Cases · Case 04 · Atlas Rentals · PROPERTY MGMT · KRAKÓW
Case study · Drawing G2-039 · Atlas RentalsSheet 04 / 04

Guest reviews 4.3 → 4.78 across 84 apartments — STR operator · Kraków

Drawn by V. K. · Auspex Co.Pilot Q1 2025 · in service · 14 mo.
Client Atlas Rentals
Vertical Short-term rentals
Size 84 units · PL/UA/EN/DE
Module M-04 Concierge
Pilot 2025-02-04 → 2025-02-18
Case №
P-09
In service · 14 mo.
§ 01The number in the contract — and the four around it
Avg. Booking review score
4.78/5
Was 4.30 · contract ≥ 4.65
Median first reply
42sec
Was 1h 11min · 24/7
Repeat-booking rate
+38%
YoY same-period
Ops time saved
19h/wk
Across 3 ops staff
§ 02Existing condition
Existing condition · before pilot

Beautiful flats. 3am check-in messages.

Atlas Rentals operates 84 short-term rental apartments across Kraków's old town and Kazimierz, listed on Booking.com and Airbnb. Their guests are international: weekend Berliners, business travelers from Lviv, summer Italians, year-round Brits.

The operations team — three people — was the bottleneck. Median first response to a guest message was 1h 11min; in practice, anything that arrived between 22:00 and 08:00 waited until morning. Lockbox codes, late check-in confirmations, "the kettle is broken", "where's the closest pharmacy" — all delayed.

The visible cost: a Booking.com average score of 4.30, dragged down by communication ratings of 4.0 and below. Founder Tomasz could trace every score-recovery review request directly to a slow night reply.

Contract number: average Booking.com score ≥ 4.65 across all 84 units, measured on rolling 60-day window after Day 30.

EXISTING · MESSAGE LATENCY · 24h CYCLE8h4h1h022:00 — 08:00 · NIGHT⌀ 1h 11minDAYEVENIGHTFIG. 02-A · ATLAS BASELINE · 8 WK · 84 UNITS
§ 03Proposed assemblyDay 1 → Day 14
PROPOSED · CONCIERGE · M-04 · 84 UNITSBOOKING.COMAIRBNBWHATSAPPEMAILHOSTAWAYPMS · 84 UNITSCODES · WIFI · WIKIG2 · CONCIERGEPer-unit knowledgeLockbox · WiFi · rulesLocal recs · routingPL · UA · EN · DE · IT⌀ 42 sec · 24/7+ REVIEW REQUESTpost check-out · day 2GUESTREPLYCLEANERDISPATCHMAINTENANCETICKETOPSESCALATIONFIG. 03-A · CONCIERGE ASSEMBLY · 5 LANG · 84 UNITS
Proposed assembly · M-04 Concierge

Per-unit knowledge. Five languages. 24/7.

The agent's brain isn't generic — it's per-apartment. We built an ingestion pipeline from Hostaway (their PMS) plus a manual annotation pass with the ops team: each of the 84 units has a structured wiki entry (lockbox code rotation, WiFi password, quiet hours, broken kettle history, the bakery downstairs, the night pharmacy, the trick to the bathroom door).

The agent answers in the guest's language of arrival: PL, UA, EN, DE and IT cover 96% of Atlas's guests. We chose those by tracing 14 weeks of historical messages, not by checklist.

Two non-obvious wins. First, the agent dispatches cleaners and maintenance directly when a guest reports something — no human in the loop for "the kettle is broken." Second, on Day 2 after check-out, the agent sends a personalised review request that references one specific moment from the stay. Communication scores moved first; cleanliness and value scores followed.

§ 04Construction sequence14 days
DayStageOutputOwner
Day 0BriefNetwork audit · 84 units · 5 languages confirmed · KPI co-signedVlad K. · Tomasz
Day 1–4Hostaway syncPMS bidirectional integration; per-unit wiki seeded from Hostaway fieldsOlha B.
Day 5Wiki annotation passOps team adds 84 unit-quirks (kettle, doors, bakeries) · 9 hrs totalIryna S. · Ola K.
Day 6ChannelsBooking.com + Airbnb + WhatsApp + Email wired with consentMarko Y.
Day 7–10Shadow modeAgent drafts; ops sends. 528 conversations evaluated.All
Day 11Dispatch rulesCleaner/maintenance auto-dispatch rules signed off · escalation thresholdsIryna S. · Tomasz
Day 12–13Live · ramp10 units → 84 units. Daily 19:00 metrics call.Marko Y.
Day 14Hand-offKPI measured rolling 60d from Day 30. Move to ongoing service.Kateryna L.
§ 05Measured outcomes

Contract said ≥ 4.65. Steady state is 4.78.

The first lift was communication scores — they crossed 4.8 inside the first 30 days because guests were getting answers at 03:00 in their own language. Cleanliness and value scores followed once the auto-dispatch was running smoothly: a broken kettle reported at 21:00 was a fresh kettle by 09:30 the next morning, with no human op pinged.

  • 4.30 → 4.78 Booking.com average across 84 units (rolling 60d, months 3–14)
  • 42-second median first reply (was 1h 11min). Night response unchanged across 24h cycle.
  • +38% repeat-booking rate YoY same-period — driven by review-driven discoverability + DM follow-ups
  • 19 hours/week saved across 3 ops staff — Ola moved fully off messaging into property growth
  • 0 escalations missed during pilot · 2 missed in 14 mo. service (both human override errors)

What we got wrong. Initially the agent suggested the closest-rated restaurant on TripAdvisor when guests asked for dinner spots. Tomasz hated it: it was generic, and Atlas had a pile of curated relationships with neighborhood places. We swapped TripAdvisor for an Atlas-curated rec list per unit. Recommendation conversion (guests actually visiting) tripled.

BOOKING.COM SCORE · ROLLING 60-DAY · 84 UNITS5.04.84.64.44.2CONTRACT · 4.65PILOT4.78COMMUNICATION (proxy)PREM1M3M14FIG. 05-A · STEADY STATE @ 4.78
Client testimonial

For the first time in seven years of running flats, my ops team is not on their phones at midnight. The guests are happier. Our reviews speak for themselves.

Tomasz W., Founder, Atlas Rentals · Kraków
§ 06Drawing metadata
Engagement
Pilot 14 days · ongoing service since 2025-02 · two named engineers · weekly review (90 days), monthly thereafter
Stack used
Claude Sonnet (reply) · GPT-4.1 (review-request copy) · Temporal · Hostaway API · Booking.com / Airbnb messaging APIs · Twilio (WhatsApp)

Run flats, hotels, or stays? Let's draw your concierge.

A 30-minute call, your booking platform audit, the score number that goes in your contract.

▸ Commission a pilot
Case 04 · Atlas Rentals — short-term rentals · Kraków · Grow2.ai