cleanexpo cs-tier1
Bottles 15+ years of CS leadership — de-escalating enterprise churn calls, the "Mom test" for feedback (recognising polite lies), recognising the difference between a complaint and a churn signal — into a deterministic daily brief + a refund dual-key gate.
Why this exists
Customer support across 11 brands cannot wait for the founder's morning cycle. The CS-tier1 bot handles inbound triage, drafts replies for review, escalates enterprise churn threats, and surfaces NPS / FCR / GRR slips before they become a board concern.
Decision rights
Autonomous:
- Compute + emit CS metrics every cycle
- Draft daily CS brief (routes through pii-redactor + draft_review)
- Draft replies for inbound tickets (routes through telegram-draft-for-review — Scribe gate)
- Approve refunds up to $100 (configurable via
TAO_CS_REFUND_CEILING) - Alert on threshold breaches
HITL (dual-key gate):
- Refund > $100 (any single transaction)
- Public apology / postmortem
- Service credits beyond standard SLA terms
- Enterprise churn-save offer (any concession to a flagged customer)
Metrics owned
| Metric | Formula | Floor / target | Alert | Critical |
|---|---|---|---|---|
| NPS | (% promoters) − (% detractors) × 100 | ≥ 50 | < 30 | — |
| FCR (First Contact Resolution) | resolved_first_contact / total_tickets | ≥ 0.75 | < 0.65 | — |
| GRR (Gross Retention Rate) | 1 − (churn / start_count) | ≥ 0.95 | < 0.90 | < 0.85 |
| Avg first-response (minutes) | mean over window | < 30m | > 60m | > 240m |
| Open enterprise churn threats | count of flagged accounts | 0 | > 0 (any) | ≥ 3 |
Cadence
| Trigger | Action |
|---|---|
| Daily 06:00 UTC | Assemble brief → pii-redact → draft_review post |
| Per inbound ticket | Draft reply → Scribe gate via telegram-draft-for-review |
| Per refund request | Dual-key gate; emit cs_refund_approved or cs_refund_blocked |
Refund approval flow
incoming refund request {amount_usd, customer_id, business_id, justification}
→ if amount_usd <= AUTONOMOUS_REFUND_CEILING_USD ($100):
emit cs_refund_approved → return approved
→ else:
draft_review.post_draft(
draft_text="💬 Refund approval — ${amount} to {customer} ({biz})",
destination_chat_id=REVIEW_CHAT_ID, drafted_by_role="CS")
emit cs_refund_blocked (queued, awaiting reaction)
return pending
Contract
Daily brief output:
💬 CS daily — {date}
Avg NPS: 47 | Avg FCR: 78% | Avg GRR: 96% | 3 alerts (1 critical) | 1 enterprise threat
🚨 Alerts:
🔴 [synthex] grr_pct: 84% < 85% — churn investigation required.
🟡 [restoreassist] nps: 22 < 30 — detractors outpacing promoters.
Per-business (worst NPS first):
- synthex: NPS 12 | FCR 71% | GRR 84% | first-response 92m
- ...
📥 2 refund approvals queued (>$100)
Metric snapshot row (per cycle, written to .harness/swarm/cs_state.jsonl):
{
"ts": "ISO-8601",
"business_id": "restoreassist",
"nps": 47.0,
"fcr_pct": 0.78,
"grr_pct": 0.96,
"avg_first_response_minutes": 22.0,
"open_enterprise_churn_threats": 0
}
Safety bindings
- All outbound replies route through telegram-draft-for-review (Scribe gate). No autonomous send.
- Refund ceiling persisted ($100). Editing requires user approval, audit-trailed.
- PII-redactor in front of every draft — customer names, emails, ticket IDs go through pii-redactor at strictness=standard.
- Enterprise churn threats are critical the moment they're opened — escalates to founder via daily brief.
- Kill-switch aware. On
TAO_SWARM_ENABLED=0, the daily brief is queued (not sent).
When NOT to use
- Tier-2 / engineering-escalation tickets — those route to CTO bot via flow_engine
- Compliance / data-deletion requests (GDPR / CCPA) — those route to GC + CCO bots (Phase B)
- Marketing-driven outreach (NPS surveys, feedback campaigns) — that's CMO bot
Verification (Wave 4 A4)
- Synthetic data → NPS / FCR / GRR computed correctly
- GRR < 85% → critical alert fires
- First response > 240m → critical alert fires
- Open enterprise threats >= 3 → critical alert fires
- $50 refund → auto-approves
- $200 refund without draft_review → blocked
- $200 refund with draft_review → pending, no real refund issued
- Daily brief assembles cleanly with all 11 businesses
References
- Blueprint:
~/Pi-CEO/Senior-Agent-Operations-Blueprint-2026-05-02.md§Role 17 Customer Success - Parent ticket: RA-1862 (epic RA-1858)
- Outbound substrate:
skills/telegram-draft-for-review/SKILL.md(Scribe gate) - HITL gate substrate:
swarm/draft_review.py - Multi-agent debate scaffold:
swarm/debate_runner.py(RA-1867)