Grow2.ai spent three weeks in summer 2025 shadowing a Berlin short-stay operator (45 units across Mitte and Friedrichshain). We did not ship an agent. The brief was: "watch us, write down what would actually help." This is the field report.
Three things we expected · and were wrong about
- We expected the bottleneck to be reply latency. It wasn't. The ops team replied to ~88% of inbound within 12 minutes during business hours. The bottleneck was 22:00–08:00, when 41% of total messages came in but the team was off.
- We expected guests to want detailed pre-arrival info. They didn't. Guests arriving from a flight wanted three things, in order: the lockbox code, the WiFi password, and confirmation that the cleaner had actually finished. Anything past that — restaurant recommendations, neighbourhood notes — was read after they'd slept.
- We expected German guests to expect formality. 60% of guests preferred informal address by message 3. The Polish, Italian and Spanish guests preferred formal in writing but warm in voice. The British guests would say "cheers mate" on message 1.
The 02:00 message problem
During the three weeks, Grow2.ai logged 167 messages received between 22:00 and 08:00. Categorised:
- Lockbox / access (41%) — guest arrived later than expected, forgot the code, or the code didn't work first try.
- WiFi (18%) — password didn't work, network not visible.
- Heating / hot water (14%) — usually a tripped breaker that the guest cannot find.
- Pharmacy / hospital (11%) — child's fever, cut finger.
- Other (16%) — cleaning issue, neighbour noise, key snapped.
Of those 167, the Berlin team replied within 30 minutes to 12. The rest waited until morning. That is the gap an agent fills cleanly — three of the five categories above are answerable from a per-unit wiki, the fourth (pharmacy) is answerable from a 200-line geocoded list, and the fifth (true emergencies) escalates to a human pager.
The bias we accidentally trained · for free
Towards the end of week 2, the operator started using a free GPT-based translator to help with a Polish family they couldn't reply to in time. The translations were grammatically clean but read as oddly formal — "Sehr geehrte"-style register on a 02:14 lockbox question. The Polish guest sent back a polite-but-baffled "Dziękuję bardzo, ale nie ma sprawy" ("thanks a lot, but no big deal") which is the Polish equivalent of cooling the temperature.
The lesson Grow2.ai took away: a translator is not a replier. Tone is downstream of context the translator doesn't have. This is precisely what the per-language tone profile (see FN-013) exists to fix.
What the operator did with the report
They did not commission a pilot. They hired a full-time Polish-speaking night-shift coordinator instead. That was, given their unit count and the cost of an agent that handles five languages around the clock, the correct economic call for them at that moment. We told them so. They have since referred two other operators to us — both became pilots.