Guide · Review and trust

AI ticket triage: what to automate and what to keep human

AI ticket triage works best when the model drafts, classifies and routes while a person decides what is real work. Relay automates the capture and the first pass, then holds every draft for approval. The tracker stays a record of decisions rather than a stream of unverified model output.

By the Relay maintainers · Last updated 19 November 2026 · Source on GitHub

What is AI ticket triage?

Triage is the sorting step between raw input and a tracked issue: deciding what the work is, how urgent it looks, who should own it and where it belongs. AI-assisted triage proposes those answers from notes, chat or transcripts. It is useful because the raw material is unstructured and repetitive to sort by hand. It is risky because a plausible-looking draft can be wrong in ways that are expensive to unwind after it reaches a shared tracker.

What should be automated?

Automate the mechanical parts. Relay captures from multiple sources, extracts candidate tickets, quotes the supporting line, suggests a type and priority, ranks up to three owners with reasons, and resolves the destination from a routing rule. Low-risk drafts can be approved in a batch under policy. These are tasks where a model is fast and a reviewer can verify the result in seconds against the attached evidence.

What should stay human?

Keep the decision. A person should confirm that a draft describes real work, that the quote supports it, and that the owner and due date make sense. Relay supports edit, reject, delegate, merge and split, with reasons recorded for learning. Approval is the gate between a suggestion and an issue. Removing it makes the tracker faster to fill and harder to trust, because nobody remains accountable for what was created.

How does triage work in Relay?

  1. Capture. Notes, chat messages, email or a transcript enters through a channel, the API or an agent.
  2. Draft. Relay extracts tickets with type, priority, owner and a source quote.
  3. Queue. Drafts land in a review queue with their evidence and assignment explanation.
  4. Decide. A reviewer approves, edits or rejects before any push happens.
  5. Push. Approved items pass through an idempotent outbox to one tracker each.

What can go wrong?

Extraction on messy multilingual fixtures measures about 0.78 precision and 0.75 recall, so some drafts need editing and some real work is missed. Live model output varies across runs. A suggested owner can be available but wrong for the task, which is why the reason is shown beside the score. The mitigation is not a larger model; it is a short review loop with the evidence in view.

Is triage useful without a hosted service?

Yes. Relay is Apache-2.0 and self-hosted first, with no feature gates or artificial limits on the open-source path. You run the API and worker on your own infrastructure, connect your trackers and chat apps, and keep your data in your own database. Hosted cloud plans are planned, not live; signup and billing are opt-in per instance. Self-hosting is the working path today.

Next step

Run the credential-free demo on a real meeting note and judge the drafts against their quotes. That tells you more about triage quality than any feature list.


Self-host Relay Read the workflow guide