A three-agent AI support desk: triage, documentation research, reply draft and a human gate

A three-agent AI support desk: triage, documentation research, reply draft and a human gate#

A single agent that receives a ticket and replies straight away does three different jobs at once: it judges how serious the issue is, finds the answer and writes. Each of these fails differently. This n8n workflow splits them across three agents with separate jobs and one gate at the end.

Triage: severity by definition, not by tone#

The first agent replies to nothing. It classifies the ticket against the severity definitions you set (from critical to low, plus needs_info). Two rules make the biggest difference:

An angry tone does NOT raise severity. A calm report of total data loss is still critical.

The second: if the ticket is too vague, the agent sets needs_info and lists what is missing rather than guessing.

Research: only pages it actually fetched#

The second agent has an HTTP tool that fetches a documentation page and returns its text. It is told to try several phrasings and follow links, up to 5 pages. It may cite only pages it fetched. Anything it inferred goes into a separate inferred_notes field.

If the docs do not answer the question, the agent sets found_answer to false and describes the gap. That is not a failure. Every such gap goes out as a separate notification, a ready-made documentation to-do.

Draft: wording only, no new facts#

The third agent writes the reply in the customer's language and to your tone guide. It may not state a fact the research did not find, and it may not promise a timeline, refund or feature. An angry customer gets a shorter, calmer reply, not more apologies.

The gate#

A reply goes out automatically only when three conditions hold at once:

Everything else goes to a person with three links: send as-is, I'll rewrite it, escalate to engineering. Every ticket, whatever the path, is written to the sheet with severity, decision and the draft.

Limitations#

The workflow does not send email. The automatic path marks the reply as ready and returns it through the webhook. Sending has to be wired on the helpdesk side or with a mail node.

The webhook responds only after the last step. On the human path that means the system that sent the ticket waits for the click or times out. In production, respond immediately and process the rest in the background.

The research tool fetches any URL the agent comes up with. Restrict it to the documentation domain.

Decision links have no login, and chat apps may open them to build previews. Turn off link previews in the channel that receives these messages.

When it does not make sense#

With a few tickets a week and small documentation, replying yourself is faster. The workflow makes sense with dozens of tickets a day, a share of repeat questions, and documentation that exists and is public. We describe how to split work between agents in every agent needs to be told what it does not do. When ticket handling should live inside your helpdesk or panel, that is an AI process automation project.