Jev 1.13 for business classification and routing

Separate the questions that lead to different actions

A complaints queue holds cases about late delivery, a wrong invoice and damaged goods, and sometimes one message concerns all three. A simple keyword rule starts to fail, and asking a large model about every case is expensive. Jev 1.13 on OpenRouter returns typed decisions with probabilities. The scenarios below are a proposal for a pilot.

For a complaint message you ask separate questions: which department handles it, what is missing and whether escalation is needed. Each answer means something different to the application. The department sends the case to a queue, and a missing document triggers a request to the customer. You allow for ambiguous situations. If the task forces one category when there are several problems, you set a priority or split the case, because without that rule even people will not agree on the right answer.

What classification and routing cost and when they pay back

The count covers the work around a case: searching, reading, writing and passing it on. What stays is the time to check the output, 50% of today's work in the scenario. Before we quote, we measure both times on 20 to 100 cases your team has already handled, in a €0 test that takes 5 working days.

We leave 50% of today's work in the count because in the example half of the cases stay below the threshold and go to an operator. A decision model is cheap per case, so the payback depends mainly on the threshold's coverage.

When a rule or a large model is enough

If the categories follow from fixed form fields, a rule is enough. If cases are rare and very varied, a large model with instructions can be simpler to maintain than a small model with a threshold. A small decision model makes sense for a large, repetitive queue, where the cost and response time of a large model grow with the number of cases. Hugging Face explains how to choose the dataset and the evaluation measures, and you choose the threshold on cases that were not used for tuning. Running small models locally is covered in the article on models inspired by Jev.

Where to start

The pilot covers one queue, for example complaints, compared with today's assignment. We choose the threshold together with the queue manager, because it is a decision about how work is shared. We build this kind of rollout as part of AI automations. For the free process review, a description of one case from recent weeks and the tools it went through is enough. You do not need to send confidential documents, and you get a written plan within 2 working days.