Can AI Build a Legal Staff Schedule? Labor-Law Compliance of Nine Language Models on Small-Business Weekly Scheduling

Small businesses build a staff schedule every week, and that schedule must comply with labor law: hour limits for workers aged 14 to 17, meal periods, overtime, and, in some cities, rest between shifts. Scheduling software now offers to build these schedules with AI, and many owners can simply ask a chatbot. This study measured how often current language models produce a schedule that follows every rule. A generator produced 592 fictional cafés, retail shops and quick-service restaurants at five levels of difficulty. Each was presented under three rule sets (federal law only, California, and a Seattle chain store) and in two input formats, giving 3,552 scenarios, of which 22.5% were made impossible by design. Nine models from five developers answered 17,478 scenarios in total. A constraint solver supplied the answer key, and every answer was graded by code. The models separated into three tiers. Claude Opus 5.5 (99.0%), Claude Sonnet 5.5 (97.5%), Claude Haiku 5.5 (91.9%) and DeepSeek V4.1 Flash (90.2%) returned a legal schedule in nearly every scenario where one existed and identified 97.9–100% of impossible scenarios. GPT-6.1 Sol (59.0%), GPT-6 Luna (46.7%) and gpt-oss-120b (19.6%) seldom broke labor law but frequently broke the business's own constraints, chiefly staff availability, or declared solvable scenarios impossible. Llama 4 Maverick (3.4%) and Gemini 3.5 Flash Lite (1.3%) broke a labor law in 34.0% and 31.2% of scenarios and returned a schedule for 81.8% and 83.2% of impossible scenarios rather than reporting that none could exist. Rule-breaking rose with difficulty for seven of nine models (odds ratio 1.60 per level). For four models, labor violations and premium pay were significantly more frequent under California and Seattle rules than under federal rules for the same business. Contrary to the pre-specified hypothesis, most models reliably recognized impossible scenarios. Giving a model a Python tool with a constraint solver reduced GPT-6 Luna's violation rate from 53.8% to 5.4% and Gemini 3.5 Flash Lite's from 86.5% to 13.5% on the same scenarios, although Gemini then failed to answer 21 of 24 impossible scenarios.