Small local AI models for internal ticket routing

The morning shows whether the hardware is enough

In the morning the helpdesk receives requests to set up a printer and for access to an application, written in employees' own words. Desktop support handles printers, and the application's administrators handle access. The model shows a short team suggestion next to the ticket, the duty agent sees the text and the proposal side by side and approves the assignment. For this work, the team name is enough.

A single suggestion on a quiet computer shows only part of the work. When many people start the day within a short span, later cases wait for the model. What counts is the time from the ticket appearing to a usable suggestion, including at the start of work, when the model is loading into memory. A correct suggestion that arrives after the cases have been split by hand comes too late.

What a small local model for ticket routing costs and when it pays back

The count covers the time spent searching, reading and compiling the material. What stays is the time to check the output against the source, 25% of today's work in the scenario. We measure both times in a trial on 20 to 100 cases the team has already handled. According to Brynjolfsson, Li and Raymond (2025), customer service agents supported by AI resolved on average 15% more issues an hour. In a helpdesk, suggesting the recipient is a narrower task, so the count covers only the time spent on assignment.

What runs locally and what needs agreeing

The Ollama documentation describes how parallel processing depends on available memory: several requests at once need extra memory, and an overloaded queue can reject the next request. The word „small” does not tell you how fast the helpdesk will get the last suggestion of the morning wave. Ollama's local mode switches off cloud models, and it does not set where the copy of a ticket sits in the helpdesk system or how the integration works. If a suggestion appears after a manual assignment, the app does not overwrite the duty agent's decision.

Some tickets can skip the model

When the form contains the chosen service, an existing rule names the team. GLPI describes ticket rules with conditions and actions on creating or changing a ticket, run in a set order. The model stays with cases described in free text where the rules did not find a recipient. If the trial shows repeated mistakes despite a clear split of responsibility, fine-tuning for ticket classification is a separate question.

Where to start

The trial covers a narrow group of routine cases and the hours of highest inflow, on the hardware intended for the work. The helpdesk team keeps its usual duty rota even when the model is unavailable. We roll out local models as part of private LLMs, and we build the view for working with sources as an AI app. For the free process review, a description of one case and the places where the material sits is enough. Confidential files stay in the company, and you get a written plan within 2 working days.