E-commerce
June 28, 2026
Monday: the help hub indicates returns within 30 days. Tuesday: the agent writes 14 days for a loyal customer. Wednesday: the bot promises a refund within 48 hours. Thursday: the customer opens a dispute. Three channels, three truths.
Intercom points out that an AI agent is only reliable if the terminology, policies, and tone remain identical across every touchpoint (Intercom, knowledge management 2026). Alhena estimates that three contradictory articles on the same policy drop bot accuracy by 15 to 25 points (Alhena, agent training 2026).
This guide #191 covers support consistency governance: avoiding contradictory answers between the Q&A hub, chatbot, and agents. Distinct from answers database (#102) (building the library) and tool comparison (#180) (choosing hub vs. bot).
Summary
Why is conflicting support expensive?
Contradictory e-commerce support responses destroy trust, CSAT, and bot efficiency in a single conversation.
Visible Symptoms
Reopened ticket: customer quotes previous agent vs. current response
"Your site says something else": help center vs. chat
Bot contradicts agent: human handoff required, low NPS
Slow agent onboarding: everyone makes up their own wording
Measurable Cost
ScreenMeet cites a target ticket reopen rate below 5%; above 10% is a systemic quality signal. Converged Hub notes that omnichannel without a single source of truth multiplies handle time and escalations (Converged Hub, omnichannel 2026).
DTC Example
Fashion brand: help center "30-day returns", agent macro "45-day VIP", bot "14 days" (hallucination). Return tickets +38%, CSAT -1.4 pt, bot unmatched return intent x2.
Domino Effect
A contradiction regarding the refund period often generates a chargeback, a 1-star review, and a reopened ticket in the same week. The customer does not distinguish between the help center, bot, and agent: to them, it is the brand that is lying.
How does it differ from the answer base and the tool comparison?
#191 covers governance and sync, not the initial content creation.
#102 REP responses base
Responses base (#102): REP sheets structure, taxonomy, macros. #191: maintaining a single truth between REP, public hub, and bot.
#180 Hub vs search vs bot
Comparison #180: which tool for which intent. Here: all coexist, how to align them.
#165 Structure by intent
Organize by intent (#165): product section architecture. #191: ops workflow when policy changes.
#103 Bot data cleaning
Clean data (#103): preparing AI corpus. #191: continuous post-launch governance, not a one-shot audit.
Promise #191
Surfaces mapping, canonical source, roles, weekly audit, contradiction KPI.
Which response surfaces should be mapped?
List each customer touchpoint / response surface before talking about consistency.
Self-service surfaces
Public help center: policy, shipping, returns articles
Product PDP sections: Q&A by SKU
Chatbot widget: RAG + system instructions
Internal site search: help center + blog index
Agent surfaces
Internal KB: Notion, Guru, Qstomy KB
Gorgias/Zendesk Macros: derived shortcuts
Agent onboarding notes: Unofficial Google Doc = danger
Passive surfaces
Site banners, footer policy, transactional Klaviyo emails, checkout reassurance page. A change from 30 to 14 days return policy must affect all of these surfaces, not just the help center.
How do you define a single source of truth?
The canonical support source is the master document from which all other formulations are derived.
Recommended Hierarchy
Official Policy: Shopify legal page or legally validated PDF
Canonical REP sheet: operational agent + bot translation
Public hub article: simplified customer version
Macro / bot snippet: short derived excerpt
Golden Rule
Never modify the macro first. Flow: policy → REP → hub → macro → bot re-indexing. Gorgias reminds us that macros are no longer used as a bot source since 2026: useful content must migrate to guidance or articles (Gorgias, AI Agent 2026).
Mandatory Canonical Sheet Fields
REP ID, linked policy, short chat response, long email response, allowed exceptions, last review date, owner, published/deprecated status. See REP structure (#102).
What governance roles and rituals should be put in place?
Consistent support governance requires an owner and a ritual, not just a shared Notion.
Typical DTC RACI
KB Owner (support lead): validates any REP modification
E-commerce Ops: syncs PDP, banners, checkout
Marketing: policy-aligned Klaviyo emails
Dev / integrator: re-index bot within 24 hours
Policy change ritual (example: return 30 → 45 days)
Notion ticket « CHANGE-RET-2026-03 » opened
REP-RET-001 update + owner validation
Help hub + macros + footer banner
Bot re-index + 10 gold questions test
Slack agents brief + archive of the old version
Ticket closure with effective date
Forbidden
Senior agent modifies macro without updating REP. Intern answers from memory regarding an undocumented exception.
How to audit contradictions every week?
The weekly support consistency audit takes 45 min with a fixed protocol.
10 gold questions protocol
Select 10 top-volume intents: delivery time, return, refund, cancellation, promo, size, WISMO, address, warranty, contact. For each intent, compare: help hub, bot, agent macro, random agent response (sample of 5 tickets).
Scoring grid
Aligned: same fact, similar phrasing
Minor discrepancy: different tone, identical fact
Contradiction: different delay, amount, or policy
Alert threshold
≥ 1 contradiction out of 10 = action within 48 hours. Alhena recommends removing contradictory duplicates before any bot retraining: three versions of a policy = tanked accuracy.
Source of frequent discrepancies
Temporary promo not removed from the bot. VIP exception in macro only. 2024 help hub never refreshed. See support post-redesign for Day 0 corpus transition.
Simple tool
Audit spreadsheet: columns for intent, hub, bot, macro, agent, score, owner fix, correction date. Shared on Friday during the 15-minute support standup.
How do I synchronize the chatbot with the canonical source?
The coherent bot cites the REP sheet, not a mix of obsolete articles.
Sync pipeline
Export published REP only (not draft)
Atomic Q&A chunking: 1 chunk = 1 fact
Metadata: intent, region, review date
Re-index + deletion of deprecated chunks
Post-index gold set test
System instructions
"If conflict between sources, prioritize the most recent REP-ID sheet. If uncertain, human handoff. Never invent refund delay." See bot instructions and anti-hallucinations.
Versioning
Alhena recommends a corpus snapshot by date for auditability: knowing what the bot knew during a disputed conversation.
How do you train agents on the canned response?
Support agents must search for the REP sheet before improvising.
Team rules
Mandatory macro for top 20 intents before customization
Documented exception: ticket note + validation request if outside of REP
Promises forbidden: lead time or commercial gesture not listed in the REP
30-minute onboarding
REP Tour: where to search, how to read escalation, example of a past contradiction and its correction. 5-question quiz: "Customer asks for return at D+40, what to reply?"
Feedback loop
Agent flags "REP-RET-001 obsolete" via the kb_gap tag. Owner processes within 5 business days. Reopened ticket due to contradiction = P2 priority REP review.
How do you handle edge cases without breaking consistency?
Support exceptions exist, but must remain traceable and rare.
Exception matrix
VIP / High LTV: pre-approved gesture in REP-ESC-VIP, not made up
Ops incident: temporary FLASH macro with end date
Warehouse error: documented immediate refund, no generic bot promise
Temporary macros
TEMP-BFCM- prefix, mandatory expiration date. Auto-removal at D+7 post-promo. Otherwise the bot and the hub will still quote the promo in March.
Multi-channel
Shopify DTC policy u2260 Amazon marketplace. Mandatory channel tag before responding. See marketplace support (#190).
Which KPIs measure the consistency of responses?
Measure the consistency and quality of support, not just ticket volume.
Leading KPIs
Ticket reopen rate: target < 5%
Gold 10 audit score: 10/10 aligned
Bot-human agreement: % of bot responses validated by agents on a sample
Open KB gaps: downward trend
Lagging KPIs
CSAT, verbatims "I was told something else", unbroken promise chargebacks, new agent onboarding time.
Monthly Dashboard
1 page: contradictions detected, average correction time, top 3 conflicting intents, ongoing policy actions. See maintain knowledge base.
How does Qstomy guarantee multi-source consistency?
Qstomy centralizes the REP corpus, bot, and human escalation on a single versioned source.
Features
Single corpus: hub, bot, agent copilot, same KB
Snapshot versioning: conversation audit, dispute
Conflict alert: detects gap between macro vs REP
Gold set test: auto post-update
Tag kb_gap: escalation from agent → owner
Encrypted DTC Scenario
Cosmetic brand, help hub with 62 articles, 48 Gorgias macros, RAG bot on obsolete help center export. Audit: 7/10 contradictions in return/delivery. Deployment of Qstomy REP corpus + 24h bot sync + weekly gold audit. After 8 weeks: contradictions 0/10, reopen rate 4.1% (vs 11%), CSAT +0.9 pt, bot resolution +18 pts, agent onboarding −2 days.
Explore AI customer support, Shopify, request a demo.
Which operational playbooks should be launched this week?
Playbook 1: Surface Mapping (2 h)
List hub, bot, macros, banners, emails. Owner per surface. Identify surface without owner = risk.
Playbook 2: 10 Canonical REP Sheets (1 day)
Top 10 volume intents. Section 4 structure. Status: published. Link to official policy.
Playbook 3: First Gold Audit (45 min)
10 questions, section 6 scoring. 48-hour correction plan if there is a contradiction.
Playbook 4: Bot Sync (2 h)
Re-index published REP. Gold test. Anti-conflict instructions.
Playbook 5: Friday KB Ritual (30 min)
kb_gap tags, modified macros without REP, 1 canonical update.
Useful Linking
A customer does not see your internal silos. They see a brand that contradicts itself, or a reliable brand. Consistency is governed, not wished for.
Enzo
June 28, 2026
