Back to Dashboard

Documentation

Learn how to use Asyntai

Getting Started
Installation Knowledge Base Test Your AI Manage Websites
Customization
Widget Colors AI Assistant Name Initial Message Widget Position Auto-Trigger Greeting Bubble Unread Badge Instruction Mode Ask Questions Use Emoji Collect Emails Collect Phone Numbers Smart Lead Capture Localization Hide Branding Privacy Link Custom Notice Transcript Download New Conversation Conversation History Tagline & Status Dot Speech to Text Chat Icon Company Logo Character Limit Temperature
AI Instructions Tips Switching From Another Chatbot Compare Plans
Features
Website Crawl Knowledge Gaps Product Cards Dynamic Product Cards Dynamic Images User Context Custom Tools Link Parameters Live Monitoring Human Takeover Escalation AI Notifications Daily Report Real-Time Data Feed Real-Time Data Feed Max Team Members Single Sign-On Two-Factor Authentication Include Images Image Vision Translation Widget Localization AI Transparency Leads Smart Lead Capture Support Tickets Bookings Embeds Exclude Pages Blocked IPs Retention Policy Zero Retention Mode PII Redaction Output Classifier Widget Version Pinning Audit Log Smarter Model Enable Thinking Reply Suggestions Follow-up Messages Speech to Text Transcript Download Embedded Chat
Analytics
Chat Analytics Chat Logs
Integrations
REST API Instagram Messenger WhatsApp Discord Zapier Email Mobile Apps
Other
Add to Home Screen Custom Open Button Ask AI Buttons Payment Failed?
Ask AI

Output Classifier

A second AI model checks every reply before it reaches the visitor

Go to Output Classifier
Enterprise

Overview

Normally the AI's reply goes straight to the visitor. With the Output Classifier on, a second, independent model reads every reply first and decides whether it is safe to send: did the bot get tricked into breaking its role, is it giving regulated advice, is it engaging with a topic you have forbidden? A flagged reply never reaches the visitor as-is — it is rewritten once, and if the rewrite is still flagged, your fallback message is shown instead.

How It Works

  1. The AI answers as usual - nothing about the main model changes
  2. The classifier reads the draft - a fast second model checks it against your enabled categories
  3. Flagged replies get one rewrite - the violation is fed back to the main model, which writes a corrected reply that is checked again
  4. Fallback as the last resort - if the rewrite is still flagged, the visitor sees your safe fallback message instead
Output Classifier
Draft reply Flagged
Visitor sees Corrected reply

What It Checks

  • Persona breaks and prompt injection - always on: the bot obeying instructions hidden in a visitor message, revealing its own rules, or dropping its support role
  • Medical advice - diagnosis, treatment or dosage recommendations; product information stays allowed
  • Legal advice - interpreting law for the visitor's situation; stating your own policies stays allowed
  • Financial advice - investment or personal finance recommendations; your own prices and fees stay allowed
  • Your forbidden topics - a free-text list you write, one topic per line — this is where you define what is sensitive for your business

The Incident Log

Every flag is recorded on the settings page: the category, the classifier's reason, the blocked draft and what the visitor saw instead. It is both your tuning tool for false positives and the audit trail you can show your own compliance team. Incident content goes through PII Redaction like every stored surface, and under Zero Retention Mode only the metadata is kept — no message content.

If the Classifier Itself Fails

A second model means a second thing that can time out. You choose what happens then:

  • Fail open (default) - the unchecked reply is delivered; availability wins
  • Fail closed - the fallback is shown instead, for sites where an unchecked reply is worse than no reply

Good To Know

  • It only acts on the reply, never the question - if your bot already refuses or answers safely, nothing is flagged, and that is correct. The classifier is a safety net for the rare reply that slips through, not a check on what the visitor asked. Trying to jailbreak a well-behaved bot will usually show no flag because the bot handled it on its own.
  • Latency - the check adds roughly half a second to every reply, and a flagged reply pays for a full rewrite on top
  • Not a guarantee - this is one model judging another; it is an additional independent layer, not an absolute barrier
  • Test chat shows verdicts - the dashboard test chat enforces exactly like the live widget and tells you when a reply was flagged, so you can try to trick your own bot
  • Write the fallback in your visitors' language - it is shown verbatim; the built-in default is English
  • Works alongside your AI Instructions - instructions steer the first model, the classifier catches what slips through; use both

Pairs with PII Redaction: one feature controls what gets stored, the other controls what gets said.

PII Redaction Personal data is masked before a message is stored anywhere

Note: The Output Classifier is part of an Enterprise agreement. Email [email protected] and we will walk you through it.