Classify it, draft it, and leave the decision to your team
A support queue is three jobs in one: work out what each message is about, route it, and write a reply. A model does the first two well and the third one adequately, under supervision.
Munito AI serves them on European infrastructure, with zero data retention by default, which matters because a support inbox is full of customer data.
What to automate, in the order that works
| Step | What the model does | Who decides |
|---|---|---|
| Classify | Topic, product area, language, urgency | Your rules |
| Route | Pick the queue and the priority from the classification | Your rules |
| Draft | Write a reply from your own articles and the ticket history | Your agent |
| Summarize | Compress a long thread before a human opens it | Nobody, safely |
Start with classification. It is the step with a measurable right answer, the lowest risk, and the fastest payback, and it tells you whether the model understands your domain before you let it write anything.
Use structured outputs so a classification arrives as fields, not prose, and your router needs no parser. The classification recipe and the customer support recipe are working starting points.
Drafts, not sends
A drafted reply that an agent edits is a productivity tool. A sent reply that nobody read is a liability. Keep a person between the model and the customer until your own numbers say otherwise, and measure the edit distance: when agents stop changing the drafts, you have earned the next step.
The economics
Support volume is steady and large, which is the shape that rewards a cheap model and a cached prompt.
Your instruction block, your tone guide and your article index are the same on every ticket. On DeepSeek-V4.1-Flash-1017 a cached prefix costs $0.0060 per million tokens against $0.30 uncached, so the fixed part of each request is close to free and you pay for the ticket.
A backlog is a batch job. The batch tiers cut the price of work that can wait overnight. See Service tiers and Pricing.
Tickets are personal data
Names, addresses, order numbers, sometimes health or payment details. A support pipeline is one of the easiest ways to send customer data somewhere you did not intend.
Send to https://api.eu.munito.ai and the request runs in the EU; the response names the region in the munito-region header. We keep no prompts and no outputs by default, train on nothing you send, and run no human review of your content. The DPA applies with no signature, at legal.munito.ai/dpa.
If your helpdesk sits inside your own cloud, the platform can too. See Enterprise.
Measure it against the queue you already have
Take a thousand closed tickets. Score the classification against the labels your team gave them, and the drafts against the replies your team sent. Then measure the two numbers that pay for the project: first-response time, and the share of tickets closed without a second human touch.
Multilingual queues are where open models earn their place: one model reads and writes the languages your customers use, so one pipeline serves every market.
Start
- Classify a message
- Build the support loop
- Answer from your own articles
- Talk to an engineer about your queue and your languages