Institutions want what AI can do and fear what it might send. Both are reasonable.
The systems we build follow one rule. The assistant drafts the step, and a named person reads it and sends it. Nothing leaves in the institution’s name without a person’s decision.
In practice the assistant fills a form from one message, reads an uploaded file, and asks one question at a time when something is missing. The person sees exactly what will be sent before it goes.
Two more choices matter. The language model can be changed, so the institution is not tied to one supplier. And it can be switched off, so that nothing leaves the server. Where the client needs it, the data stays in the country.
The result is quieter than most AI projects. The work gets faster, and the responsibility stays where it was.