Mid-Conversation Reclassification Reroutes Actively-Owned Ticket

Issue: A Routing Agent That Re-Runs Full-Context Intent Classification on Every New Customer Message Silently Reassigns a Ticket Already Claimed and Being Worked by a Human Agent, Because Growing Conversation Context Shifts the Classification Output Without Any Check on the Ticket’s Current Ownership State

Frequency: Occasional

Symptoms

• A ticket already assigned to and actively being worked by a human agent is autonomously reassigned to a different team mid-conversation, after the customer adds a new message that shifts the topical balance of the conversation
• The reassignment happens with no check of the ticket’s current status field (claimed, in-progress, assigned-to) before the reroute action executes
• Reclassification-triggered reassignments cluster on longer conversations, where later messages introduce enough new topic-relevant tokens to outweigh the original classification’s supporting content in the full-context re-run
• The human agent who was actively working the ticket loses access or context ownership without notification, and the newly-assigned team receives a ticket with no awareness that meaningful work (diagnosis, partial fix) had already happened
• Re-running the identical conversation transcript through the classifier in a single pass (rather than incrementally, message by message) produces a different, and often more stable, classification than the turn-by-turn re-classification did

Root Cause The routing agent is architected to re-invoke intent classification on the full conversation context after every new customer message, rather than maintaining a sticky classification that only changes on an explicit, gated re-evaluation trigger. Because LLM classification outputs are sensitive to the token-level composition of their input, each new message shifts the effective context the classifier sees, and a conversation that starts on-topic for Team A but later includes phrasing more associated with Team B’s domain (a tangential mention, a follow-up question) can flip the classification even though the ticket’s actual, primary issue has not changed. With no ownership-state check gating the reassignment action, every classification flip is treated as a fresh routing decision rather than a signal to be reconciled against the fact that a human is already mid-resolution.

Example

Turn 1: Customer: "My invoice shows a charge I don't recognize"
Classification: Billing -- routed to Billing team, claimed by agent Maria, who begins investigating
Turn 4 (after some back-and-forth): Customer: "Also, when I try to log into the account to check
  past invoices, I get a 'session expired' error every time"
Full-context re-classification (triggered by the new message): Now weighs "session expired,"
  "log into," "error" heavily -- reclassifies as Technical Support
Reroute action: Ticket autonomously reassigned to the Technical Support queue
No ownership check: Maria's active investigation status is not checked before reassignment
Impact: Maria's partial billing investigation is lost from the new team's context; ticket
  restarts from scratch in Technical Support, which cannot resolve the original billing charge

Key Statistics

Eval Recipes

Test Cases

Evaluation Dataset

• Source: Historical multi-turn support conversations with known human-assigned ownership timestamps and final correct team assignment
• Size: 200+ conversations, weighted toward longer threads (5+ turns) where topic drift is more likely
• Key variations: claimed vs. unclaimed ticket at time of new message; genuine topic change vs. tangential mention; single primary issue vs. genuinely multi-issue conversations

Metrics

Automated Checks

def check_unowned_reroute(trace: list[dict], ticket_state: dict) -> dict:
    """Flag a reroute action taken on a ticket that was claimed and in-progress."""
    reroute_calls = [c for c in trace if c["tool"] == "reassign_ticket"]
    flagged = [
        c for c in reroute_calls
        if ticket_state.get("status") == "in_progress"
        and not c.get("args", {}).get("ownership_check_passed", False)
    ]
    return {
        "unowned_reroute_count": len(flagged),
        "risk": "silent_reassignment_of_claimed_ticket" if flagged else None,
    }

Mitigation Strategies

Prevention

1. Ownership-State Gate on Reroute: Require the routing agent to check the ticket’s current status (claimed, in-progress) before executing a reassignment, and route any classification flip on a claimed ticket to a human-reviewed escalation instead of an automatic reroute
2. Sticky Classification with Explicit Re-Evaluation Trigger: Maintain the ticket’s classification as sticky once assigned, only re-evaluating on an explicit trigger (customer states “actually, this is a different issue”) rather than re-running full classification on every new message
3. Multi-Issue Detection and Split, Not Silent Reroute: When a new message introduces a genuinely separate topic, offer a ticket-split action (spinning off a second ticket) rather than reclassifying and rerouting the original, preserving the in-progress work

Detection & Response

1. Claimed-Ticket Reroute Scanning: Automatically flag any reassignment action executed against a ticket with an active human owner at the time of reassignment
2. Classification-Flip Frequency Monitoring: Track how often a given ticket’s classification changes across its lifetime; tickets with 2+ flips are a strong signal of drift-driven misrouting rather than genuine topic evolution

Architecture Patterns

• Ownership-Aware Routing Gate: Insert an explicit ownership-check step between classification output and reroute execution, structurally preventing a classification result alone from triggering reassignment
• Incremental-Delta Classification: Classify based on what changed in the new message relative to the established classification, rather than re-deriving intent from the full accumulated context each time, reducing sensitivity to later-turn token composition

Production Signals

Key Metrics

Alerts

References

• Agent Drift: Quantifying Behavioral Degradation in Multi-Agent LLM Systems Over Extended Interactions
• Token Statistics Reveal Conversational Drift in Multi-turn LLM Interaction