What “AI scheduling” actually means

Every scheduling tool sells AI now. Very little of it is.

This is worth pulling apart, because the distinction is not pedantry — it determines whether you can check the software's work, whether it behaves the same way twice, and what happens when it gets something wrong.

Three different things wearing one word

When a scheduling product says AI, it is almost always one of these:

1. A constraint solver. Rules filter who can work a shift; a scoring function picks between whoever is left. Entirely deterministic.
2. A statistical forecast. Historical averages, usually weighted toward recent weeks. Also deterministic.
3. A language model. Genuinely generative, genuinely non-deterministic.

The first two are the ones that actually build your schedule. Neither involves a language model, and neither has for the decades they have existed under the name "optimisation".

Determinism is the feature

There is a reason serious scheduling engines are deterministic: the same inputs must produce the same schedule.

If they do not, you cannot reproduce a result to investigate a complaint. You cannot tell whether last week's odd assignment was a bug or a coin flip. You cannot A/B a weight change, because you cannot hold anything else constant.

A language model in that position is a liability, not a feature. Asked to build a roster twice, it gives you two rosters. Asked why, it produces a fluent explanation that may or may not describe what it actually did — because a generated justification is not a trace.

Sofia's Auto-Fill is deliberately a scorer, not a model. Hard constraints run first: availability, qualification, approved time off, overlaps, minimum rest, daily and weekly hour caps, minor-labor limits, consecutive-day limits. Only then does weighted scoring choose between the people who remain. Preferences never override rules, because preferences are not consulted until the rules have already filtered.

The weights are published. Every assignment carries its score, the factors that produced it, and the alternatives that were considered. That is not a generated narrative — it is the arithmetic, shown.

What to ask a vendor

Four questions separate the categories quickly:

• "If I run this twice on the same data, do I get the same schedule?" If not, ask what you are supposed to do when it is wrong.
• "Show me why this person got this shift." A real answer names factors and their contributions. A generated one reads well and cites nothing.
• "What are the weights?" If they cannot be stated, they cannot be tuned to your business.
• "Can I undo it?" An automated decision you cannot reverse is one you have to review before accepting, which removes most of the benefit.

Where a language model does earn its place

None of this means the technology is useless here. It means it belongs in a different job.

Asking a question in plain language and getting a straight answer about your own data — how many shifts are open, what the week costs, who is close to overtime — is exactly what language models are good at. The work is comprehension, not optimisation.

Two conditions make it safe. It has to be grounded: answering from your real figures, supplied with the question, rather than from what the model has absorbed. And it should be read-only: reporting on the schedule rather than changing it. A wrong sentence is a misunderstanding you can correct. A wrong mutation is a shift somebody does not know about.

That is the whole of Sofia's use of AI: one assistant, manager-gated, grounded in the current week, which cannot touch your schedule. When the model is unavailable it falls back to computed answers and says so in the message, so you always know which you are reading.

The honest version

Most scheduling AI is a solver with a marketing budget. Solvers are good — better than a language model for this job. The problem is not the technology, it is the label, because the label obscures the questions you should be asking.

Ask whether it is reproducible. Ask to see the reasoning. Ask what the weights are. Ask if you can undo it.

A tool that answers those four questions well is worth more than one that calls itself intelligent.