E-commerce
July 1, 2026
"Return" interpreted as "delivery". "Cancel" routed to "exchange". Same customer phrasing, three times, three bad bot responses. Without an improvement loop, misunderstandings repeat and fuel useless chatmis_ tickets.
Improving an AI chatbot from customer misunderstandings means collecting comprehension failures, prioritizing poorly routed intents, adjusting synonyms and confidence thresholds, and then measuring the decrease in repetition loops. Distinct from misunderstanding support (#879): here the product and bot ops team fixes the model, not the agent handling the ticket.
This guide #880 covers bot_misund_* intents, flow MISUNDb MUB-1 to MUB-8 and KPI misund_bot. Consumes CHATMIS-MAP #879 exported weekly.
Summary
Why a loop dedicated to chatbot misunderstandings?
The chatmis_ tickets signal a symptom. Without a product pipeline, the bot team fixes issues based on gut feeling or waits for a spike in complaints. A structured loop transforms each LOG-FOR-QA #879 into a prioritized backlog: incorrect intent, missing entity, confidence threshold too low, missing synonym.
Five targeted levers for improvement
Intent confirmation: validate the topic before giving an off-topic response
Entity synonyms: map customer phrasing to stable intents
Handoff threshold: break the loop after N comprehension failures
In-chat feedback: a "the bot did not understand" button feeds QA
Weekly review: top 10 chatmis_wrong_topic fixed during the sprint
DTC retail example
DTC fashion, 22 chatmis_wrong_topic cases/month. After MISUNDb: misund_bot_wrong_topic_rate -41%, chatmis_ tickets -29% in 6 weeks.
MISUNDb #880 vs CHATMIS #879, TRIAGE, HALLU #123 and products #109
Five comprehension quality contents, five distinct roles.
Quick Matrix
#880 MISUNDb: product loop to correct recurring misunderstood intents
CHATMIS #879: agents manage CM-5 misunderstanding tickets
TRIAGE: general support intent triage
HALLU #123: prevent false bot info distinct from bad intent
HANDOFF #12: transfer rules after bot failure
PROD #109: misunderstood products PDP analytics distinct from bot NLP
Pipeline: #879 collects and resolves → #880 fixes bot → decrease in chatmis_repeat_loop.
Promise #880
Policy MISUNDBOT-SUP, flow MUB-1 to MUB-8, 8 intents bot_misund_*, templates TPL-MISUNDb, KPI misund_bot_wrong_topic_rate.
Which bot_misund_* intents should be configured?
Eight intents reduce misunderstandings and feed the feedback loop.
Eight bot_misund intents
bot_misund_confirm_intent: confirm_copy validate topic before mapping response
bot_misund_clarify_low_conf: clarify_copy if confidence score is below threshold map
bot_misund_rephrase_hint: rephrase_copy guide short formulation map
bot_misund_loop_break: loop_break_copy exit repetition loop map
bot_misund_handoff_threshold: handoff_copy after N failed turns map
bot_misund_feedback_capture: feedback_copy misunderstanding report button map
bot_misund_synonym_resolve: resolution via synonym_map entities map
bot_misund_unknown_graceful: unknown_copy honest out-of-scope map
Each intent logs misunderstood_event for weekly review #880.
How to consume CHATMIS-MAP #879?
The bot and the product team read CHATMIS-MAP #879 exported: chatmis_wrong_topic, chatmis_repeat_loop, chatmis_not_understood, transcript, intent_detected, intent_expected.
Bot comprehension guardrails
CONFIRM-INTENT-BOT: confirm_copy if confidence is between seuil_bas (lower threshold) and seuil_haut (upper threshold)
LOOP-BREAK-BOT: loop_break_copy after 2 turns of the same incorrect intent
HANDOFF-THRESHOLD-BOT: handoff after 3 map comprehension failures
FEEDBACK-LOG-BOT: each feedback_capture → backlog #880
SYNONYM-GROUND-BOT: synonym_map return refund cancellation linked
NO-BLAME-USER-BOT: never attribute misunderstanding to the customer alone
HALLU-REROUTE-BOT: false info detected → #123 guardrails
MISUNDBOT-SUP policy in six rules
Six rules for a responsible improvement loop.
CONFIRM-INTENT-BOT: confirm subject if confidence is uncertain
LOOP-BREAK-BOT: propose reformulation or handoff before 4th identical turn
FEEDBACK-LOG-BOT: every customer report logged with transcript
WEEKLY-REVIEW-BOT: review top erroneous intents every week
SYNONYM-GROUND-BOT: synonyms documented by critical customer service intent
NO-BLAME-USER-BOT: neutral tone, never "reformulate better" alone
Flow MISUNDb MUB-1 to MUB-8
Eight-step flow: real-time comprehension then weekly continuous improvement.
MUB-1 Ingest message: NLP confidence score of extracted entities
MUB-2 Synonym check: SYNONYM-GROUND resolves ambiguous wording
MUB-3 Confidence gate: high → respond; medium → CONFIRM-INTENT; low → CLARIFY
MUB-4 Respond or clarify: response or confirm_copy clarify_copy
MUB-5 Loop watch: if repeat → LOOP-BREAK then HANDOFF-THRESHOLD
MUB-6 Feedback capture: feedback_copy if customer reports misunderstanding
MUB-7 Weekly export: CHATMIS-MAP aggregation → bot sprint backlog
MUB-8 Deploy fix: synonym threshold template → measure misund_bot KPI
Example TPL-MISUNDb-CONFIRM
" [confirm_copy map : return order {{no.}} ?] CONFIRM-INTENT-BOT. "
TPL-MISUNDb templates and touchpoints
Four short templates to reduce embed misunderstandings.
TPL-MISUNDb-CONFIRM
[confirm_copy map.] CONFIRM-INTENT-BOT. Wait for yes/no before business response.
TPL-MISUNDb-CLARIFY
[clarify_copy map.] [rephrase_copy map optional.] No off-topic reply until clarified.
TPL-MISUNDb-LOOP-BREAK
[loop_break_copy map.] [handoff_copy map if threshold reached.] LOOP-BREAK-BOT.
TPL-MISUNDb-FEEDBACK
[feedback_copy map.] FEEDBACK-LOG-BOT to backlog #880.
Touchpoints
Medium confidence: confirm before delivery/return response
2nd identical turn: loop_break + rephrase_hint
3rd failure: handoff #12 with transcript #155
Post-conversation: optional feedback_capture button
Edge cases and reroutes
Five cases out of the standard flow.
Mixed return + after-sales question: clarify split, not a forced intent
Bot gave incorrect info: HALLU #123 reroute not a synonym on its own
Customer language or slang: log to enrich synonym_map, handoff if blocked
Topic out of bot scope: unknown_graceful + immediate handoff #12
Sales peak new intents: accelerated weekly review, do not wait a month
Essential bot misunderstanding KPIs
Five MISUNDb monitoring metrics and correlation with tickets #879.
misund_bot_wrong_topic_rate: % wrong intent turns / total turns
misund_bot_loop_break_rate: % loops broken before handoff
misund_bot_confirm_save_rate: % confirm_intent avoiding wrong answer
misund_bot_feedback_captured_rate: reported cases logged / detected misunderstandings
misund_bot_chatmis_ticket_delta: evolution of chatmis_ tickets after fix deployment
Target: wrong_topic_rate decreasing and negative chatmis_ticket_delta after each synonym sprint.
MISUNDb anti-patterns
Five common errors in the misunderstanding loop.
Fixing without data #879: FEEDBACK-LOG and CHATMIS export mandatory
Forcing low-confidence response: CLARIFY or CONFIRM, do not hallucinate intent
Infinite loop: LOOP-BREAK and HANDOFF-THRESHOLD not configured
Blaming the user: NO-BLAME-USER, supportive rephrase_hint
Monthly review only: WEEKLY-REVIEW on critical CS intents
MISUNDb with Qstomy
Qstomy on Shopify: weekly CHATMIS-MAP export, misund_bot_* dashboard, editable synonym_map, configurable CONFIRM-INTENT and LOOP-BREAK, chatmis_ ticket correlation.
Scenario: retail DTC, return/delivery confused in 18% of turns. Synonym_map + confirm_intent. misund_bot_wrong_topic_rate -41%, chatmis_ tickets -29% in 6 weeks.
Explore AI support and request a demo.
Checklist, FAQ and going further
MISUNDb Checklist (8 steps)
Sync CHATMIS-MAP #879: weekly export of wrong_topic repeat_loop transcripts
Policy MISUNDBOT-SUP: 6 rules CONFIRM LOOP-BREAK FEEDBACK-LOG
8 intents bot_misund_*: flow MUB-1 to MUB-8
4 templates TPL-MISUNDb-*: CONFIRM CLARIFY LOOP-BREAK FEEDBACK
synonym_map Critical Customer Service: return refund cancellation delivery
Confidence thresholds: high medium low + HANDOFF-THRESHOLD 3 rounds
Red team loop: 2 rounds same mistake → loop_break handoff test
KPI Dashboard: misund_bot_* section 9 + delta chatmis_
FAQ
Difference with #879?
#879 = agents handling misunderstanding tickets. #880 = product fixing recurring bot issue.
Difference with #123?
#880 = wrong intent. #123 = factually incorrect answer.
How many synonyms to add?
Top 10 chatmis_wrong_topic wordings each sprint. Measure KPI delta.
Does Confirm intent slow down the bot?
Yes, but it avoids wrong answers and chatmis_ tickets. monitor confirm_save_rate.
Going further
This week: activate CONFIRM-INTENT on return/delivery, export top 10 chatmis_wrong_topic #879, deploy a test synonym_map, measure misund_bot_wrong_topic_rate.
Enzo
July 1, 2026
