Ticket triage is a strange corner of ITSM: it's high-volume enough to demand automation, but every request is worded differently enough that keyword-matching categorization rules max out at maybe 60–70% accuracy in practice. That gap between "rule-based" and "actually understands the request" is exactly where agentic AI earns its place.
What a rules engine gets wrong
Classic triage automation matches keywords or regex patterns against the ticket description to assign category, priority, and assignment group. It works fine for the requests users already know how to phrase in the system's expected vocabulary, and fails silently on everything else — a user describing "my laptop won't connect to the VPN after the update" doesn't necessarily hit a "network" keyword rule, and gets misrouted to a generic queue where it sits until a human re-triages it.
What an agent does differently
An agentic triage system reads the full ticket text, reasons about intent rather than matching surface keywords, and can take actions beyond classification — querying the CMDB for the user's assigned assets, checking if there's a known related incident already open, or asking a clarifying question back to the requester before finalizing a category. It's not just smarter pattern matching; it's a system that can gather more information before deciding, the way a human triage agent would.
tools available to the triage agent:
- search_knowledge_base(query)
- get_user_assets(user_id)
- find_related_incidents(description)
- assign_category(category, priority, group)
- ask_clarifying_question(question)
Where the guardrails matter
The risk with letting an agent triage autonomously is confidently wrong routing at scale — a rules engine fails in predictable, debuggable ways; an agent can fail in ways that look plausible until someone actually checks. The mitigation that works in practice:
- Confidence thresholds with human fallback. Below a set confidence level, route to a human triage queue instead of guessing — an honest "I'm not sure" beats a wrong assignment.
- Explainability logged with every decision. The agent should log why it picked a category, not just the category — this is what makes a wrong routing debuggable instead of a mystery.
- A feedback loop from re-routes back into evaluation. Every time a human re-categorizes a ticket the agent triaged, that's a labeled example — feed it into your eval set, not just a shrug.
The honest tradeoff
Agentic triage isn't more reliable than a well-tuned rules engine on the cases the rules engine was built for — it's more reliable on the long tail of phrasing the rules engine never anticipated, which in most real ticket queues is a third or more of total volume. That's the actual business case: not replacing what already works, but covering what doesn't.
# comments
loading comments...