Guide · Start here
From meeting notes to tracked tickets
AI auto-ticketing turns a meeting note, chat thread or transcript into a ticket you can inspect before it reaches a tracker. Relay drafts each item with the exact source quote, suggests an owner with reasons, and waits for a person to approve. Nothing is pushed without that approval.
What is AI auto-ticketing?
AI auto-ticketing is the step between a conversation and a tracker. A tool reads the source, drafts candidate tickets, attaches the sentence or transcript segment behind each one, and holds the drafts for review. Relay is an open-source, Apache-2.0 implementation of that pattern. It runs on your own infrastructure and keeps the evidence chain from source quote through to approved ticket, so a ticket's origin is never lost.
What can Relay capture?
Relay accepts plain text, Markdown, VTT and SRT transcripts, DOCX, PDF and transcript JSON. Email in .eml form and notetaker exports go through the capture API; audio needs a transcription endpoint or a supplied transcript. Every draft links back to the quote or segment that produced it. The tables below show the inputs and the four trackers available after approval.
| Source | Accepted form | Evidence kept |
|---|---|---|
| Notes | Text, Markdown, DOCX, PDF | Quoted sentence from the note |
| Transcripts | VTT, SRT, transcript JSON | Segment with speaker and timestamp |
.eml via the capture API | Quoted message line | |
| Chat | Slack and Discord messages | Channel and message reference |
| Tracker | What an approved ticket becomes |
|---|---|
| GitHub | Issue |
| Linear | Issue |
| Jira | Ticket |
| Notion | Database task |
How does a note become a ticket?
- Capture the note. Post it in a watched channel, mention the bot in a thread, use
/relay log, or send a file through the REST API. - Extract drafts. Relay proposes a title, type, priority and owner, each tied to a source quote.
- Review. A person edits, rejects, delegates or approves the draft.
- Push. Approved drafts enter an idempotent outbox that produces exactly one issue per draft.
Why keep a person in the loop?
Nothing reaches a tracker until a person approves it. The approval step is where a reviewer sees the source quote, the suggested owner and the destination together, then decides. Relay supports edit, reject, delegate, merge, split and bulk approval for low-risk drafts under policy. The point is not to slow work down: it is to keep the tracker a record of decisions rather than a list of model guesses.
How are owners suggested, and can you trust the suggestion?
Smart Assignment ranks up to three candidates and prints reasons beside each score. The documented signals are expertise from past work, recent context from the conversation, availability and a fairness share. A reviewer can ask why a name appeared, then accept or change it; every manual change is stored as feedback. Confidence is a prompt for a decision, not a verdict.
In an illustrative run, a billing-webhook ticket lists a candidate who owns that module, a second who reviewed recent changes, and a third on the rota. The scores are demo values, not measurements.
Illustrative demo data.
Can an AI agent create tickets?
Relay exposes a REST API and an MCP server with 28 tools. Reads are open to scoped keys, but every external write needs confirm: true on the tool call. An agent can therefore draft tickets from a transcript or summarise decisions without being able to push to a tracker on its own. A viewer key gives read-only access. The same gate applies whether the agent runs locally or against your self-hosted instance.
Self-hosted or hosted?
Self-hosting is the working path today. Relay is Apache-2.0, runs on your own database and has no feature gates or artificial limits. Hosted Cloud Free and Cloud Pro or Team plans are planned, not live: plan limits are enforced in the API, while signup and billing are opt-in per instance and the waitlist is closed. Do not expect a managed service you can sign up for today.
What are the limits?
Extraction on messy multilingual fixtures measures about 0.78 precision and 0.75 recall, so expect to edit some drafts. Live model output is noisy across runs. Live chat and tracker smoke tests still need real credentials. Audio capture needs a transcription endpoint or a supplied transcript. Dedupe and digest state are in-process, so multi-instance deployments need shared state. These are documented constraints, not hidden behaviour.
Where to next?
Each guide below covers one part of the loop in more detail, from the first note to the approval gate and the agent surface.