Handing appliance scheduling to AI: it captures 96.7–98% of the savings — and still needs a validator
Five LLMs acting as home energy managers: cheap is not the same as safe or feasible, models both invent schedules and explain a constraint correctly while committing an invalid one.
Key numbers
- System: a tool-calling ReAct agent converts natural-language requests into half-hourly schedules for multiple household appliances, using dynamic tariffs, weather forecasts, PV generation estimates and household demand data.
- Evaluation: five LLM backends benchmarked against a mixed-integer linear programming oracle across dynamic tariffs, constraint conflicts, weather-aware scheduling and a seven-day rolling deployment.
- Results: native function calling achieved high scheduling success and near-optimal cost on ordinary tariff days while text-parsed actions reduced reliability; low cost did not guarantee safe or feasible behaviour — Claude Sonnet 4.6 performed best in power-cap and infeasibility scenarios, and Qwen-3 exceeded GPT-4o-mini on overall constraint compliance.
- Failure modes: fabricated schedules, failed commitments, and reasoning-to-action failures where the model explains a deadline correctly but commits an invalid schedule.
- Value captured: across the seven-day window the agents captured 96.7%–98.0% of the savings available between an off-peak timer and the MILP oracle, beating rule-based policies — with the authors insisting every committed schedule pass an independent deterministic feasibility validator.
Method and sample
arXiv paper “LLMs for Agentic Home Energy Management” (2026-07-06): a tool-calling ReAct agent converts natural language into appliance schedules, benchmarked against MILP across dynamic tariffs, constraint conflicts, weather-aware scheduling and a seven-day rolling deployment, with failure modes analysed.
What it means for brands
In home energy, appliances and home services, AI is shifting from answering questions to executing actions. That raises the bar: being right is not enough, it has to be safe and feasible — so the moat is machine-verifiable specifications and service constraints.
How SourceSight handles it
For home and appliance projects SourceSight uses purchase and usage prompts (which to choose, pricing, what home size, power draw, warranty) to see how AI describes you, while the content factory structures specs, energy use and installation conditions for exactly these agentic workflows.
Limits
The experiment uses one tariff mechanism and a seven-day window, so the sample is narrow; tariffs and equipment differ widely by region, making 96.7–98% a result of those conditions rather than a general guarantee.
References
Run it on your own project
What the same prompts return across 10 engines is only knowable by running them.
Related research
Cross-category recommendations: +36% add-to-cart on item pages, +15% on cart pages
Recommending general merchandise to grocery shoppers — a milk frother with the milk — has long been hard; LLM retrieval plus real-time cart context moved production A/B metrics.
Industries · 2026-10-07Home services: 22% of homeowners already use AI to find a provider, and 80% of businesses don't know how to be found
The local services funnel is moving upstream: before a homeowner opens a website or calls, AI has already synthesised reviews, reputation and service descriptions into one answer.