Nest Concierge

Context

For a short-stay portfolio, the agent answers FAQs (wifi, check-in, parking), qualifies emergencies and hands off a summary when humans must act. Playbooks are versioned per listing: a wrong access code is an incident, not a “creative approximation”.

Challenge

Night messages, per-property rules, costly error risk. Support that does not scale with headcount.

Solution

Per-property playbooks, intent taxonomy, escalation matrix (refund / access / safety = never auto), reply runtime + dashboard (auto rate, re-open, complaints).

Baseline

Without an agent: 3-5h/day messaging for many multi-unit ops (market claims). Realistic pro automation often caps near 60-70%; vendor 90% assumes mature playbooks and few ambiguous cases.

Results

• Design target ~70% of messages without human after burn-in
• Forced escalation on sensitive cases with context summary
• Critical alerts in minutes (webhook / push)
• Portfolio scales without 1 support hire per N listings

Architecture decisions

• Choice: Versioned playbooks per listing

  Why: Listing context is the source of truth; a global prompt invents codes.

• Choice: Escalation matrix before the model

  Why: Without it, 80% auto becomes an incident generator.

• Choice: Strong HITL burn-in (weeks)

  Why: Serious vendors use approval gates before full auto.

Limits

• Price negotiation, disputes and safety emergencies stay human.
• A flaky channel connector (email vs API) kills auto rate faster than the model.
• Auto % must be defined: reply sent vs thread closed without human.

Checklist to rebuild this

1. Inventory channels (Airbnb, Booking, WhatsApp, email…)
2. Write a minimal playbook per unit (access, wifi, check-in/out)
3. Intent taxonomy + auto / draft / human matrix
4. Burn-in: human approval before send
5. Dashboard: auto rate, re-open, complaints, escalation time
6. On-call owner + alert SLA
7. Ban auto on access codes and refunds
8. Weekly review of intent false positives