Inbox Triage Agent
An agent reads incoming email in Gmail or Outlook, drafts replies for routine cases, and flags anything ambiguous for a human to review and send.
The problem
Shared inboxes fill with a mix of routine requests, genuine emergencies, and noise. Someone has to read everything to find the messages that matter, and reply latency grows with volume. Routine replies eat the time that should go to the hard cases.
How it runs on Neblex
The agent watches a Gmail or Outlook mailbox through connectors, with poll-based change detection picking up new mail. It reads each message, decides whether it is routine, and for routine cases writes a draft reply in the mailbox: a draft, not a sent message.
Sending can require human approval: the reviewer sees the original email, the agent reasoning, and the proposed reply in a task inbox with SLA timers and escalation, and approves or rejects. Ambiguous or sensitive messages skip drafting entirely and get flagged to a human, with a Slack notification so they are seen quickly.
The agent runs as a step in a deterministic flow, so the labeling, folder moves, and notifications around it are exact and repeatable, with retries and replay on failure. Every read, draft, and send is logged to a tamper-evident hash-chained audit trail.
Step by step
Connect the mailbox
Gmail or Outlook connects through a connector, with poll-based change detection watching for new mail.
Agent reads and sorts
Each message is judged routine, ambiguous, or urgent.
Routine replies drafted
The agent writes a draft reply directly in the mailbox.
Human reviews sends
Drafts route through approval, with the reasoning and the proposed reply shown to the reviewer before anything sends.
Ambiguity escalates
Unclear messages flag to a human with a Slack notification and full context attached.
Platform capabilities used
- Gmail and Outlook connectors as tools
- Poll-based change detection
- Human-in-the-loop approval before sending
- Agent inside a deterministic flow
- Slack notifications
- Hash-chained audit trail
Common questions
Can the agent send email without anyone checking?
Only if you remove the approval checkpoint. Any send can require a human decision first, with the reasoning and the proposed reply shown to the reviewer, and every action lands in the tamper-evident hash-chained audit trail.
How does it decide what is routine?
You define the categories and guardrails when you build the agent, which you can do by describing it in plain English. Anything the agent cannot place confidently is flagged to a human rather than answered.
Which model reads our email?
The one you choose. Neblex connects to any OpenAI-compatible endpoint, including self-hosted models on your own infrastructure, or Anthropic, so mail content can be processed by models you control.
Related use cases
Support Triage Agent
A triage agent reads each Zendesk ticket, judges urgency and sentiment, drafts a reply, and escalates anything it is unsure about to a human for review.
AgenticKnowledge Agent
An agent answers employee and customer questions from your docs, wikis, and past tickets, citing sources and flagging gaps in the documentation.
FinanceExpense Categorization and Approval
Categorize submitted expenses, check them against policy, and route each report to the right approver based on amount and type, all in one governed flow.
Want this running on your stack?
Neblex Integration Fabric is in beta: full platform, free while in beta. Bring this workflow and we will map it to your systems.