ITSM in the Age of AI Agents
Autonomous agents are collapsing incident triage, change approvals, and knowledge management into seconds. What survives is the operating model — and most were never designed for this speed. A field guide to redesigning ITSM before the agents redesign it for you.
The queue is the first casualty
Traditional ITSM is built around a queue: tickets arrive, humans triage, work is assigned, status trickles back. Every step assumes human reading speed. An AI agent that classifies, prioritises and routes in seconds does not make the queue faster — it makes the queue irrelevant for a large share of volume.
The practical consequence: the metrics most service desks report — first-response time, backlog age, triage accuracy — stop measuring the bottleneck. The bottleneck moves to exception handling: the tickets the agent is unsure about, and the ones it should never touch.
What agents are genuinely good at today
Notice what is not on the list: judgement calls under ambiguity, changes with blast radius, and anything where being wrong is expensive. Those stay with people — and the operating model must say so explicitly, in writing, per workflow.
- Classification and routing against historical ticket data — consistent, tireless, and better than a rotating duty roster.
- Drafting knowledge articles from resolved incidents — humans approve, agents assemble.
- First-line resolution of known, low-risk request types: password resets, access requests with policy-based approval, status queries.
- Summarising long ticket threads for the engineer picking up a handover.
Redesign the operating model, not just the tooling
The organisations getting value from service-desk AI share one habit: they redesigned roles before they switched on the automation. Agents handle the flow; humans handle the exceptions, the approvals that carry risk, and the continuous correction of the agent's behaviour.
That means new responsibilities that most ITSM org charts do not have yet: someone owns the agent's accuracy the way a service owner owns a service. Someone reviews what the agent resolved and what it escalated, weekly, with the authority to change its instructions. Without that loop, automation quality drifts quietly until users route around it.
A sensible adoption order
- Start with classification and routing — measurable, reversible, low risk.
- Add knowledge drafting with human approval — builds the article base agents will later answer from.
- Automate resolution for one well-understood request type — prove the pattern end to end.
- Only then expand, one workflow at a time, each with a named human owner and an explicit rollback.
The takeaway
AI does not remove the operating model from ITSM — it exposes whether you had one. Automate the flow, keep humans on the exceptions, and assign someone to own the agent's behaviour like a service.
Related
AI & workflow automation in practice