An AI Scheduling Assistant for Window Cleaning That Never Changes Anything Without Asking

What can an AI scheduling assistant do?

The fiddly operations an office does many times a day, asked for in plain words. In the live system the office's assistant can:

• move a visit to another day or time
• top up a van that is finishing short
• change which postcodes a van covers
• add a cleaner or change their account
• act on a single visit

It also answers questions: how does today look, which vans are under target, who has not been assigned a van, what is still unscheduled. It follows the office around the app, so it knows which screen they are on.

Why must a human confirm every change?

Because a confident wrong change to a live round is expensive, and an AI will occasionally be confidently wrong. So the assistant never writes anything directly. It shows exactly what it intends to do, for example the customer, the old slot and the new one, and waits.

The guard rails that make it trustworthy:

• Proposals are held server-side and only an explicit Confirm executes them.
• Rejecting does nothing, and the proposal stays visible so you can see what it would have done.
• Proposals expire after about fifteen minutes; a quarter-hour-old instruction should be asked again.
• Failures are reported honestly. If Jobber refuses, you get the reason, not a success message.

Should crews have an assistant too?

Yes, but a narrower one. A cleaner's assistant only sees their own van's day. It can tell them what is left, how they are doing against target, whether they are on for a bonus and how the last hour went.

It deliberately cannot mark a job completed or missed. Those go through the buttons on the job card, because the cleaner's location is recorded at the moment they press, and that is what protects them if anyone asks whether they were at the house. Telling the cleaner to use the button is the assistant working as designed.

What about questions about how the system works?

Answer them from the same guide the office reads. The live system has an in-app guide, written per screen, and a help bot that answers only from it. One source with two readers means the bot and the guide can never disagree. The help bot also refuses questions the guide does not cover, rather than answering off-topic questions confidently, because nobody can tell a confident wrong answer from a right one.

How should you introduce AI into scheduling?

1. Recommendation mode first. The assistant suggests, a person approves every change. This is where the live system started.
2. Semi-automatic next. Low-risk, obvious changes can run on their own while exceptions are flagged.
3. Automatic only for proven rules. The end-of-day roll-over runs unattended because its rule was proven on real days first. See missed visits.

Frequently asked questions