Intent Boundary Confusion

Issue: Similar Intents Misclassified Due to Overlapping Definitions

Frequency: Common

Symptoms

• “Not interested” treated as “Do not contact”
• “Busy now” routed to rejection instead of callback
• Hesitant responses marked as clear decline
• Soft objections escalated to hard objections
• Qualification status inconsistent across similar calls

Root Cause Voice agents often have multiple intent categories with subtle semantic boundaries. Without clear, mutually exclusive intent definitions, the classifier—whether rule-based or LLM-based—confuses adjacent intents. A caller saying “I’m not sure” could be hesitant (continue pitch), not interested (stop pitch), or requesting more info (provide details). Overlapping intent instructions compound the problem.

Example

Intent definitions in system:
- Not Interested: "Caller declines or refuses the offer"
- Do Not Contact: "Caller asks for removal or no future contact"
- Callback: "Caller asks to be called later"

Scenario 1: Ambiguous decline
Caller: "I don't think this is for me right now"

Interpretation A: Not Interested (decline the offer)
Interpretation B: Callback (not right NOW = call later)
Interpretation C: Hesitant (needs more info)

Agent chose: Do Not Contact ✗ (over-escalation)
Correct: Not Interested or Hesitant

---

Scenario 2: Busy vs rejection
Caller: "No, I can't talk right now"

Agent interpreted: Not Interested ("No, I can't")
Correct: Callback (can't talk NOW = busy)

---

Scenario 3: Soft vs hard objection
Caller: "Please don't WhatsApp me"

Agent interpreted: Not Interested (declined offer)
Correct: Do Not Contact (no-contact request)

---

Intent confusion matrix (500 calls):
                    Classified As:
                    NotInt  DNC   Callback  Hesitant
Actual NotInt        78%    12%     5%       5%
Actual DNC           15%    80%     2%       3%
Actual Callback      20%     3%    72%       5%
Actual Hesitant      25%     2%     8%      65%

Cross-category errors: 22% average
Most confused: NotInterested ↔ DNC (13.5%)

Key Statistics From Voice Agent Intent Classification (2026):

• Intent boundary confusion: 15-25%
• Adjacent intent misclassification: 10-20%
• Over-escalation rate (soft → hard): 8-15%
• Under-escalation rate (hard → soft): 5-10%
• Impact on customer satisfaction: -25% for misclassified

Common Intent Boundary Confusions

Contributing Factors

• Overlapping intent definitions
• Missing temporal qualifiers (now vs. ever)
• No intensity gradient (soft vs. hard decline)
• Single-label classification for multi-intent
• Tone/prosody ignored in classification
• Insufficient negative examples in training

Eval Recipes

Test Cases

Metrics

Mitigation Strategies

Prevention

1. Mutually exclusive definitions: Ensure intents don’t overlap semantically
2. Hierarchical classification: First classify category, then sub-intent
3. Temporal markers: Explicitly check for “now”, “later”, “ever” qualifiers
4. Intensity gradient: Separate soft/hard variants of each intent
5. Multi-turn context: Use conversation history, not just last turn
6. Confidence thresholds: Route low-confidence to clarification

Intent Hierarchy

class IntentClassifier:
    """Hierarchical intent classification with boundary handling"""
    
    INTENT_HIERARCHY = {
        "negative": {
            "hard": ["do_not_contact", "wrong_number"],
            "soft": ["not_interested", "hesitant"],
            "temporal": ["busy", "callback"]
        },
        "positive": {
            "confirmed": ["qualified"],
            "partial": ["interested", "curious"]
        },
        "neutral": {
            "unable": ["unable_to_continue", "voicemail"]
        }
    }
    
    TEMPORAL_MARKERS = ["now", "right now", "at the moment", 
                        "currently", "abhi", "filhaal"]
    HARD_MARKERS = ["don't call", "don't contact", "remove", 
                    "stop calling", "mat karo"]
    
    def classify(self, transcript: str, context: list) -> dict:
        # First: check for hard markers (highest priority)
        if self.has_hard_markers(transcript):
            return {"intent": "do_not_contact", "confidence": 0.95}
        
        # Second: check temporal qualifiers
        has_temporal = self.has_temporal_markers(transcript)
        
        # Third: classify with context
        base_intent = self.llm_classify(transcript, context)
        
        # Adjust based on temporal
        if has_temporal and base_intent == "not_interested":
            return {"intent": "callback", "confidence": 0.8,
                    "note": "temporal_override"}
        
        return base_intent
    
    def has_hard_markers(self, text: str) -> bool:
        text_lower = text.lower()
        return any(marker in text_lower for marker in self.HARD_MARKERS)
    
    def has_temporal_markers(self, text: str) -> bool:
        text_lower = text.lower()
        return any(marker in text_lower for marker in self.TEMPORAL_MARKERS)
    
    def get_clarification_prompt(self, ambiguous_intent: str) -> str:
        """Generate clarification question for ambiguous cases"""
        clarifications = {
            "not_interested_or_callback": 
                "Just to confirm—would you prefer I call back later, "
                "or is this not something you're interested in?",
            "not_interested_or_dnc":
                "Got it. Should I also make sure we don't reach out again?"
        }
        return clarifications.get(ambiguous_intent, "")

Intent Definitions Best Practice

intents:
  do_not_contact:
    definition: "Caller explicitly requests removal from contact lists 
                 OR refuses any future WhatsApp/call follow-up"
    triggers:
      - "don't call me"
      - "remove my number"
      - "don't WhatsApp me"
      - "stop contacting"
    NOT: "Simple decline of current offer without contact restriction"
    
  not_interested:
    definition: "Caller declines the current offer but does NOT 
                 request removal from future contact"
    triggers:
      - "not interested"
      - "not for me"
      - "don't want this"
    NOT: "Busy/unavailable OR requesting no future contact"
    
  callback:
    definition: "Caller indicates they ARE interested but cannot 
                 talk NOW, with temporal qualifier"
    triggers:
      - "call me later"
      - "busy right now"
      - "in a meeting"
      - "not now"
    REQUIRES: Temporal qualifier (now, currently, right now)

Detection & Response

1. Intent-classification audit logging with boundary-accuracy tracking: For each call, log: {call_id, detected_intent (enum: soft_no|hard_no|interested|interested_but_later|not_qualified), intent_confidence_score (0-1.0), boundary_classification_correct (Y/N), escalation_triggered (Y/N), escalation_type (if yes: soft_to_hard|interested_to_action|unqualified_push), caller_satisfaction_indicator}. Daily audit: sample 10% of calls, manually verify intent classification against actual caller intent. Alert if: intent_accuracy <85%, or soft_to_hard escalation rate >5%, or hard_no misclassification rate >2%.

2. Escalation-compliance monitoring with temporal-qualifier validation: Track every escalation: when escalation happened, what temporal qualifier was present (if any). Verify: was temporal qualifier present before escalation? If no temporal qualifier, was escalation still triggered? Alert if agent escalates without temporal qualifier presence (potential over-escalation). Monthly report: “Escalations by intent type”, “Over-escalations (missing temporal qualifier)”, “Misclassifications (soft_no treated as interested)”.

Architecture Patterns

1. Intent Classifier with Boundary Enforcement: On caller response, classifier analyzes: (a) explicit keywords (“stop”, “no”, “interested”), (b) temporal qualifiers (“right now”, “later”, “eventually”), (c) intent score and confidence threshold. Maps to intent_enum: soft_no → callback eligible; hard_no → do not contact; interested → pitch; interested_but_later → callback; not_qualified → soft exit. Boundaries enforced: if soft_no detected, NO escalation unless temporal_qualifier present.

2. Escalation Gate with Temporal-Qualifier Validation: Before agent escalates from soft_no to hard_no or push, check: temporal_qualifier present? If absent, block escalation. If present, proceed with qualified escalation. All escalations logged with temporal_qualifier reference.

Key Metrics

Alerts & Escalation

References

• Intent Classification Survey - Multi-intent challenges
• Dialog State Tracking - Boundary handling
• Voice Agent Failures - Classification issues
• Conversational AI Research - Intent design