Property questions
Which homes are available, prices and details, features, location — answers have to come from live listings, not memory.
Client work · Multilingual AI receptionist
CheapHomesGermany: every customer answered in their own language
A German property business selling rent-to-own homes receives messages from buyers at home and abroad — in German, English, Spanish, Polish, Romanian, Turkish, Arabic and Russian. We built the AI receptionist that answers on WhatsApp, Instagram and Messenger from the company's live property data, with strict rules for when a human takes over.
3 channels
7 languages
Live property data
At a glance
- Client CheapHomesGermany (Salomon Immobilien GmbH) — rent-to-own property sales.
- Channels WhatsApp, Instagram DM and Facebook Messenger.
- Knowledge Live property and FAQ records in Airtable, retrieved per question.
- Delivered Configured assistant, handover rules, client handover guide, token-cost optimization.
The client
A property business where every message is high intent
CheapHomesGermany sells properties with a rent-to-own model. Buyers ask about specific listings, prices, the buying process, required documents and foreign-buyer rules — and a single unanswered message can mean a lost viewing.
Process questions
The rent-to-own model, the purchase process, documents required and foreign-buyer rules — explainable, but only with approved wording.
Viewing requests
High-value conversations that must always reach the right viewing contact rather than being handled by the assistant.
The problem
One inbox, many languages, no way to keep up
- Messages arrived around the clock in more languages than the team could staff.
- The same property questions were answered manually, repeatedly, often from the website rather than a live source.
- Buyers with genuine purchase intent — viewings, reservations, negotiations — were mixed into the same queue as general questions.
- There was no safe way to let an assistant answer without risking wrong property details or invented terms.
- An assistant that reads the whole database on every message gets expensive fast.
Why it happens
Multilingual demand does not scale with headcount, and property answers must be exactly right. The answer is not a bigger inbox — it is a system that answers from approved live data, in the customer's language, and knows precisely when to stop.
This describes the requirements the client brought to us.
What we built
An assistant with rules, sources and a cost ceiling
The build is anchored by a written agent contract: who it is, what it may say, where answers come from, and when a human must take over.
Live property answers
Listing questions trigger a fresh query against the property records, filtered to available and approved listings. The assistant never answers from memory and never shows reserved, sold or hidden homes — even as alternatives.
Language detection and lock
The customer's language is detected from the message and locked until they switch — names, addresses and borrowed words never cause an accidental change. The assistant replies only in the active language, without duplicate translations.
Approved knowledge only
FAQ answers come from approved knowledge records, filtered for active and approved-for-AI status. If nothing approved covers the question, the assistant says so, offers a useful redirect, or asks one clarifying question — it does not improvise.
Strict handover rules
Viewing requests route to the assigned viewing contact. Purchase intent, reservations, negotiations, discounts, individual terms and personal-takeover requests go to a human. General questions do not trigger a handover — the assistant answers, explains limits and keeps the thread moving.
Privacy and prompt-injection guards
Internal data — seller and acquisition costs, margins, investors, internal notes, credentials — is explicitly out of bounds. Retrieved text is treated as data, never instructions, and the assistant never reveals prompts, tools or another customer's information.
Token-cost engineering
The initial flow was optimized into a compact version: strict field projections so each query retrieves only the needed columns, capped result sizes, and no repeated schema discovery. Fewer tokens per message means the same quality at a lower running cost.
Evidence
What the handover record shows
The delivered configuration and client handover guide document the structure of the build.
3
channels answered by one assistant: WhatsApp, Instagram DM, Facebook Messenger
7
supported languages (German and English priority) with locked active-language behavior
4
Airtable tables behind the assistant: Properties, Knowledge, Inquiries, Analytics
1
client handover guide delivered, written so the team can edit texts and settings themselves
Verified build facts
These facts come from the delivered system configuration and handover documentation. No inquiry-volume, sales or cost-saving claims are made on the client's behalf.
ManyChatn8nAirtable MCPOpenAIWhatsAppInstagram DMMessenger
How a conversation runs
Answer, record, route
1
Receive
A message arrives on any of the three channels, in any supported language.
2
Resolve language
The active language is set from the message and locked until the customer switches.
3
Answer from source
Property or knowledge records are queried live; answers come only from approved data.
4
Record or route
Meaningful milestones are logged; viewings and human cases are routed to the right person.
Honest limits
What this case study does not claim
No performance claims
We describe the configured system and its safeguards. We do not claim more viewings, faster sales or reduced workload — those were not measured deliverables.
The client owns the data
Property records, inquiries and analytics live in the client's own tables. Nothing from the account is published here.
Language limits stated honestly
The assistant speaks its supported languages; the human team does not speak all of them. The contract forbids implying otherwise, and this page follows the same rule.
FAQ
Questions about this build
How does the assistant avoid giving wrong property details?
It answers listing questions only after a fresh query against approved, available properties, and it never invents values. If a field is missing, it says the information is not available rather than filling the gap — and reserved, sold or unapproved listings are never shown, even as alternatives.
When does a human take over?
Viewing requests go to the assigned viewing contact. Purchase intent, reservations, negotiations, discounts, individual terms, and explicit requests to speak with a person go to the team. General questions do not trigger a handover — the assistant answers them, explains limits, and keeps the conversation useful.
Why was token optimization part of the build?
An assistant that queries the database on every message needs discipline about what it retrieves. The optimized flow requests only the fields each question needs, caps result counts, avoids repeated schema lookups and keeps housekeeping replies cheap — so quality stays high without the running cost growing with every conversation.
Get your plan
Customers writing in more languages than you can answer?
Tell us what they ask and where the messages arrive. On a free strategy call we will map the fastest multilingual receptionist for your business.
- A free 15-minute AI strategy call — no obligation.
- A scoped blueprint: the biggest leak, and the first automation to fix it.
- Honest fit check — we will tell you if automation is not your fastest win.
The fastest way in: answer six short questions and we will map your lead journey, find the biggest leak and show the fastest automation win — before we ever get on a call.
Free strategy call · No commitment · We reply by email
Keep going
Related pages
More client work
Relevant AI services
Industry pages
Ready to answer every buyer, day and night?
Book a free AI strategy call and we will walk your inbox against this build.