Letting the assistant plan the week
The job
Deciding what happens when. With several clients running, Monday used to open with an hour of triage across boards, and the plan rarely survived to Wednesday. The assistant reads the full queue and proposes the week: what comes first, and which tasks to group so a day can be spent inside one client’s work instead of scattered across five.
The systems involved
The project management system, and only that. Every open task across every client board lives there, so the assistant sees the whole queue at once. It reads the tasks with their fields and writes back one thing, a proposed running order for the week. It moves nothing on its own.
How it works
The precondition is data quality. Every task carries a client, an estimate and a deadline; the ranking needs all of them. Deadlines set what must land this week, estimates decide what fits in a day, and the client field is what makes grouping possible. None of that is an AI problem. It is discipline in how tasks get written, and it had to exist before the assistant was any use here. Tasks missing a field come back first, as questions. Then the proposal: the week in day-blocks, deadline work pinned, the remainder gathered by client so the cost of switching context is paid between days. It stays a proposal, and the overrides are worth reading afterwards, because each one marks something the fields failed to capture.
Where this goes wrong
The ranking inherits whatever is in the fields. A task written tidily, with a clean estimate, ranks well; a large one written vaguely ranks badly, and the vagueness is usually a sign it was never broken down. So a queue left alone drifts toward the work that is easy to describe. The correction sits in the data rather than the ranking: anything large, vaguely written and untouched for weeks belongs out of the ordering and in a breakdown. And the veto stays human, exercised in the plan itself.
When it’s a poor fit
The fit is poor when the fields are thin; a queue without estimates and deadlines produces a plan that is confident and unfounded, and fixing the data is the real project. It is poor when one client dominates the week and the order is obvious. And it is poor in weeks governed from outside the system, by a launch or by travel, when the queue is background.
My experience
I let the assistant order the week and keep the veto, and the vetoes are where I find out what my own task data has been hiding.