IT Service Management Modernisation: Where AI Helps and Where It Does Not
AI deflection works when the knowledge base is good. Otherwise it just delays the ticket.
Service desks are an obvious target for AI: high volume, repetitive requests, and abundant historical data. Results in practice are mixed, and the split is predictable. Where the underlying service management fundamentals are sound, AI produces substantial gains. Where they are not, it produces a chatbot users learn to bypass.
What works reliably
| Application | Typical effect | Depends on |
|---|---|---|
| Categorisation and routing | Fewer misroutes, faster assignment | Consistent historical categorisation |
| Knowledge retrieval for agents | Faster resolution, less escalation | Accurate, maintained articles |
| Response and resolution note drafting | Less administrative time per ticket | Reasonable ticket data |
| Ticket summarisation for escalation | Cleaner handover, less customer repetition | Notes captured during handling |
| Self-service deflection | Lower volume on documented issues | Good knowledge base and clear escalation path |
| Incident clustering | Earlier detection of a common cause | Consistent data capture |
Routing is the most reliable early win because it is bounded, measurable, and low-risk. Deflection is the most attractive and the most dependent on prerequisites.
Knowledge quality is the constraint
Retrieval-based assistants can only answer from what exists. Most organisations discover their knowledge base covers the issues that were easy to document rather than the ones users actually raise.
Fix that with data: take the top 50 request types by volume, check whether a current, accurate article exists for each, and write the missing ones. Assign owners and review dates, because an assistant confidently surfacing a two-year-old article is worse than surfacing nothing. This work delivers value on its own and is a prerequisite for anything automated.
Design the escape hatch first
User tolerance for self-service collapses when there is no visible way to reach a human. Make escalation available at every step, pass the full conversation context into the resulting ticket so nothing is repeated, and set the assistant's scope honestly — admitting it cannot help is a better outcome than a confident wrong answer.
Measure deflection honestly too: a session only counts as deflected if the user did not raise a ticket about the same issue shortly afterwards. Vendor-reported deflection numbers routinely ignore this.
Automate fulfilment, not just conversation
The larger prize is not answering questions but completing requests: password and access resets through verified identity flows, software provisioning within licence policy, standard access requests with approval routing, and environment or resource provisioning through existing automation.
Each of these requires an integration and an authorisation model, which is why they are often skipped in favour of a conversational layer. They are also where the measurable time saving actually is.
Do not automate around root causes
The risk of efficient ticket handling is that recurring problems become cheap enough to tolerate. If the same incident is resolved 400 times a month with impressive speed, the win is eliminating the cause, not accelerating the response.
Keep problem management explicitly funded, and report repeat incident volume alongside handling metrics so leadership can see the difference between fast service and fewer failures.
Measure what matters
Track genuine deflection rate, first-contact resolution, mean time to restore, reopen rate, misroute rate, repeat incident volume, and cost per ticket including AI running costs. Report user satisfaction separately for AI-handled and human-handled interactions — a blended figure hides exactly the signal you need to tune the programme.
Frequently asked questions
Where does AI genuinely help ITSM?
Categorisation and routing, agent knowledge retrieval, response and resolution note drafting, ticket summarisation, and deflection for well-documented repetitive requests.
Why does deflection often fail?
Because it depends on knowledge base quality; missing or outdated articles for common issues mean the assistant only delays the ticket.
How should this be measured?
Genuine deflection (no follow-up ticket), first-contact resolution, time to restore, reopen and misroute rates, and repeat incident volume.
What should be automated first?
Routing, then request fulfilment for high-volume standard requests such as access and provisioning — that is where measurable time is recovered.