Proton has released Lumo Desktop, a version of its Lumo AI assistant for Windows, macOS and Linux. It can work with the files on your computer and with the apps you connect to it, such as Proton Mail, Google Drive, Jira and Confluence. The part that deserves your attention is the permission screen, because it asks a better question than most desktop AI does.
## The useful question is what the AI may do Most privacy talk about AI is about what it can see. On a desktop, what it can do matters at least as much. An assistant that reads a document or an email can be steered by instructions hidden inside it, and what happens next depends on its powers. If it can only read, the worst result is a wrong answer. If it can also send and delete, a hidden instruction can email your files to a stranger or wipe a folder.
Lumo sets those powers one action at a time. For each connector you pick automatic, ask first or blocked, and each scheduled task has its own set of rules. Proton's own example is to let Lumo search your inbox, require approval before it sends an email, and block it from deleting messages.
Apple is moving the same way for the whole Mac. It has announced that a coming macOS update will let you make an app ask permission every time it wants broad access. That is the folder-level version of the idea, and it can get tedious. Sorting by action avoids most of that tedium. An assistant reads files hundreds of times a week and sends an email a few times, so asking only before the rare risky action keeps the prompts rare enough that you still read them.
## What it looks like in a small business Picture a shop owner who keeps supplier invoices in one folder and card receipts in Google Drive. She connects both, lets Lumo read them freely, and schedules a monthly review that flags charges that look wrong. She also connects her mail and sets sending to ask first and deleting to blocked.
The review runs while she is serving customers. If it finds a double charge and drafts a complaint to the supplier, the email waits for her approval instead of going out in her name. If checking invoices by hand takes an hour a week, the assistant hands back most of about four hours a month, and she still decides on every message that leaves.
Lumo is not the only assistant that asks first. Anthropic's Claude Cowork always asks before permanently deleting files, in any mode. The habit is spreading, and it is a rule you can carry from one assistant to another.
## What the rules do not cover The permission rules control the assistant's hands. They do not control where your files go when it reads them. Lumo is built on open-source language models that run in Proton's own European data centers, so whatever it reads has to travel there for the model to work with it. Proton promises that it keeps no server-side logs, never trains on your files or connected apps, and encrypts chat history so only you can read it. Proton also publishes the code, so outsiders can check how the app behaves.
That makes the two halves different kinds of control. The rules are yours and you can see them working. The privacy promise rests on Proton, and the public code is what lets others hold it to that promise.
Give an AI assistant what you would give a new colleague in their first week: read access to the files the job needs, a rule to check with you before anything goes out in your name, and no power to delete.