What is AI ticket triage?
Every support request has to be read and sent somewhere: billing, shipping, technical support, complaints. In most companies someone does this by hand, scanning a shared inbox or a help desk queue and moving messages along. AI ticket triage does that step with a large language model (LLM). The model reads free text, compares it with your list of categories and assigns the request.
Keyword rules have done a version of this for years. They break as soon as a customer writes "the payment document never showed up" instead of "invoice". An LLM copes with badly written messages, requests that mix two problems and synonyms nobody planned for.
Triage is the first link in the support chain. Drafted replies, priority and escalation all come after it, so a wrong first assignment delays everything downstream.
How accurate is automated ticket routing compared to human agents?
My best number comes from a project I led as tech lead at a major Italian bank. Its internal help desk receives more than 1 million tickets a year, written by employees in free form, and each one must land in one of 400+ hierarchical categories. The LLM agent my team of 4 built picks the right category 81% of the time. Manual routing was at 75%.
Six points sounds small. Across a million tickets it is roughly sixty thousand fewer misroutes a year, each of which used to bounce between teams. After go-live, average resolution time fell from 5 or 6 days to 2 or 3. The full write-up is in the LLM ticket routing case study.
The reason the model beats people is mundane. No agent remembers 400 category definitions. For each ticket the system retrieves the definitions of the likely categories (a technique called retrieval-augmented generation, or RAG) and chooses among them. It also reads ticket number 900 of the day as carefully as the first.
What are the options for automating ticket routing?
There are three realistic routes. Which one fits depends mostly on how many categories you have and how sensitive the data is.
| Option | Good fit | Limits |
|---|---|---|
| Keyword rules (inbox filters, help desk automations) | A handful of categories, very standard requests | Break on new phrasing; constant upkeep |
| Built-in AI triage in your help desk platform | Dozens of categories, flat taxonomy, no special data constraints | Little control over where data runs and how errors are measured |
| Custom LLM classifier with RAG over your taxonomy | Hundreds of categories, hierarchies, regulated data, high volume | Needs labelled history and a few weeks of work |
Keyword rules still earn a place as a safety net, for example to send anything that mentions legal action straight to a person. The classifier handles the rest.
What do you need before automating support ticket triage?
- A history of tickets that were already routed, even an imperfect one. It is the yardstick for measuring the system before launch.
- A written list of categories with a one-line definition each. Often it only exists in the heads of the people routing today.
- An environment where the model can run on your data. At the bank we used Azure OpenAI inside the client's EU cloud environment, and sensitive data stayed out of the logs.
- Someone who owns the decision on uncertain cases.
If tickets contain customer personal data, your usual data protection obligations apply (GDPR in Europe): a lawful basis, a processor agreement with the model provider, and data stored where the contract says. Have your compliance team confirm how the EU AI Act classifies your use case.
How much does AI ticket triage cost?
There are two costs: building it and running it. The build depends on the number of categories, your help desk integration and how clean the history is. Running cost is mostly model calls, which are cheap per short ticket but scale with volume.
Compare it with what manual triage costs today: the hours spent reading the shared queue, plus the time lost when tickets reach the wrong team. At the bank the second cost was the larger one, because every misroute added days to resolution.
What mistakes do teams make when they automate ticket routing?
- Starting without measuring how accurate people are today. Without the 75% baseline, the bank's 81% would have meant nothing.
- Putting every category in the prompt. With 400 categories that does not work; retrieve only the relevant ones.
- Trusting every prediction. Low-confidence tickets should go to a person.
- Leaving categories undefined. The model fails where agents fail too: between categories nobody ever told apart clearly.
- Never checking again after launch. New products and team changes mean the taxonomy needs updates.
Where should you start with AI ticket triage?
I would start with an export of a few thousand tickets that were already routed. Run the classifier on them and compare its choices with the human ones. Within a couple of weeks you know whether it beats your current routing and which categories it confuses. If it does not, you have spent little and ended up with a cleaner taxonomy.
To see a system like this in production, read the bank help desk case study. For a custom build, see AI agent development, and for regulated finance, AI consulting for banks.
Want to know if AI can triage your support queue?
Send me a short description of your queue: volume, number of categories, help desk tool. I will tell you whether a backtest on your history is worth running.
Let's talk
