CX Decisions v2 — 20 CX decision types for 2026 practices
2,400 LLM-inferred (state, question, answer) rows in strands-decider Example format (kind/state/instructions/options/label/task/weight/instruction_variants), split evenly: 120 rows × 20 decision types, all states unique, all labels in range, all rows schema-validated.
Generation: four parallel subagents, each inventing realistic 2026 CX states (subscriptions, SSO, renewals, invoices, chat, fraud, SLA, PII) and judging the correct label by model inference. No scripts, no templates, no rule-based synthesis. Score labels spread evenly across levels (unlike v1's urgency rows, which were 99% one class).
The 20 decisions
| slice | task | kind | question |
|---|---|---|---|
| A routing/triage | cx_intent | choice/8 | route intent (billing, technical, account, order_status, product_info, cancellation, complaint, other) |
| A | cx_channel | choice/5 | best next channel (phone, email, chat, sms, self_service) |
| A | cx_category | choice/5 | case category (billing, technical, account, shipping, product) |
| A | cx_human | noul | needs a human agent vs AI can resolve |
| A | cx_selfserve | noul | eligible for self-service/KB deflection |
| B risk | cx_escalation | score/4 | escalation level |
| B | cx_churn | score/4 | churn risk |
| B | cx_complaint | noul | formal complaint vs inquiry |
| B | cx_approval | noul | needs supervisor approval |
| B | cx_fraud | noul | fraud / account-takeover signs |
| C customer state | cx_sentiment | score/5 | customer sentiment |
| C | cx_priority | score/4 | service priority |
| C | cx_csat | score/4 | satisfaction risk without fast action |
| C | cx_vip | noul | needs VIP handling |
| C | cx_callback | noul | needs a callback scheduled |
| D compliance/flow | cx_pii | noul | contains PII needing redaction |
| D | cx_translate | noul | needs translation |
| D | cx_repeat | noul | repeat contact, unresolved |
| D | cx_sla | noul | at risk of breaching response SLA |
| D | cx_nba | choice/5 | next best action (apologize_resolve, offer_refund, schedule_callback, escalate, provide_instructions) |
Intended use
Continual-train checkpoint for a strands-decider model (append to train_files; the 20 task names stratify held-out splits). Companions the cx_lam loop-gating rows (v1, 33,148 rows in the training repo) — together: 4 loop + 20 applied = 24 CX decision types.
Limits
LLM-judged labels: realistic but not human-adjudicated; spot-check before high-stakes use. English only. No PII in states (names are invented). Scores are ordinal judgments, not calibrated severities.