AI for Treuhand work, on terms you can defend

01

Why the usual AI tools are a non-starter for a Treuhänder

Your obligations do not stop at good intentions. The FADP governs how client data is handled, and under the revised FADP, fines target the responsible individuals — which for a small fiduciary means the managing director personally. Pasting a client’s file into a foreign consumer chatbot is a data-handling decision you would have to own, and most Treuhänder rightly refuse to.

Autonomous agents raise the stakes further: an agent that can send email or write to records is not a drafting aid, it is an actor. In 2026 the exposed-instance incidents around self-hosted agents showed what that looks like without governance. For a profession built on discretion, the requirement is specific: know where the data sits, know which model saw it, and hold every outward action for a human.

• Persistent data stored in Zurich, Switzerland; processing on EU infrastructure
• Swiss track: Swiss-hosted open-weight models (Infomaniak) — inference stays in Switzerland when you choose it
• Outward actions wait as pending proposals; the run blocks until you decide

02

What governed autonomy looks like in a fiduciary practice

The agent takes the recurring load: it drafts client correspondence and deadline reminders and holds each one as a proposal for your sign-off; it summarises and organises the documents you upload; it chases missing records with drafted follow-ups you approve before they send; it runs research and keeps mandate notes in long-term memory; and scheduled runs carry the month-end routine without you re-briefing it. It works through whatever tools you connect via MCP — it does not require you to change your stack.

The controls carrying the load are architectural, not promises: the approval gate is default-deny, so anything mutation-capable pauses for you; credentials sit in a write-only vault the model can never read back; residency-aware routing keeps storage in Zurich and lets you pin inference to Swiss-hosted models; and the audit trail records every proposal, approval, rejection, and dispatch against Ed25519-signed checkpoints. The EU AI Act’s Article 14 asks for human oversight as an external control rather than a prompt instruction — the proposal gate is exactly that kind of control, and it is the architecture monopea was built around.