It only suggests and asks
The assistant asks the author of a task what is missing to start it, reminds about tasks nobody has touched, answers @AI in chat and proposes task drafts and a plan for the month. It does not decide anything on its own: a draft becomes a task only when a person accepts it, and nothing reaches the calendar until someone reviews the plan and applies it.
Off until you switch it on
Everything starts switched off. Each function, each chat channel, each task status and each client is included deliberately, and a single task can be hidden from the assistant with one switch in the form. A project rule keeps the new tasks of a project away from it. The assistant starts working only after a person managing it accepts the data processing terms.
What it sees and what it never sees
For a task the assistant reads the title, description, type, status and deadline. It never sees internal notes, attachments, rates, quotes or settlements, and in chat it does not read time entries - the monthly planner gets only the total hours of finished tasks it compares with. Before anything is sent, email addresses, phone numbers, account numbers and things that look like passwords or keys are removed, and known people are replaced with labels. This is a heuristic, not full anonymisation: a password typed without any label can get through, so a task description is not a place for secrets.
About the work, not about people
The assistant writes about what was agreed or requested, not about who did it - its questions and drafts do not name people, and the estimate accuracy report has no people in it at all. It is there to keep tasks clear, not to assess the team.
Analyses and price
Every call to the language model is one analysis. The module includes a monthly number of analyses that depends on the size of the team, and when they run out you can buy extra packs. The module is sold outside the packages, and its price grows with every other paid module.