E-commerce
July 1, 2026
"Do you train on my messages?" "Does my order history feed the bot?" "Does OpenAI keep our conversations?" Three questions where chatbot training data lack a clear support response.
The e-commerce training data support questions covers KB sources, catalog, conversations, and exclusions, without improvising or denying the actual usage. Distinct from AI response transparency (#860): here the focus is on what trains the model, not just the status of a chat response.
This guide #907 deploys policy TRAINDATA-SUP, flow TD-1 to TD-8, and matrix TRAINDATA-MAP. Paired with the future bot training data (#908).
Summary
Why does training data generate tickets?
Wary customer after a poor bot response. Agent says "we don't use anything" even though the Shopify KB powers the RAG. Without TRAINDATA-MAP, there is confusion with disclosure #860 or privacy #154.
Five typical training data frictions
Vague sources: customer does not know which data
Conversations used: past chats for learning?
AI Provider: OpenAI Anthropic vendor retention
Opt-out: can we opt out of data contribution
Catalog orders: purchase data in the bot
DTC retail example
DTC Fashion, 7 traindata_ tickets/month. After TRAINDATA-MAP: traindata_trust_resolution_rate 90%, contradictory responses -58%.
TRAINDATA #907 vs AITRAN #860, AIGENCON #859, Governance #142 and bot #908
Seven AI transparency contents, seven distinct angles.
Quick Matrix
#907 TRAINDATA: training data sources exclusions conversations catalog
AITRAN #860: disclose AI status chat response distinct training
AIGENCON #859: shop content generated by AI distinct bot training
Governance #142: internal validation rules supervision
Privacy #154: cross-functional data usage
CONVREC #899: conversation recording distinct training
Opt-out #909: exclusion of conversations from training distinct sources
Bot #908: explain sources limits exclusions widget
#860 = is this response AI? #907 = what was the bot trained on?
Promise #907
TRAINDATA-SUP policy, TRAINDATA-GATE tree, 8 macros, register of sources and exclusions, KPI traindata_trust_resolution_rate.
Which typologies of traindata_* to classify?
Action-oriented classifier: sources ≠ conversations ≠ opt_out ≠ vendor.
Eight TRAINDATA-MAP typologies
traindata_sources_ask: what data trains the bot
traindata_conversations_used: customer chats used for training
traindata_catalog_orders: ordered products in sources
traindata_vendor_question: AI provider OpenAI data retention
traindata_exclusions_ask: what is not used
traindata_opt_out_intent: refuse data contribution
traindata_human_review: human supervision of training
traindata_escalate_dpo: formal GDPR request
Policy TRAINDATA-SUP: Agent Rules and Source Registry
The TRAINDATA-SUP policy sets answers from the training registry, not technical improvisation.
Six TRAINDATA-SUP rules
REGISTRY-FIRST: TRAINDATA macro from sources registry
Explicit exclusions: TRAINDATA-EXCLUDE what is not used
Do not deny actual usage: KB catalog honesty if applicable
Vendor copy: TRAINDATA-VENDOR documented provider DPA
Opt-out → #909: dedicated traindata_opt_out handoff opt-out
Formal GDPR → DPO: traindata_escalate_dpo
Typical documented sources
KB help center: FAQ policy articles indexed by RAG
Shopify Catalog: product sheets for product answers
CS Macros: validated agent templates
Exclusions: customer conversations unless documented opt-in
Flow TD-1 to TD-8: standard resolution
Eight sequential steps, SLA P3 traindata < 24 h, escalate DPO if escalate_dpo.
Flow TD-1 to TD-8
TD-1 Triage: training data vs disclosure #860 vs privacy #154?
TD-2 Registry lookup: sources exclusions vendor registry
TD-3 Classify: traindata_* via TRAINDATA-MAP
TD-4 Inform: SOURCES EXCLUDE VENDOR macros
TD-5 Route: opt_out #909 disclose #860 DPO
TD-6 Confirm: macro TRAINDATA-DONE
TD-7 Log: registry version used
TD-8 Close: KPI traindata_trust_resolution_rate + brief #908
Eight ready-to-paste TRAINDATA-* macros
Macros aligned registry sources exclusions vendor.
TRAINDATA-* Library
TRAINDATA-ACKNOWLEDGE: "Thank you for your question about the chatbot data."
TRAINDATA-SOURCES: "Sources: {{sources_list}}. No training on out-of-scope data."
TRAINDATA-EXCLUDE: "Not used: {{exclusions}} unless explicit opt-in."
TRAINDATA-CONVERSATIONS: "Conversations: {{statut_conv}} according to policy {{policy_link}}."
TRAINDATA-VENDOR: "AI Vendor: {{vendor}}. DPA and retention: {{vendor_copy}}."
TRAINDATA-OPT-OUT-ROUTE: "To opt out of contribution: {{opt_out_process}}."
TRAINDATA-HUMAN-REVIEW: "Supervision: {{review_process}} before production release."
TRAINDATA-DONE: "Summary: {{question}}. Answer: {{résolution}}. Reference: {{id}}."
TRAINDATA-GATE tree and agent-ready registry
Decision tree before improvising or denying catalog usage.
TRAINDATA-GATE
Training sources question? → SOURCES + EXCLUDE registry
Conversations used? → CONVERSATIONS + opt_out route #909
OpenAI vendor? → VENDOR copy DPA
AI-only response status? → handoff AITRAN #860
AI product sheet content? → handoff AIGENCON #859
Minimum internal registry
Document: indexed sources, conversation exclusions, vendor and DPA, opt-out process, human review pipeline. Train agents: traindata ≠ aitran #860 ≠ convrec #899.
KPI, QA and handoff to bot #908
Measuring TRAINDATA detects contradictory responses and obsolete registry.
Four TRAINDATA KPIs
traindata_trust_resolution_rate: tier 1 resolved questions / total traindata_
traindata_registry_compliance_rate: % responses aligned with registry
traindata_contradiction_rate: contradictory responses between agents low target
traindata_opt_out_route_rate: % opt_out routed #909 correctly
Handoff #908
Export TRAINDATA-MAP to bot: traindata_sources_ask traindata_exclusions_ask high priority. Guardrail TRAINDATA-REGISTRY-GATE brief #908 sources_copy widget.
Edge cases: fine-tuning, vendor change, minor
Three cases outside the standard flow.
Fine-tuning conversations
If fine-tune is active: honest CONVERSATIONS + opt-out #909. Do not deny if the registry indicates it.
AI vendor change
VENDOR macro updated registry. Do not promise the old vendor.
Minor's data
ESCALATE-DPO. EXCLUDE conversations of minors if policy requires it.
Agent training: 20 minutes TRAINDATA
Module: REGISTRY-FIRST, SOURCES EXCLUDE VENDOR, routing #860 #909 #908.
Exercises
Ticket A: "train on my chats?" → CONVERSATIONS + EXCLUDE
Ticket B: "is it an AI?" alone → handoff AITRAN #860
Ticket C: learning opt-out → OPT-OUT-ROUTE #909
How Qstomy structures TRAINDATA in your stack
Qstomy route traindata_*, sync training registry macros, handoff #908 tier 1 and #909 opt-out.
Three bricks
Routing: intent training_data vs ai_disclose vs privacy
Training registry: sources exclusions vendor review_process
Bot #908: sources limits exclusions widget
Scenario: DTC, 6 tickets/month traindata. REGISTRY-FIRST agents, bot #908 tier 1. traindata_trust_resolution_rate goes from 72% to 91% in 4 weeks.
FAQ and TRAINDATA deployment checklist
FAQ
Is saying "we do not use any data" false?
No. REGISTRY-FIRST. KB catalog = source if documented.
Difference #860?
#907 = training data. #860 = is this response AI-generated.
Difference #909?
#907 informs sources. #909 executes learning conversations opt-out.
Difference #908?
#907 = agents. #908 = bot explaining sources limits exclusions.
7-day Checklist
D1: TRAINDATA-SUP + TRAINDATA-MAP + vendor opt-out sources registry
D2: 8 helpdesk macros
D3: routing matrix #860 #859 #909
D4: 20 min agent training REGISTRY
D5: tags traindata_* + KPI
D6: test SOURCES vs handoff #860 vs #909
D7: bot brief #908 REGISTRY-GATE
Interlinking
Enzo
July 1, 2026
