Building an AI Assistant for Swiss Property Managers in 3 Days
From zero to a working managed agent on Anthropic's new platform — what it takes, what it costs, and what I learned.
This week I shipped the skeleton of an AI assistant for small Swiss Hausverwaltungen. Three days. Live agent on Anthropic's Managed Agents beta. Real system prompt. Real pricing model. Zero reference customers — yet.
Here's the build log.
The gap
Switzerland has roughly 9'000 property management firms. Most of them are tiny — five to twenty employees, family-run, drowning in email. The German AI incumbents (casavi, Cambioo, Nestermind) are built for enterprise clients with 200+ staff and procurement departments. Nobody is serving the long tail in Schweizer Hochdeutsch with a managed-service pricing model.
That's the gap.
The stack
- Platform: Anthropic Managed Agents (beta `managed-agents-2026-04-01`)
- Model: Claude Sonnet 4.6
- Tools: `agent_toolset_20260401` (bash, file ops, web search, MCP)
- Prompt caching: aggressive, ~70% token reduction on repeat sessions
- COGS per client: ~CHF 100–150/month at realistic usage (500 emails/mo)
- Sell price: CHF 1'800/month managed service
- Gross margin: ~92%
The system prompt
7KB of CH-German conventions — Sie-Form, kein ß, no comma after Anrede, 5 email categories, P1/P2/P3 priority levels, 8 hard constraints. The non-obvious part: Swiss Hochdeutsch isn't just "German without ß." It's a specific register that avoids certain loanwords, uses different verb forms, and has tone expectations that differ from German German.
I stress-tested the prompt against 10 synthetic email fixtures (tenant complaints, damage reports, visit requests, admin forms). The failures taught me more than the successes.
What I got wrong on day one
1. I pitched the service at CHF 390/month. Three parallel research agents politely told me I was undercharging by 4–5x. The floor for a Swiss managed AI service is CHF 1'800/month. Anything lower signals "hobby project" to buyers who are paying CHF 8'000/month for an FTE.
2. I thought domain expertise was the blocker. It isn't. The blocker is zero reference customers, and the fix is one good pilot — not ten years of SVIT experience.
3. I overbuilt the toolset. First draft had 14 tools. The actual product needs 4: read email, draft response, search templates, escalate. Everything else is premature.
What's next
Outreach starts next week. Twenty cold emails per day, Tuesday–Thursday 08:00–10:00. Founder-honest positioning: "I'm not a 20-year veteran, I'm an AI engineer who finished the SVIT course and wants to build a real product with you. Pilot price is half retail. Cancel any time."
I'll publish the reply numbers here next Friday.