Artificial intelligence in IT support is presented as an unavoidable step. The demonstrations are convincing: instant conversational agents, automatic classification, resolution with no human involvement. The reality of deployments is more nuanced. The projects that work have one thing in common: they did not start with the AI — they started with the process.
The uses with proven value share one prerequisite: a process that is already structured
- Automatic classification of requests: effective when the monthly volume is sufficient and the categories are coherent. The system classifies correctly what the teams have qualified correctly in the past — it does not rescue a badly structured history.
- Suggested answers and a searchable knowledge base: useful when the base is up to date and well structured. An outdated base produces suggestions nobody can use, and that is the most common obstacle on this kind of project. It is exactly what a RAG and knowledge management approach addresses.
- Automation of first-line requests: password resets, account unlocking, standard access requests. These cases offer the fastest return and the lowest risk — and they account for a large share of first-line volume.
What AI does not do well: understand ambiguous requests, handle exceptions, deal with complex multi-system problems. Human supervision remains indispensable, at least to check the quality of the automated answers through the first months.
AI does not improve a failing process — it industrialises it
The first cause of failure is not technical: it is data quality. A system fed with badly qualified requests — incoherent categories, vague descriptions, undocumented resolutions — will produce mediocre results however sophisticated the algorithm.
If the service levels are not defined, if the escalations are not clear, if the knowledge base is not maintained, introducing AI will only speed up the existing dysfunctions. An automatic resolution system that answers beside the point generates more frustration than slow human support.
Before choosing a solution, one question is enough
Is our support process structured enough for AI to improve it? If the answer is not a clear yes, that is where to start: coherent request categories, defined service levels, a knowledge base kept up to date. Structuring the service catalogue and the handling circuits belongs to digital workflow before it belongs to AI.
Once those foundations are in place, the natural sequence is to start with repetitive first-line requests, measure the results on a narrow perimeter, then widen. Putting a feedback mechanism in place from launch — so that agents can correct the system's mistakes easily — is what keeps quality up over time.
