Confidentiality kept by architecture, not by policy

01

Why an ungoverned agent cannot touch client matters

An autonomous agent is an actor, not a search box. Give it an inbox and it can send; give it a document store and it can write. For a profession bound by professional secrecy, the failure mode of an ungoverned agent is not embarrassment — it is a privileged communication leaving the firm without a lawyer having decided it should. The 2026 wave of exposed self-hosted agent instances showed how quickly "autonomous" becomes "unsupervised" when governance is a system-prompt instruction instead of an external control.

The requirements are concrete: know where matter data rests, keep credentials out of the model’s view entirely, be able to work on something sensitive without leaving a durable trace, and put a human decision in front of anything that leaves the building. Each of those is an architectural property of monopea, not a usage guideline.

• Matter data stored in Switzerland, processed on EU infrastructure
• Write-only vault: a secret can be stored but never read back — by you, the API, or the model
• Incognito turns run with no durable writes for the genuinely sensitive questions

02

What the agent does inside a practice — and where it stops

In practice, the agent carries the drafting and organising load: it summarises and structures case documents you upload, drafts client letters and engagement correspondence as pending proposals, prepares research memos on the questions you frame, keeps matter context in long-term memory so you brief it once, and runs scheduled routines — deadline sweeps, status digests — without being re-asked. It works through the tools you connect via MCP rather than demanding a new stack.

Where it stops is the point of the product: anything outward-facing waits at the approval gate, where the responsible lawyer sees the exact draft and the exact tool call before it is dispatched. The vault means the model references credentials as {{secret:NAME}} placeholders and never sees the value; incognito mode covers the conversations that should not persist; and the audit trail — every proposal, approval, rejection, dispatch, signed in checkpoints — is the record that lets a firm say precisely what its AI did on a matter.