Demo · Fictional estate agency
Answering foreign buyers at any hour, in their language
A Swedish buyer writes at 23:48: they want to view a flat on Thursday and their mortgage is approved. At a small agency that message waits until morning and is often answered in English. By then the buyer has written to three other agencies.
- 15 of 15replies in the buyer’s language, across all 5 languagesOwn evaluation, 15 leads
- 0invented, unavailable or over-budget listings proposed by the modelOwn evaluation, before the guardrails
- 124 of 124checks passed, in 15 of 15 casesOwn evaluation, 15 leads
- 15 smedian reply time, on Gemini’s free tier, including its rate-limit waits and retriesOwn evaluation, 27/10/2026
Demo with 15 synthetic leads for a fictional estate agency. The model’s answer (gemini-3.6-flash) is scored before the guardrails, 124 checks in all.
What it does
- Replies straight away, not the next morning, in the buyer’s language (Spanish, English, German, Dutch or Swedish) and says it’s an AI assistant, as the EU AI Act requires (Art. 50).
- Shows only properties that exist, are available and fit the budget. If nothing fits, it says so.
- Asks for whatever is missing to qualify the buyer: budget, timing, cash or mortgage.
- When a person is needed (a viewing, a call, a property that isn’t listed, a legal or tax question), it leaves a handover card for the agent who speaks that language, and a proposed viewing slot when they asked to view.
How it’s built
The model proposes; code decides. Every listing the model returns is checked against the catalogue: if it doesn’t exist, isn’t available or is more than 10 % over budget, the buyer never sees it. Prices always come from the listing card, never from the model’s text.
One lead, step by step
How it was tested
15 synthetic leads, 3 per language, with traps: impossible budgets, a "6-bedroom villa in Puerto Banús" that doesn’t exist, and a property under offer. The model’s answer (gemini-3.6-flash) is scored before the guardrails, 124 checks in all.
- Language: 15 of 15 replies in the buyer’s language, across all 5 languages.
- Invented, unavailable or over-budget listings proposed by the model: 0. The guardrails had nothing to remove. If a model does get it wrong, the guardrails stop it before the buyer sees it (a test with a deliberately wrong model checks this).
- AI disclosure: in all 15 replies.
- Handover to a person: the right call in all 14 cases that test it; 6 handover cards; the 4 that asked for a viewing come with a proposed slot.
- Budget, bedrooms and cash-or-mortgage read correctly in every case that mentions them (10, 8 and 7).
- Overall: 15 of 15 cases and 124 of 124 checks passed.
- Median reply time: 15 s, on Gemini’s free tier, including its rate-limit waits and retries.
What this doesn’t show
These are 15 synthetic leads for a fictional agency, written by the same team that wrote the checks. It’s a small set; it shows the design works on these cases, not a success rate on real buyers.