The AI scheduling agent

What it is

A chat panel that sits beside your schedule and can do the things you can do. Not a wrapper that writes text for you to copy — it performs the actual operations, through exactly the same permission, validation and audit path as the buttons.

Two consequences follow, and they are the whole design:

• It cannot exceed your permissions. An employee talking to the assistant gets the employee tool set. A manager gets the manager one. There is no privileged AI account.
• Everything it does is audited the same way your own clicks are, attributed to you, with the fact that it came through the assistant recorded.

It is included on every paid plan rather than reserved for the top tier, because it is the reason to use this product at all.

How a week gets drafted

Ask for a week and three things happen in order.

1. It gathers context. Your roster and positions, stated availability, approved time-off, skills and their expiry dates, and — importantly — how the load actually fell over the previous six weeks: who took the weekends, who took the late finishes, who went over their cap, who got the biggest share.

2. It proposes. The agent decides the assignments; the proposal step enriches and validates them. You get a complete week presented for review, with the conflicts it could not resolve called out rather than hidden.

3. You apply it. Nothing is written until you say so. Applying puts it into the draft week — still invisible to your team until you publish.

The six-week fairness window is what separates this from a solver that balances one week in isolation. A week that is perfectly fair on its own can still be the fourth Saturday in a row for the same person, and that is the version staff quit over.

Talking to it well

It reads what is on your screen — which week you are looking at, which schedule, which view — so you rarely need to restate that. "Give Tom Friday off" means this Friday, in this week.

Things that make a first draft land:

• State coverage, not assignments. "Two on the floor at all times, three from 6pm Friday and Saturday" beats naming people — naming people is the part it is good at.
• Say the constraints out loud even if they are already in the system. It costs you nothing and removes the ambiguity.
• Correct it conversationally. "Too many hours on Ana, move one of hers to whoever is light" works, and is faster than editing the grid.

The permanent constraints belong in the system rather than in every prompt: put recurring limits in compliance rules, qualifications in skills, and personal patterns in availability. Then the agent honours them every week without being told.

Beyond drafting

Managers can also ask it to:

• Fill the gaps in a week you built by hand — same eligibility checks, applied only to the empty slots.
• Explain an assignment — why this person on this shift, with the actual reasons cited: availability, skill match, cost, or load balancing against recent weeks.
• Check the week before publishing and summarise what would fail.
• Answer questions about your own data — who is closest to overtime, who has not acknowledged next week, which certifications expire this month.

Employees get a smaller, self-service set:

• "What am I working this week?"
• "Request Thursday and Friday off."
• "Who can take my Saturday?" then offer the swap.
• "Clock me in." / "Start my break."
• "Set me unavailable on Monday mornings."

A manager is also somebody's employee, so there is a perspective switch — hold the conversation as yourself-the-staff- member, with its own history. An Employee cannot switch the other way.

Limits worth knowing

It reasons over your organization's data and nothing else. It does not know your local employment law — that is what compliance rules encode — and it has no forecast of next month's demand.

It will tell you when it cannot do something rather than improvising. A week it cannot fill comes back with the holes named and the reason each one is a hole.

Conversations are stored per organization and per perspective, so you can pick up where you left off; the drafting work happens against your data on our servers and is not used to train models.