Bad Clarification Behavior
Agent asks unnecessary questions or misses critical clarifications.
Customer Service
15 patterns in this category
Customer-service failures happen across three distinct workflows: reactive support handling inbound customer requests (conversation resolution), proactive outreach reaching out to at-risk customers (retention outreach), and specialized financial-risk handling (refund and billing disputes). The shared underlying failure modes are the same conversation-quality problems documented in general agent-interaction — clarification miscalibration, state tracking, tone — but customer-service adds domain-specific constraints: support agents select canned responses by account state (not just topic), escalate by attempt count (not just confidence), and handle financial claims that carry monetary risk if wrong. A support-specific failure is when the agent’s general conversation is fine but the canned response was for the wrong account tier, the escalation threshold was calibrated to bot-solvable requests, or a prior credit was issued but not carried through a handoff.
Key Takeaways
- 15 patterns are documented across 3 goals: Conversation Resolution (12 patterns), Proactive Retention Outreach (1 pattern), Refund and Billing Disputes (2 patterns).
- Conversation Resolution is the largest goal (12 patterns) because it covers general conversation-quality failures applied to support (clarification, escalation, tone) plus support-specific failures (canned-response retrieval mismatch, multi-agent handoff context loss).
- Embedding-retrieval-selects-similar-but-wrong-canned-response shows that canned-response libraries contain responses for similar-but-distinct scenarios (different tiers, different bugs), and retrieval by text similarity alone can rank the wrong response first if the library contains topically-overlapping templates.
- Multi-agent-handoff-drops-de-escalation-context and multi-agent-handoff-drops-partial-credit-already-issued together show that support handoffs between triage and specialized agents lose critical context (frustration level, prior credits issued) because handoff schemas lack fields for conditional information.
Customer-Service Goals
| Goal | Covers | Patterns |
|---|---|---|
| Conversation Resolution | Canned-response selection, clarification behavior, escalation timing, tone and emotional state, multi-agent handoff quality | 12 |
| Proactive Retention Outreach | Fabricated personalization when usage-analytics data is unavailable, overriding the fallback-to-generic-messaging policy | 1 |
| Refund and Billing Disputes | Retrieval-mismatched billing-dispute templates, dropped prior-credit information at handoffs | 2 |
Total: 15 patterns
How the Goals Relate
The three goals represent different support workflows: Conversation Resolution handles inbound customer-initiated requests, Proactive Retention Outreach reaches out to at-risk customers before they churn, and Refund and Billing Disputes is a specialized financial-risk domain applying resolution principles to billing scenarios. Conversation Resolution and Refund/Billing Disputes overlap on shared customer-service mechanics (escalation, state tracking, information gathering) but Refund/Billing adds financial verification (ledger queries, scope-matching of dispute templates). Proactive Retention Outreach is distinct in that the agent initiates contact with no prior customer message, so credibility is not yet built — a fabricated justification damages trust on first touch, whereas the same fabrication in reactive support is at least responsive to customer context. To localize an incident by symptom: agent asks repeatedly for information the customer already gave → Conversation Resolution; customer is routed to a specialist but has to re-explain everything → Conversation Resolution handoff failure; a churn-risk account is contacted with inaccurate usage metrics → Proactive Retention Outreach; customer disputes a $60 charge, receives a full refund despite already having $15 credit issued → Refund/Billing Disputes handoff failure.
Frequently Asked Questions
How are conversation-resolution support failures different from general conversation-quality failures?
Support-specific failures add domain constraints: canned-response libraries (selection must match account state), escalation thresholds (must be calibrated per request category, not uniformly conservative), and financial claims (carry real money at stake). Embedding retrieval selects similar-but-wrong-canned-response and multi-agent handoff drops de-escalation context are support-specific because they involve structured handoff schemas and account-state matching, not just conversation quality.
Can proactive retention outreach use a generic message instead of personalization?
Yes. Fabricated usage-decline justification argues for exactly that when analytics data is unavailable. Generic outreach (“we value your business”) loses personalization but retains credibility; fabricated personalization gains apparent specificity but damages trust if the customer knows it’s inaccurate.
What is the highest-severity customer-service failure?
Multi-agent handoff drops partial-credit-already-issued creates direct financial loss (duplicate refunds, overpayments). Conversation Resolution failures affect customer trust and escalation volume; Proactive Retention failures affect churn. Refund/Billing failures have immediate monetary impact.
Are all customer-service failures about agent capability, or are some about configuration?
Most are about configuration and architecture: canned-response retrieval needs account-state filtering (configuration), escalation thresholds need calibration per request type (configuration), handoff schemas need conditional fields (architecture). A support agent with good conversation quality can fail at all three if configuration is wrong.
Related Categories
- Conversation Quality — general conversation-quality failures (clarification, state tracking, tone) that appear in customer-service contexts but apply across all agent types
- Knowledge Retrieval — retrieval-augmented generation failures that affect support agents when canned-response or knowledge-base selection is involved
Bad Refusal
Agent refuses safe requests or gives unsafe help.
Document-Level Retrieval Mismatch Pulls Wrong Billing-Dispute Template
A Billing-Dispute Agent That Retrieves a Dispute-Handling Template or Policy Article From a Knowledge Base Via Embedding Similarity to the Customer's Complaint Wording Pulls an Entire Document That Is Topically and Lexically Close -- Covering a Structurally Similar but Different Dispute Type, Product Tier, or Region -- Rather Than the Document That Actually Governs the Customer's Account, and Applies That Wrong Document's Resolution Steps and Dollar Thresholds Confidently
Embedding Retrieval Selects Similar but Wrong Canned Response
A Support Agent That Selects a Canned Response or Macro From a Library Via Semantic/Embedding Similarity to the Customer's Message, Rather Than by Matching the Customer's Actual Account State or Issue Category, Retrieves a Response That Is Lexically and Topically Close to What the Customer Wrote but Answers a Different Underlying Situation -- Sending Confident, On-Topic-Sounding Guidance That Does Not Actually Apply to the Customer's Case
Escalation Too Early
Agent gives up despite solvable request.
Fabricated Usage-Decline Justification When Analytics Tool Returns Empty
A Proactive Retention Agent That Calls a Usage-Analytics Tool to Determine Why a Customer Has Been Flagged as At-Risk of Churning Receives an Empty or Partial Result -- Because the Customer's Product Tier Is Not Instrumented for That Metric, or the Analytics Service Timed Out -- and Composes a Specific, Plausible-Sounding Usage-Decline Narrative ("You Haven't Logged In Since Early Last Month and Your Team's Usage Dropped 60%") to Justify the Outreach, Rather Than Stating the Actual Usage Data Was Unavailable
False Completion Claim
Agent says done when action was not completed.
Inconsistent Answers
Agent gives different answers across turns.
Multi-Agent Handoff Drops De-Escalation Context Between Triage and Billing Agent
A Triage Agent That Determines a Customer Is Already Frustrated and Has Explicitly Requested Not to Repeat Their Account Details Again Records That Context Only in Its Own Conversational Reasoning, and When the Conversation Is Routed to a Downstream Specialized Billing Agent That Operates on a Structured Ticket Object Containing Only the Stated Issue Category, the De-Escalation Context and Already-Provided Details Never Cross the Handoff Boundary, So the Billing Agent Re-Opens the Conversation by Asking the Customer to Re-Authenticate and Re-Explain Everything From Scratch
Multi-Agent Handoff Drops Partial-Credit-Already-Issued Flag Between Triage and Billing Agent
A Triage Agent That Learns During Intake That a Customer Has Already Been Issued a Partial Credit for a Disputed Charge by a Prior Agent Records That Fact Only in Its Own Conversational Summary, and When the Conversation Is Routed to a Downstream Specialized Billing-Dispute Agent That Operates on a Structured Dispute-Case Object Containing Only the Disputed Amount and Category, the Already-Issued Partial Credit Never Crosses the Handoff Boundary -- So the Billing Agent Calculates and Approves a Second, Full-Amount Refund on Top of the Credit the Customer Already Received
Over-Apology Loop
Agent apologizes repeatedly without solving.
Poor Escalation
Agent fails to hand off to a human at the right time.
Repeated Question Loop
Agent asks for information already provided.
Tone Mismatch
Agent sounds rude, robotic, too casual, or too formal.
Unclear Next Step
User does not know what happens next.