PropertyPilot.ai
From a conversation with a host to an AI host, tested with 2 holiday rentals
The story
Before
A property rental company with 2 holiday homes in the Netherlands answers the same questions all day. WiFi, towels, check-in. Via WhatsApp, Booking, Airbnb. Being available 24/7 is exhausting.
Obstacle
Guests ask questions in broken English, German, French. Simple chatbots fail on nuance. Generic AI hallucinates answers. They don't dare to trust it without supervision.
Insight
Start small. Two homes. One host who dares to test. Build the AI host that learns from real guest questions. Validate the concept before scaling.
After
14 days later: AI host live at 2 holiday homes. Question comes in, AI analyses, answers or escalates. The host monitors, teaches the system. Now 2 months later: 90%+ answered automatically.
The first 14 days
From concept to working prototype.
Discovery: which questions keep coming back?
We analyse historical messages. "Where is the WiFi password?" comes back 50x per month. "Extra towels?" 30x. We make a top 20 of repetitive questions.
Designing the knowledge base structure
How do we organise the knowledge? Per home, per category, per question type. The host must be able to update it easily without tech skills.
AI model selection and RAG setup
We choose GPT-4 with RAG (Retrieval Augmented Generation). The AI searches the knowledge base first before it answers.
Multilingual support integration
DeepL API for detection and translation. Question comes in in German? AI answers in German. Knowledge base stays Dutch, translation happens on the fly.
Building the knowledge base editor
Simple interface for hosts. Type "WiFi password is on the fridge", save, done. No complicated admin panels.
Confidence scoring and escalation
AI gives every answer a confidence score. Below 90%? Question is forwarded to the host via chat. Safety built in.
First test with real questions
We feed in 50 historical questions. How does the AI react? Which answers are good? Where does it doubt? We fine-tune the prompts.
Booking.com webhook integration
New Booking messages come in automatically. The AI reads, assesses and answers, within 30 seconds.
Airbnb and Natuurhuisje connections
Same flow for other platforms. One knowledge base, all channels. The host only has to update one place.
Learning loop: from escalation to knowledge
When the host answers a question, he can directly say: "Save this for next time." The AI learns with every interaction.
Response time optimisation
From 30 seconds to 5 seconds. Caching, smart routing, parallel processing. It has to feel instant.
Dashboard for hosts
Overview: how many questions answered? Which escalations? What knowledge is missing? Insights to improve the system.
Tone of voice fine-tuning
The AI must sound like a friendly host, not like a robot. We test different tones. Warm but professional.
Go-live with first holiday home
We go live with one property. Real guests, real questions. The host monitors along. It works. First 10 questions: 100% correctly answered.
1-3 months later
Two months later: understanding broken English
After two months about sixty real guest questions have been answered, and the AI was also trained on a year of the host's old messages. The big lesson: guests often write in broken English. "Where toilet paper extra?" Grammatically wrong, but the intent is clear. The confidence score now also judges language quality: with poor English the confidence has to be even higher (95 percent or more) before an answer goes out.
1 year later
What comes next
PropertyPilot grows host by host, with a pilot for each new one. On the list: seasonal knowledge (heating in winter, barbecue in summer), information per region and connections to smart devices in the home.
"I slept badly because guests still asked questions at 11 PM. Now the AI answers immediately, in their own language. My reviews are up, my stress is gone. This is magical."