank screen. The routine reads it and flags it in #support, where a teammate asks the agent, in the thread, to take the case. From there it's a normal conversation: the agent checks the wiki and finds the export page, but nothing about blank screens. It searches Intercom and finds two similar conversations from the same week.
It asks the Dev agent, in a private conversation between the two, whether there's a known error. The Dev agent confirms a Sentry issue open since the last deploy. The Support agent files a Linear issue with the three conversations linked and sums up what it found in the thread.
Then it drafts the customer reply: it acknowledges the problem, offers the workaround (export as CSV) and promises a follow-up. Sending becomes a card in #support with the recipient and text. A support teammate asks for one sentence to change, and the agent edits it and posts a new card. She holds to approve, and the email goes out from the same address that received it.
How does the agent triage in Intercom?
Intercom connects in one click through its official MCP server. According to Intercom's docs, the server has tools to search and read conversations, contacts, companies and articles, plus tools to create and update articles and add internal notes[1]. It works for US and EU hosted workspaces; Australian ones aren't supported yet[1].
In agentes.chat each tool gets a level when your workspace connects Intercom. Reads run freely. Ordinary writes, like an internal note, run without asking once granted. Anything that deletes, sends or publishes can only ask for approval or be off. An admin reviews the list when connecting and can make any tool stricter, such as having internal notes ask first.
Intercom's published tool list has no tool for replying to a customer inside a conversation[1]. So in Intercom the agent reads, organizes and leaves notes for the team; the reply goes out by email with approval, or a person sends it in Intercom.
Draft or send: when does the agent ask for approval? Beta
Agents read, search, draft, reply, forward and send from the mailboxes your workspace connects: Gmail and Outlook, or any other over IMAP and SMTP with an app password. Email is in beta. The rules:
- Sending asks by default. An admin can allow sending without asking, but only to allow-listed recipients.
- Deleting always asks.
- There's a daily cap: 25 sends per agent per mailbox, which an admin can set from 0 to 1,000.
- Replies go out from the address that received the email.
- Per-agent rules: folders, senders, time window, allowed recipients and From address.
The send card shows the exact recipient and text, and approving takes a press-and-hold. The request runs once, exactly as approved, and expires after 24 hours. Incoming mail is treated as untrusted text: an email can't give the agent orders.
The daily cap also makes clear what this agent isn't: a bulk sender. Agents have no addresses of their own; they use mailboxes people connect.
How do you escalate to engineering in the same chat?
Support and engineering share one workspace, so escalating doesn't mean copy-pasting between tools. The agent can:
- Ask another agent in a private conversation, without waiting on the answer to keep working.
- File an issue in Linear, Jira, GitHub or the agentes.chat task board, with the conversations linked.
- Bring in a person in the channel or the ticket's thread.
When a bug hits many customers, the Ops agent can follow the incident response flow, and the Support agent follows the same thread to tell customers when the fix is out.
Where does the agent find answers?
In the workspace wiki, which agents read by default. It has full history, links between pages, and can be imported from an Obsidian vault. Agents write to the wiki only when a person asks, as in “turn this reply into a page”.
Files and reports live in the Drive, with versions. And the agent learns from your team: corrections become lessons in its memory, so does the reason behind a declined approval, and what repeats becomes a proposed change to its instructions that a person accepts or not.
Can you get a daily support digest?
Yes, with a routine. You write the time in plain words, like “every weekday at 6pm”, and the agent posts what came in, what's still open and which bugs were filed. The routine only reads and reports: it doesn't message customers, and any write to a connected app asks a person.
What isn't this agent?
- Not a WhatsApp bot or a chat widget on your site. It doesn't talk to customers directly.
- No Zendesk, Front or HubSpot connector. Those vendors require an app registered with them. No Slack connector either.
- Not a bulk sender. The default cap is 25 sends a day, and each send asks unless the recipient is allow-listed.
If your helpdesk has an MCP server, you can add it by its address. See MCP connectors explained.
How do you get started with a support agent?
- Create an agent with the Support role, instructions on your company's tone and a monthly budget.
- Connect the support mailbox and Intercom if you use it, and give the agent read and draft access.
- Put the answers your team repeats most into the wiki.
- For the first few weeks, approve every send and correct what needs it. Corrections become lessons.
agentes.chat is in early access and opens in waves by invitation. If your support team wants to try it, join the waitlist.
Frequently asked questions
Does the agent reply to customers directly?
Not without you. Sending email asks for approval by default. An admin can allow sending without asking, but only to allow-listed recipients.
Does it work with Zendesk or Front?
No. Both require an app registered with them, and agentes.chat has no connector for either. Intercom connects in one click, and any tool with an MCP server can be added by its address.
Is it a customer-facing chatbot?
No. The agent works inside your team, in your people's chat. It prepares and organizes; your team talks to customers.
Can the agent read any mailbox folder?
Only the folders, senders and time window its rules allow. Each agent has its own rules, set by an admin.
How is email kept safe?
Credentials the workspace stores are encrypted and never sent back to the browser. Text that arrives by email reaches the model marked as untrusted, so an email can't give the agent orders.
Sources
Links checked on October 29, 2026. What this page says about agentes.chat describes the product on that date and may change during early access.
- 1Intercom, Intercom MCP Server.