insights

The most important line in our triage agent: the agent can’t write

The most important design decision in our AI ticket-triage agent: the agent itself can’t write to the ticket.

It classifies the ticket and returns a confidence score. A plain, deterministic step then decides what happens:

  • If confidence is above the threshold, the fields get set and routing takes over.
  • If it’s below, nothing changes and a human triages as usual.

Why this split works:

  • The model can be wrong without being dangerous.
  • The threshold is one number we can tune from real accuracy data, field by field.
  • Audits are easy, because every write comes from code we can read.

“Let the LLM call the API” is the fastest demo and the slowest path to production. Put the judgement in the model and the action in boring code.

All insights