ECIS 2026 (2026) AI Processed
SCHNITZEL-PREDICTION: DESIGNING HUMAN–AI COLLABORATION FOR CAFETERIA DEMAND FORECASTING
This study reports on a nine-month action design research project that developed a collaborative human-AI forecasting system for corporate cafeteria operations. The researchers combined machine learning models with large language models to categorize dishes semantically, while integrating human override features to maintain chef autonomy.
Problem
Cafeterias generate substantial food waste due to inaccurate demand planning, which is traditionally treated as either a purely automated AI problem or an isolated human task. Traditional forecasting models struggle with the cold-start problem when new dishes are introduced, and automated models fail to adjust for local contextual exceptions known only to managers.
Outcome
- The algorithmic machine learning model achieved a 30% reduction in forecasting errors compared to a naive 7-day lag baseline.
- The hypothesis that the algorithmic model would statistically outperform experienced human planners across all menu items was not supported, as direct statistical comparisons showed no overall significant difference (p = 0.454).
- Experienced human planners statistically outperformed the algorithmic model for novel menu items with limited historical data (MAE of 10.46 for humans versus 17.00 for the algorithm), supporting the hypothesis that human judgment remains superior for high-uncertainty contextual forecasting.
- Acknowledgeable limitations of this research include its single-case design at a German firm, a small quantitative evaluation scope of 24 operating days, and qualitative insights derived from two think-aloud sessions.
- The hypothesis that the algorithmic model would statistically outperform experienced human planners across all menu items was not supported, as direct statistical comparisons showed no overall significant difference (p = 0.454).
- Experienced human planners statistically outperformed the algorithmic model for novel menu items with limited historical data (MAE of 10.46 for humans versus 17.00 for the algorithm), supporting the hypothesis that human judgment remains superior for high-uncertainty contextual forecasting.
- Acknowledgeable limitations of this research include its single-case design at a German firm, a small quantitative evaluation scope of 24 operating days, and qualitative insights derived from two think-aloud sessions.
What it means for you
- CIO / IT Executive: Issue a directive to your enterprise architecture team to halt all pure-automation forecasting initiatives and mandate that any future demand-planning software deployments integrate a 'human-in-the-loop' interface, specifically reserving override capabilities for experienced frontline staff in high-uncertainty scenarios.
- IT Manager: Create a task in your team's active development sprint to build an LLM-powered semantic tagging feature for your menu database, and pair it with a simple UI override slider on the kitchen manager's dashboard to let them manually adjust AI-generated forecasts for newly introduced dishes.
- Business Strategist: Review your cafeteria's historical food waste logs and restructure the kitchen managers' weekly KPIs to incentivize them to manually audit and adjust automated algorithmic forecasts, focusing their human intervention specifically on novel menu items where historical data is sparse.
- Researcher: Draft a grant proposal for a multi-site, longitudinal study that replicates the 'SCHNITZEL-PREDICTION' methodology across five different corporate catering firms over a six-month period to overcome the single-case and 24-day evaluation limitations of this study.
- Policymaker: Draft an update to the public sector sustainable procurement guidelines requiring that any government-funded cafeteria software tenders mandate human-AI collaborative interfaces rather than fully automated solutions to actively prevent food waste caused by algorithmic 'cold-start' forecasting failures.
Transcript
Host: Welcome to A.I.S. Insights - Turning IS Research into Business Action. I am your host, Aaron Ivan Sterling, and today we are discussing a study presented at the European Conference on Information Systems 2026. The study is titled "SCHNITZEL-PREDICTION: DESIGNING HUMAN–AI COLLABORATION FOR CAFETERIA DEMAND FORECASTING". Joining me to unpack these findings is our expert analyst, Ava Irene Solis. Welcome, Ava.
Expert: Thank you, Aaron. I am glad to be here to discuss this research.
Host: To start us off, what is the core focus of this study?
Expert: This study reports on a nine-month action design research project conducted at a German financial services association. The researchers developed a collaborative demand forecasting system for the firm's corporate cafeteria. What makes this study interesting for business practitioners is that it combines machine learning models with large language models to categorize dishes semantically, while integrating human override features to keep the cafeteria managers in control.
Host: Food waste is a major challenge for corporate cafeterias. Can you elaborate on the specific planning problems this study addresses?
Expert: Absolutely. In communal catering, cafeterias often prepare food in large batches under high demand uncertainty, which contributes to about twenty percent of prepared meals ending up as food waste. Traditionally, this planning is treated either as a purely automated algorithmic task or as an isolated human task relying on intuitive guessing. The issue is that standard machine learning models face a "cold-start" problem when a new dish is introduced because there is no historical sales data for it. On the flip side, fully automated models struggle to adapt to local, sudden contextual changes that only the local managers know about, like local construction noise that might drive employees away, or special on-site meetings.
Host: So the study looks at how we can get the best of both worlds. How did the researchers design their approach to solve this?
Expert: They developed a web-based forecasting system called "Schnitzel-Prediction." The core prediction engine is built on an XGBoost machine learning model. To handle the cold-start problem, they set up a semantic pipeline using Google's Gemini 2.0 Flash model. The large language model analyzes the name and description of a new dish and translates it into structured binary flags, such as whether the dish contains meat, is vegan, or is a stew. This allows the forecasting model to transfer historical demand patterns from similar dishes. Finally, the system includes a web interface that allows the manager to review the prediction and manually adjust it based on their local knowledge.
Host: Let's talk about the outcomes. How did this collaborative system perform in practice?
Expert: The evaluation showed distinct strengths for both the technology and the human planners. First, the machine learning model achieved a thirty percent reduction in forecasting errors compared to a naive seven-day lag baseline, which simply assumes today's demand will match the demand of the same day last week.
Host: That sounds like a solid improvement over the baseline. But how did the algorithm compare to experienced human planners?
Expert: This is where we see the value of human-AI collaboration. The hypothesis that the algorithmic model would statistically outperform experienced human planners across all menu items was actually not supported. The direct statistical comparison showed no overall significant difference, with a p-value of zero-point-four-five-four.
Host: That is an important finding. It suggests the algorithm is not a complete replacement for human expertise.
Expert: Exactly. In fact, for novel menu items with limited historical data, the experienced human planners statistically outperformed the algorithmic model. For those new dishes, the humans achieved a Mean Absolute Error of ten-point-four-six, compared to seventeen-point-zero-zero for the algorithm. This supports the idea that human judgment remains superior for high-uncertainty contextual forecasting.
Host: Are there any specific limitations to these findings that practitioners should keep in mind?
Expert: Yes, there are some boundaries. The research was a single-case design at a German firm, so the specific local workflows and culture heavily influenced the design. Additionally, the quantitative evaluation was conducted over a limited scope of twenty-four operating days, and the qualitative insights were derived from two think-aloud sessions with internal stakeholders.
Host: Looking at the big picture, what are the actionable takeaways for business leaders and technology developers?
Expert: For managers in operational contexts like hospitality, retail, or logistics, the key takeaway is to avoid complete automation of complex tasks where human local context is highly valuable. Instead, seek out or co-design systems that respect human expertise through override mechanisms. For information systems developers, the study provides a blueprint showing that success is often found in the interfaces connecting the algorithm to the human rather than just building more complex models. Specifically, we can use large language models as feature engineers to help traditional models understand novel data, and we should implement attributable dashboards that store and visualize both the machine's predictions and the human's overrides separately. This helps teams analyze over time when to trust the algorithm and when to rely on human intuition.
Host: A clear reminder that technology is most effective when it supports, rather than replaces, human capability. Thank you, Ava, for sharing these insights today.
Expert: Thank you, Aaron.
Host: And thank you to our listeners. We hope these research-backed insights help you take practical action in your organization. Join us next time on A.I.S. Insights.
Expert: Thank you, Aaron. I am glad to be here to discuss this research.
Host: To start us off, what is the core focus of this study?
Expert: This study reports on a nine-month action design research project conducted at a German financial services association. The researchers developed a collaborative demand forecasting system for the firm's corporate cafeteria. What makes this study interesting for business practitioners is that it combines machine learning models with large language models to categorize dishes semantically, while integrating human override features to keep the cafeteria managers in control.
Host: Food waste is a major challenge for corporate cafeterias. Can you elaborate on the specific planning problems this study addresses?
Expert: Absolutely. In communal catering, cafeterias often prepare food in large batches under high demand uncertainty, which contributes to about twenty percent of prepared meals ending up as food waste. Traditionally, this planning is treated either as a purely automated algorithmic task or as an isolated human task relying on intuitive guessing. The issue is that standard machine learning models face a "cold-start" problem when a new dish is introduced because there is no historical sales data for it. On the flip side, fully automated models struggle to adapt to local, sudden contextual changes that only the local managers know about, like local construction noise that might drive employees away, or special on-site meetings.
Host: So the study looks at how we can get the best of both worlds. How did the researchers design their approach to solve this?
Expert: They developed a web-based forecasting system called "Schnitzel-Prediction." The core prediction engine is built on an XGBoost machine learning model. To handle the cold-start problem, they set up a semantic pipeline using Google's Gemini 2.0 Flash model. The large language model analyzes the name and description of a new dish and translates it into structured binary flags, such as whether the dish contains meat, is vegan, or is a stew. This allows the forecasting model to transfer historical demand patterns from similar dishes. Finally, the system includes a web interface that allows the manager to review the prediction and manually adjust it based on their local knowledge.
Host: Let's talk about the outcomes. How did this collaborative system perform in practice?
Expert: The evaluation showed distinct strengths for both the technology and the human planners. First, the machine learning model achieved a thirty percent reduction in forecasting errors compared to a naive seven-day lag baseline, which simply assumes today's demand will match the demand of the same day last week.
Host: That sounds like a solid improvement over the baseline. But how did the algorithm compare to experienced human planners?
Expert: This is where we see the value of human-AI collaboration. The hypothesis that the algorithmic model would statistically outperform experienced human planners across all menu items was actually not supported. The direct statistical comparison showed no overall significant difference, with a p-value of zero-point-four-five-four.
Host: That is an important finding. It suggests the algorithm is not a complete replacement for human expertise.
Expert: Exactly. In fact, for novel menu items with limited historical data, the experienced human planners statistically outperformed the algorithmic model. For those new dishes, the humans achieved a Mean Absolute Error of ten-point-four-six, compared to seventeen-point-zero-zero for the algorithm. This supports the idea that human judgment remains superior for high-uncertainty contextual forecasting.
Host: Are there any specific limitations to these findings that practitioners should keep in mind?
Expert: Yes, there are some boundaries. The research was a single-case design at a German firm, so the specific local workflows and culture heavily influenced the design. Additionally, the quantitative evaluation was conducted over a limited scope of twenty-four operating days, and the qualitative insights were derived from two think-aloud sessions with internal stakeholders.
Host: Looking at the big picture, what are the actionable takeaways for business leaders and technology developers?
Expert: For managers in operational contexts like hospitality, retail, or logistics, the key takeaway is to avoid complete automation of complex tasks where human local context is highly valuable. Instead, seek out or co-design systems that respect human expertise through override mechanisms. For information systems developers, the study provides a blueprint showing that success is often found in the interfaces connecting the algorithm to the human rather than just building more complex models. Specifically, we can use large language models as feature engineers to help traditional models understand novel data, and we should implement attributable dashboards that store and visualize both the machine's predictions and the human's overrides separately. This helps teams analyze over time when to trust the algorithm and when to rely on human intuition.
Host: A clear reminder that technology is most effective when it supports, rather than replaces, human capability. Thank you, Ava, for sharing these insights today.
Expert: Thank you, Aaron.
Host: And thank you to our listeners. We hope these research-backed insights help you take practical action in your organization. Join us next time on A.I.S. Insights.