AI Customer Service Routing Engine

Routing Pipeline

Incoming volume 40/min

Bot confidence threshold 0.70

Classifier accuracy 0.85

Human agents 4

Negative-sentiment escalation 0.60

Automation rate

–

Avg resolution time

–

Agent queue depth

–

CSAT proxy

–

How it works

Every message spawning at the intake pad carries a hidden intent (billing, technical, order status or complaint) and a sentiment. It first passes through the NLP intent classifier, which produces a confidence score — how sure the model is about what the customer wants. That score is compared against the confidence gate's threshold: above it, the message is routed to the bot for instant self-service resolution; below it, or whenever a negative-sentiment message trips the escalation rule regardless of confidence, it joins the human agent queue instead.

confidence ≈ accuracy ± noise  (higher when the classification is actually correct)
route = bot        if confidence ≥ threshold AND NOT forced-escalation
route = human queue otherwise
forced-escalation ~ Bernoulli(sentimentSensitivity) when sentiment is negative

Incoming volume — how many messages arrive per minute; push it high enough and the agent queue starts backing up.
Bot confidence threshold — how sure the model must be before it's trusted to answer alone; raise it to protect quality at the cost of more escalations.
Classifier accuracy — the underlying NLP model quality; a low-accuracy model still produces confident-looking scores sometimes, so it silently mis-resolves a fraction of bot answers.
Human agents — the size of the escalation pool; too few and the queue grows even at a threshold that looks reasonable on paper.
Negative-sentiment escalation — the probability an angry-sounding message gets routed straight to a human regardless of the model's confidence, a common real-world safety rule.

Real-world relevance: this is the automation/quality trade-off every deployed support chatbot has to tune — a threshold set too low resolves more automatically but silently damages CSAT on mis-classified cases, while a threshold set too high floods human agents and erases the labor savings automation was meant to deliver.

Billing Technical Order status Complaint

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.