Every ITSM vendor now claims AI will transform service delivery. Some of these claims hold up. Most don't. Here's an honest breakdown of what AI can actually do in ITSM today, what it can't, and how tools like GLPI fit into the picture.
What AI does well in ITSM right now
Ticket classification from natural language
An LLM can read a ticket description like "my laptop screen flickers when I connect the dock" and correctly assign it to Hardware > Display > Docking Station with roughly 85-90% accuracy. This is a real, deployable use case. The model reads the text, matches it against your category tree, and writes the classification back via API. It saves dispatchers time on the 80% of tickets that are straightforward.
Suggested solutions from knowledge base
When a ticket arrives, an LLM can search your knowledge base and suggest relevant articles to the agent. This works best with well-maintained KB articles that have clear titles and structured content. It works poorly with a KB full of outdated or duplicate articles -- the AI will confidently suggest wrong answers.
Chatbot for L0 self-service
A chatbot that handles password resets, VPN instructions, and "how do I install the printer" questions can deflect 15-25% of tickets before they reach a human. This is a genuine workload reduction, not a gimmick. The key is limiting the chatbot's scope to topics where wrong answers are harmless and correct answers are well-documented.
Pattern detection in incident data
Analyzing incident data over time to spot trends -- "printer failures spike every Monday morning" or "VPN tickets increase after each Windows update" -- is a legitimate AI application. It requires clean, consistently categorized historical data, but it surfaces insights that humans miss because no one reads 10,000 ticket records manually.
What AI cannot do reliably
Replace L2/L3 diagnosis
Complex troubleshooting requires understanding network topology, application dependencies, and organizational context that no model has access to. AI can suggest possibilities, but it cannot replace the engineer who knows that the ERP server shares a VLAN with the backup system and they conflict during nightly jobs.
Make judgment calls on changes
Change advisory decisions involve risk assessment, business impact analysis, and political awareness. An AI cannot evaluate whether deploying a patch during quarter-end close is acceptable. These decisions require human judgment and organizational knowledge.
Handle edge cases
The 10% of tickets that don't fit any pattern -- the ones that make experienced agents pause and think -- are exactly where AI fails. It's trained on the common cases. Edge cases are, by definition, underrepresented in training data.
Where GLPI fits
GLPI is not an AI engine. It's the data source and the integration point. A practical AI integration with GLPI looks like this:
- New ticket arrives in GLPI via email or self-service portal
- A webhook or scheduled script picks up the ticket via GLPI's REST API
- An external service (n8n workflow, custom Python script, or cloud function) sends the ticket description to an LLM
- The LLM returns a category, priority suggestion, and optional KB article links
- The script writes the results back to GLPI via API -- updating category, adding a note with suggestions
The entire flow runs outside GLPI. GLPI just provides the data and accepts the results.
The prerequisite nobody mentions
AI on messy data produces messy results. If your GLPI categories are inconsistent, your ticket descriptions are one-word entries like "broken," and your resolution notes are empty, no amount of AI will extract useful patterns. Clean your data first: standardize categories, enforce description quality, and require resolution notes. Then AI has something to work with.