AI shift planning for events: what the model proposes, what rules check and what a person decides

AI shift planning at events is filling open shifts with a proposal from a language model after rules have filtered the candidates by qualification, availability, rest period and statutory limits. The model names a person, the reason and a confidence; the planner books. What vendor brochures call "AI plans your shifts" is three different things, and whether a tool is any good depends on whether it keeps them apart.

Three layers that have to stay separate

Layer	What it does	Example	Who decides
Rules	Check every shift deterministically, same result for the same input	Rest period 11 hours, Minijob limit, 70 days, qualification, availability	Nobody; the result is fixed
Model	Weighs the rest that rules cannot decide and proposes	"M. Keller: bar licence, free on Saturday, 41 of 70 days, confidence 92 %"	The model proposes
Person	Accepts or discards, books, approves	Accept the card, the shift is filled	The planner

A tool that mixes the three layers has two typical failures. It lets the model calculate what a formula does better and then produces rest-period violations with 90 percent confidence. Or it books without a click, and nobody can say later why the person was in that shift.

What is checked before the model

Five checks have an unambiguous answer and therefore belong in rules, not in a model:

1. Qualification. The shift requires a bar licence, a forklift licence or first-aid training; the person has it or not. That is a tag on the profile.
2. Availability. The person named the day in the application or not.
3. Rest period. Between the end of the same person's last shift and the start of the next there must be 11 hours, § 5 ArbZG, 10 in event operations with compensation. That is a subtraction. What applies to build, show day and strike is in Rest periods at events.
4. Minijob limit. The planned pay of the month against 603 euros. A multiplication and a comparison.
5. 70-day limit. Worked plus planned days of the calendar year against 70, § 8 (1) no. 2 SGB IV. A count across all events, see Short-term employment.

Out of 200 people in the pool, perhaps twelve remain for a Saturday bar shift after these five checks. Only now is a model worth it.

What the model adds

Twelve candidates who are all allowed: who fits best? Here a language model helps, because the criteria are soft. Who wrote "preferably in a team with Lisa" in the application? Who did the same bar last year? Who has used 41 of 70 days and who 68? Who already has four night shifts in a row? The model reads these signals and proposes one person, with two sentences of reasoning and a confidence. The reasoning is the actual value: it makes the proposal checkable. A proposal without a reason is a die with a user interface.

Two further tasks suit a model because they read across a whole list instead of filling one shift:

• Same person, three spellings. In the applicant list and across the lists of all contractors in accreditation, the model recognises typos, double names and swapped fields and flags what belongs together. Merging is a click.
• Check of the monthly close. Rules calculate markers per time entry: night work, missing break, overlap, deviation of start and duration from the plan. The model reads which combinations are payroll-relevant and would rather name three solid findings than twelve weak ones. Every finding shows its evidence.

What a person decides, and why that is not a formality

A shift booking has legal consequences: it creates working time, pay and social-insurance contributions. Art. 22 GDPR gives every person the right not to be subject to a decision based solely on automated processing that produces legal effects. A shift planner that books without confirmation moves into that territory. A shift planner that proposes and waits for the click does not. The difference is not the technology but the click, and it has to be deliberate: an "accept all proposals" button is the same automation with a detour.

The same applies to the month-close check. The model names findings; approving the hours stays with HR. And for the look-alikes: the model flags the pair, whether it is one person is decided by someone who knows them.

Data protection: what the model sees and where it runs

Question	The answer you should hear
Where does the model run?	In the EU, with the region named. A US endpoint is a transfer to a third country.
Are inputs used for training?	No, contractually guaranteed. For models via a cloud provider, that is in the provider's terms.
Does the model see real names?	For assignment proposals yes, otherwise the proposal cannot be checked. For the month-close check no: people reach the model as references.
Is there a data processing agreement?	Yes, Art. 28 GDPR, with the model provider as sub-processor.
Can I switch the AI off?	Yes, per function, without the shift plan ceasing to work.

The last row is the test. A shift planner that no longer plans without the model has moved the rules into the model.

Checklist: choosing an AI shift planner

Criterion	How to recognise it
Rules before the model	Rest period, Minijob, 70 days are flagged in the plan without AI too
Reason per proposal	Two sentences and a confidence, not a list without a reason
No booking without a click	Proposal as a card, accepting is an action
EU region and no training	Named in the privacy policy and the data processing agreement
Pseudonymisation where possible	Month-close check without real names
Switchable per function	A switch in the settings, not a support ticket
Price without an AI module	Included in the platform fee or transparent per call
Finding nothing is a result	The model would rather report nothing than weak findings

Whoever ticks these eight points does not get a tool that makes the shift plan. They get one that replaces the look across 200 names and flags the error before the person stands at the gate. That is the part of shift planning that used to happen at night.