Almost every enterprise I talk to has run a GenAI chatbot pilot for IT support. Almost none of them are still running it a year later. The pattern is consistent enough that I can usually predict, within the first ten minutes of a conversation, whether a GenAI-for-ITSM initiative will survive contact with production.
Decision 1: Classification, Not Conversation
The highest-value, lowest-risk GenAI use case in ITSM isn't a chatbot users talk to — it's a classifier that runs silently on every incoming ticket. Accurate classification and prioritization compounds across thousands of tickets without ever exposing an end user to an AI interaction that might go wrong.
Decision 2: Confidence Thresholds, Not Auto-Everything
Production systems route low-confidence classifications to a human, full stop. The temptation to auto-resolve everything the model touches is exactly how pilots become incidents. On Smart Desk, the AI ticket triage system I built using Claude API and n8n, only well-understood, low-risk categories were enabled for auto-resolution initially — everything else routed through normal human-reviewed paths.
Decision 3: Draft, Don't Send
For anything customer- or stakeholder-facing — incident updates, resolution summaries, RCA drafts — the AI should produce a draft a human reviews and sends, not an autonomous message. This single design choice is the difference between a tool operations teams trust and one they quietly route around.
"The fastest way to kill trust in an AI tool is one bad autonomous action nobody approved."
Decision 4: Integrate, Don't Replace
Every successful implementation I've built sits inside the existing ServiceNow workflow rather than next to it. Tickets still live in ServiceNow. Escalations still follow the existing matrix. The AI is a layer that makes the existing system faster, not a parallel system competing for the team's attention.
The chatbot demo and the production system solve different problems. Demos are built to impress; production systems are built to survive a bad day. Design for the second one from the start.
None of this is exotic engineering. It's mostly discipline about scope and a refusal to let the AI take actions that haven't earned trust yet. That discipline is also exactly what turns a pilot into something that's still running — and still saving real hours — a year later.