Case study · Harbor Hotels Group

Voice AI that books rooms.

A 24/7 voice agent across 47 properties that handles reservations, upgrades, and guest requests at human-grade quality — for 31% less than the prior BPO.

31%COST REDUCTION VS BPO
3.1MCALL MINUTES HANDLED
72%FIRST-CALL RESOLUTION
4.7/5GUEST SATISFACTION
Engagement details
CLIENT
Harbor Hotels Group
INDUSTRY
Hospitality · 47 properties
DURATION
16 weeks · ongoing operate
TEAM
4 Mindlytic seniors + 2 Harbor stakeholders
STARTED
Q1 2025
STATUS
Live · 24/7 across all properties
THE PROBLEM

A BPO that couldn't book rooms after 9pm.

Harbor's reservation calls were routed to a third-party BPO that staffed up during day hours and shed capacity overnight. Result: 38% of calls between 9pm and 6am went to voicemail, and conversion on those callbacks was 11% — a quarter of daytime numbers.

Harbor had spent $2.3M on the BPO contract the prior year. Their leadership wanted 24/7 coverage at quality parity with daytime, without doubling spend.

WHAT WE BUILT

A voice agent on every Harbor reservation line.

We deployed a Twilio-fronted voice agent backed by GPT-4o Realtime, with property-specific knowledge bases (room types, amenities, policies) and live integration with Harbor's PMS for availability + booking. Warm transfer to a human concierge for VIPs and exceptions.

The agent runs 24/7 across all 47 properties. Daytime, it handles overflow when humans are busy. Overnight, it handles 100% of inbound until the morning shift starts.

  • Inbound + outbound calling on Twilio Programmable Voice
  • Per-property knowledge index in Postgres + pgvector (rates, room types, amenities)
  • Live PMS booking integration via Harbor's existing API
  • Warm transfer with full conversation summary to night managers
  • TCPA + PCI-DSS compliance for payment capture
WHAT WE FOUND

The first 300ms of silence is the whole game.

Our first prototype had an 1100ms median turn latency. Guests hung up. We rebuilt the orchestrator to start TTS streaming on the first sub-clause from the LLM, dropped median to 600ms, and the hang-up rate fell 84% overnight.

We also discovered that hotel guests interrupt agents constantly — to add a date, change a room type, ask about parking. Without barge-in tuned to the right VAD threshold, the agent talked over guests and felt robotic. Two weeks of acoustic tuning fixed it.

THE OUTCOME

Live across 47 properties. 31% less spend.

Six months in, the agent handles 78% of inbound calls without human assistance. First-call resolution sits at 72%, ahead of Harbor's daytime BPO benchmark of 64%. Guest satisfaction (post-call survey) averages 4.7/5.

Harbor cancelled the BPO contract in Q3 2025 and routed the savings into expansion at four new properties — all of which launched with the voice agent live on day one.

Voice AI that our guests actually enjoy. We thought that was impossible until night four of the pilot.
Rafael Ortiz
Director of Operations, Harbor Hotels Group

What we used.

LLM
GPT-4o Realtime
STT
Deepgram Nova-2
TTS
ElevenLabs Turbo
Carrier
Twilio Voice
Backend
FastAPI
DB
Postgres + pgvector
Infra
AWS ECS
Obs
Datadog + LangSmith

Other case studies.

Have an AI project in mind?

Book a 30-minute discovery call. We'll tell you if we're the right fit — and if not, who is.