Multi-Agent Handoff Drops Prior Attempted Fix Between Bot and Human Agent

Issue: A Bot-Tier Support Agent’s Chat Transcript With a Customer Shows the Customer Already Tried Two Specific Troubleshooting Steps Without Success, but the Structured Ticket-Summary Payload the Bot Hands Off to the Human-Agent Queue Does Not Carry That Field, So the Human Agent Re-Suggests One of the Same Already-Tried Fixes

Frequency: Common

Symptoms

• The human agent’s first response to an escalated ticket suggests a troubleshooting step that the bot-tier transcript, for the same conversation, shows the customer already tried and reported as unsuccessful
• The structured handoff summary the human agent sees contains the customer’s stated problem and category but no field listing steps already attempted, even though the full bot transcript (often not read in full by the time-pressured human agent) contains that information
• Customer explicitly states “I already told the bot I tried that” in response to the human agent’s repeated suggestion, a verbatim or near-verbatim complaint pattern that recurs across escalated tickets
• Re-handoff with an explicit “steps already attempted” structured field populated from the bot transcript eliminates the repeated suggestion, isolating the failure to the handoff contract rather than the human agent’s competence
• Customer satisfaction scores for escalated tickets are measurably lower when the bot-to-human handoff repeats an already-tried step, compared to escalations where the handoff correctly carries that context forward

Root Cause The bot-tier agent and the human-agent queue interface communicate through a structured ticket-summary payload, not through the human agent reading the full bot transcript before responding. When the bot’s output contract allows “steps already attempted” to exist only within the free-text transcript without being extracted into a structured field the handoff summary surfaces, that information is technically available but practically invisible to a human agent working through a queue under time pressure, who reasonably relies on the structured summary rather than re-reading an entire prior bot conversation for every escalated ticket.

Example

Customer chats with the bot-tier support agent about a sync failure, and the transcript shows: "I already tried restarting the app and clearing the cache, neither worked"
Bot escalates the ticket to a human agent; the structured handoff summary includes the customer's stated problem category ("sync failure") but no "steps already attempted" field
Human agent, working from the structured summary without reading the full transcript, opens with: "Have you tried clearing the app cache?"
Customer responds: "I already told your bot I tried that," registering frustration at having to repeat information already provided
Resolution is delayed by the redundant exchange, and the customer's satisfaction rating for the interaction reflects the repeated-information frustration rather than the underlying issue alone

Key Statistics

Contributing Factors

• Bot agent’s output contract permits “steps already attempted” to exist only in the free-text transcript without requiring extraction into a structured handoff field
• Human-agent queue interface surfaces the structured handoff summary as the primary working context, with full bot-transcript review treated as optional given time pressure in a busy queue
• No automated reconciliation compares troubleshooting-step language in the bot transcript against the structured handoff summary’s fields before the ticket reaches a human agent

Mitigation Strategies

1. Mandatory Structured “Steps Attempted” Field: Require the bot agent to populate a structured steps_attempted field listing every troubleshooting step the customer reports having tried, extracted from the transcript at escalation time, and block the handoff from completing without it
2. Transcript-to-Field Reconciliation Check: Run an automated scan of the bot transcript for attempted-step language and verify a corresponding structured field entry exists before the ticket reaches the human-agent queue, flagging any mismatch
3. Human-Agent Interface Surfaces Attempted Steps Prominently: Display the structured steps_attempted field prominently in the human-agent’s working view, positioned so it is seen before any troubleshooting suggestion is composed, rather than buried in an optional transcript link
4. Suggestion-Repeat Detection: Before a human agent’s suggested step is sent, automatically check it against the structured steps_attempted field and warn the agent if the suggestion duplicates an already-tried step

Metrics

• Rate of escalated tickets where the human agent’s first suggestion duplicates a step listed in the bot transcript as already attempted
• Percentage of bot-to-human handoffs with a populated structured steps_attempted field versus transcript-only
• Customer satisfaction score for escalated tickets, segmented by whether the handoff repeated an already-tried step

Alerts

• A human agent’s response duplicates a step the structured steps_attempted field (or an unreconciled transcript) shows the customer already tried → P2
• Reconciliation check finds attempted-step language in the bot transcript with no corresponding structured field, and the ticket reaches a human agent before resolution → P2
• Suggestion-repeat rate across escalated tickets exceeds baseline for two consecutive reporting periods → P3

References

• Demystifying the Lifecycle of Failures in Platform-Orchestrated Agentic Workflows
• Why Do Multi-Agent LLM Systems Fail? (MAST)
• CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions