Newsletter
Stay Ahead
Occasional, engineering-led notes on applied AI — what we're building, running, and learning in production.
Occasional, engineering-led notes on applied AI — what we're building, running, and learning in production.
Use case: Route support tickets by intent, urgency, and topic
Support inboxes and ticket queues mix billing questions, bug reports, sales enquiries, and outright emergencies in one undifferentiated stream, and manually tagging each one before routing it to the right team is a chore nobody wants. Webanto's chat endpoint can classify each inbound message against a fixed taxonomy you define, intent, urgency, topic, and return the label as JSON in a single call, so routing logic downstream just reads a field instead of guessing from free text. Because Qwen3.6 reasons before committing to a label, it copes better with sarcasm, mixed-intent messages ('also, is this normal, and also please cancel') and non-native phrasing than a keyword-matching rule set, though edge cases still need a fallback queue rather than blind auto-routing.
Every inbound ticket gets a consistent intent, urgency and topic label without a human triaging the queue first.
Mixed-intent messages, a billing question buried inside a bug report, get flagged rather than silently routed to only one team.
Classification runs fast enough to sit inline in your ticket-creation webhook, so routing happens before an agent ever opens the ticket.
A fixed taxonomy keeps your reporting consistent over time, instead of drifting as different agents apply tags differently.
Define your taxonomy up front: a closed list of intents (billing, bug, feature-request, sales, complaint), a 3-level urgency scale, and a topic list. Closed lists classify far more reliably than open-ended labels.
Write the system prompt with the taxonomy spelled out as an enum and an instruction to return only JSON matching { intent, urgency, topic, confidence }.
Call POST /api/v1/chat with model 'qwen', the ticket subject and body as the user message, and a reasonably tight max_tokens — classification output is short, but still allow headroom for reasoning tokens or you'll get a truncated or empty response.
Hook the call into your ticket-creation webhook, or a queue consumer, so classification happens the moment a ticket lands rather than on a batch delay.
Add a confidence threshold: route anything below it to a manual-triage queue instead of trusting a low-confidence auto-classification.
Budget credits per ticket: a typical support email, subject plus a few paragraphs, runs around 800-1,200 total tokens with reasoning, comfortably under 1 credit at 2,000 tokens per credit, so a queue of 5,000 tickets a month stays well inside the Starter plan's 10,000 monthly credits.
Sample and spot-check classifications weekly against what your agents actually did with the ticket, and adjust the taxonomy or prompt wording where the model and your team disagree.
For the bulk of routine tickets, yes — clear billing or how-to questions route reliably. For anything ambiguous, sarcastic, or legally sensitive, such as a complaint threatening legal action or a safety issue, keep a confidence threshold that routes to manual review. Full auto-routing without a human safety net for edge cases isn't recommended.
Spell the taxonomy out explicitly as an enum in the prompt and ask for JSON only, then validate the returned value against your allowed list in code. If it returns something outside the list, treat that as a low-confidence result and route to manual triage rather than accepting it.
It weighs the actual content and context rather than just urgency-flavoured words, which tends to be more reliable than a keyword match on terms like 'urgent' that get overused. It isn't infallible on subtle cases, so keep the confidence threshold and manual fallback in place.
Credits are metered by usage, not seats, so a quiet month costs less and a busy month costs more. You only need to upgrade plans if you're consistently running into the monthly cap; check your usage dashboard against your plan limit.
Yes, Qwen3.6 handles multiple languages, though classification accuracy on less common languages hasn't been benchmarked internally, so it's worth spot-checking non-English tickets more closely at first.
Shoppers who spot a product in the wild, on someone else's feed, or on a competitor's site rarely know the right keywords to find it in your catalogue. Visual search closes that gap: a shopper uploads
Large media libraries accumulate redundancy fast — the same product shot re-exported at a different crop, a stock photo uploaded twice under different filenames, a banner re-saved at a lower compressi
Reverse image search answers a narrower question than catalogue browsing: given one photo someone just handed you — a screenshot, a competitor's product shot, a customer's snapshot of a damaged item —
Start a 14-day free trial on any paid plan. No credit card required.