The Problem
Every large technical support team has a quiet but critical step: someone has to read each incoming ticket, work out which domain it belongs to, and route it to the right team. In a cloud support organization serving hundreds of engineers, tickets arrive across several closely related categories – data pipelines, batch compute, cloud infrastructure, access and permissions – and the difference between them can be razor-thin. Is this a data-tooling problem, a spot-termination failure in a batch run, an IAM issue, or a Kubernetes infrastructure fault?
Manual triage eats senior engineers’ time, delays actual handling, and sometimes misroutes a ticket so it needs an extra hop between teams.
The Solution
The team built a triage agent that runs 24/7 against the Jira board, picks up new tickets, and performs full semantic analysis using an LLM.
The agent doesn’t rely on a rigid keyword list. It understands bilingual technical context in Hebrew and English and picks up on the real signals:
- Mentions of the organization’s data-pipeline tooling route to the data platform team.
- Spot-termination errors and batch-run identifiers route to the batch compute team.
- Cloud accounts and cluster issues route to cloud operations.
- SSO, SAML, and IAM questions route to the access team.
It also handles edge cases – for example, distinguishing a local tooling error that isn’t the batch team’s problem from a genuine batch failure.
How It Works in Practice
- A new ticket with a cloud component enters Jira.
- The agent analyzes title, description, and components.
- The model produces a routing recommendation with a confidence score.
- The agent labels the ticket, assigns it to the right on-call engineer, and updates its status.
- A one- or two-sentence message goes to the team’s chat channel so everyone sees what was detected and why.
- When confidence is low or the case is ambiguous, the agent flags for human review instead of guessing.
The Results
- Triage happens within seconds of a ticket opening, instead of hours.
- Every routing decision is visible to the whole team in real time.
- Senior engineers spend far less time on manual triage and more on actually solving problems.
- The triage agent became the first gate in a growing system of specialist agents, each handling a specific request type.
Why It’s Worth Talking About
This is an AI agent making a real operational decision, not generating a text reply. By using an LLM to understand bilingual technical nuance that regex-based scripts can’t catch, it turns a tedious bottleneck into a seamless automated workflow.
