AI intake

What it does

With AI intake on, every new ticket from the sources you choose is sent to a language model together with your category list and your own guidance. The model returns a category, priority, severity, size (1–5), a one-line summary and a reason. Intake never waits on the model; triage runs in the background a few seconds after the ticket exists.

• Suggest mode shows the result as a card on the ticket with an Apply button and a differs / matches tag. Start here — it tells you how often you'd have agreed.
• Apply mode sets the fields on arrival and leaves an internal note saying what changed and why. Switch to it when you'd trust the model with a new hire's judgment.

Ticket text goes to whatever endpoint you configure. With a model on your own hardware, nothing leaves the building.

Setting it up

Admin → AI intake. You need an OpenAI-compatible chat endpoint:

• Ollama on a PC with a GPU (the local option). Install Ollama, run ollama pull qwen3:14b, set the environment variable OLLAMA_HOST=0.0.0.0 so other machines can reach it, and restart it. Endpoint: http://<that-pc>:11434/v1, model qwen3:14b, no key. A 16 GB card runs a 14B model comfortably; on 8 GB try qwen3:8b.
• A hosted API (OpenAI-compatible): endpoint https://api.…/v1, your key, and the provider's model name.

Save, then Test connection. Then tick the intakes it should watch (web form, email, staff-created), choose the mode, and Save again.

Shop guidance

The free-text box is the most useful control. Write your rules of thumb in plain English:

• Anything mentioning the walk-in freezer or cooler is critical priority and major severity.
• Facilities handles plumbing and HVAC; IT never does.
• Password resets are size 1, low priority.

Add a line each time you catch it getting something wrong.

Backlog and re-runs

Triage next N waiting runs the model over open tickets it hasn't seen. Every ticket has Ask again. The status box shows how many were triaged, applied, the average response time, and the last error.