How should AI help with scheduling and dispatch?

Scheduling looks like an optimisation problem and is mostly a constraint problem with a policy question hidden inside it. The constraints are stateable: who is qualified for what, travel time between locations, working hours, equipment availability, appointment durations, commitments already made. Given those, proposing an allocation is genuinely useful work and largely mechanical.

The hidden question is what to do when the constraints conflict, which they do daily. Two jobs need the same person, an emergency arrives into a full day, a customer wants a slot that would strand someone across town. Resolving these is a judgement about which customer relationship, which margin and which promise matters more, and it is a policy the business holds whether or not it has written it down. A system that resolves them without being told will resolve them consistently and invisibly, which means the business has a scheduling policy it cannot state.

So the productive arrangement is a proposal plus the conflicts named. Here is an allocation satisfying the constraints, here are the three cases where something had to give, here is what each option costs. A dispatcher then decides in seconds with the trade-off visible, rather than either doing the whole allocation by hand or accepting a schedule whose compromises are unlabelled.

The genuinely valuable second application is the reshuffle. Schedules break — a job overruns, a van breaks down, someone is ill — and rebuilding the day under pressure is where errors and unhappy customers come from. Producing a revised allocation quickly, with the affected customers listed and the messages drafted, addresses the part of the problem that is hardest under time pressure.

Two constraints are consistently underestimated in these systems and are worth stating explicitly. Travel time is not distance and varies by time of day, so a schedule built without real travel estimates will be wrong in the direction of over-optimism every time. And the durations in your system are what someone typed when the job was booked, not what the job takes, and a scheduler working from those inherits every optimistic estimate.

The customer-facing half deserves its own decision. Confirmations, reminders and change notifications are assembled from records and are safe to send automatically; a message that offers a new time or explains a delay is making a commitment and belongs behind a person, at least until the business has watched what it says for a while.