AI-Powered Sales Forecasting & Labor Optimization for the Restaurant Industry
Transforming day-part demand visibility through machine learning and conversational AI
In the highly competitive quick-service restaurant (QSR) and food service industry, the margin between operational excellence and costly inefficiency often comes down to one question: how accurately can an organization anticipate customer demand? Staffing decisions, food preparation volumes, and inventory procurement are all downstream of that single forecast. Yet for many restaurant operators, that forecast remains stubbornly imprecise. It’s too broad in its time horizon and too unreliable to drive confident, proactive action.
This use case examines how Smartbridge partnered with a food service client to rapidly prototype a machine learning-powered sales forecasting and labor optimization solution during an intensive one-day hackathon. Built entirely within Microsoft Fabric and surfaced through a Power BI dashboard and conversational AI agent, the solution demonstrates that granular, day-part-level demand forecasting is not only achievable, but it is achievable quickly, with modern cloud-native tooling and the right data foundations.
The result is a working proof of concept that extends beyond passive reporting into proactive, alert-driven intelligence (notifying store managers of demand-staffing misalignments before they become operational problems).
The Challenge: Flying Blind in a Demand-Driven Industry
The client’s existing forecasting infrastructure centered on a legacy system that produced high-level daily sales projections. While the data existed, confidence in it was low. Frontline teams and managers had largely stopped relying on it, defaulting instead to paper-based records updated every 30 minutes. This was a reactive, labor-intensive workaround that introduced further delays and inconsistencies.
The operational consequences were tangible and recurring:
The core gap was not a lack of data. The client had three years of historical sales and labor records. The gap was in how that data was being used. Daily aggregates obscured the variation within a day, and without day-part granularity, no forecast could meaningfully guide the staffing and prep decisions that play out hour by hour across a restaurant’s operations.
The Solution: Day-Part Intelligence, Built in a Day
Smartbridge designed and built a machine learning solution that restructures the forecasting problem around five operationally meaningful day parts: breakfast (opening through mid-morning), mid-morning, lunch, afternoon, and dinner/evening. Rather than asking “what will this store sell today,” the model asks “what will this store sell during lunch?” — a question whose answer is far more actionable for a store manager making staffing and prep decisions.
Data Inputs
The predictive models were trained on a multi-source data set that included:
Technology Platform
The entire solution was developed within Microsoft Fabric, which served as the unified data engineering and analytics platform. Fabric’s integrated environment allowed the team to move rapidly from raw data ingestion to model training to visualization without context-switching across disconnected tools. GitHub Copilot was used as an AI-assisted development accelerator throughout the notebook development process, enabling the team to produce, review, and iterate on code substantially faster than traditional development cycles would allow.
Outputs were delivered through two surfaces: a Power BI dashboard for visual analytics and trend monitoring, and a conversational AI agent that allows store managers and operations teams to query forecast data in natural language. Basically, asking questions like a colleague rather than running reports like an analyst.
Forecast Accuracy
Model performance was validated against a held-out test data set. The R² score — a measure of how closely predicted values track actual values, where 1.0 represents a perfect fit — ranged from 0.85 to 0.90. Translated to practical terms, the model explained 85 to 90 percent of the variance in sales outcomes, a meaningful improvement over a baseline forecast that provided no day-part differentiation at all.
Conversational Agent Demonstration
During the hackathon showcase, the team demonstrated the conversational agent using a representative store and a forward-looking query date. A store manager persona was used to illustrate realistic, role-appropriate use cases:
The agent’s ability to return granular, day-part-level labor guidance from a conversational query rather than requiring a manager to navigate a reporting interface represents a meaningful shift in how operational intelligence is accessed and applied.
Proactive Intelligence: The Next Layer
Beyond the reactive query-and-response model, Smartbridge also built a proactive alerting agent as a stretch deliverable and completed it within the hackathon timeframe. Where the conversational agent answers questions, the proactive agent surfaces insights unprompted. It monitors forecasted sales against scheduled labor and sends alerts when misalignments are detected: for example, notifying a manager that current staffing levels fall short of what the lunch forecast suggests will be needed, or that scheduled labor for a slow afternoon period is tracking above projected demand.
This shift from reactive to proactive represents an important evolution in how AI is applied in operational settings. Rather than requiring managers to know what to ask, the system monitors on their behalf and flags what matters, reducing cognitive load and increasing the likelihood that forecast intelligence actually reaches the decisions it is meant to inform.
Business Impact
The operational value unlocked by day-part forecasting accuracy maps directly to the cost and experience drivers that matter most in QSR environments:
| Impact Area | Operational Outcome |
|---|---|
| Labor Optimization | Align staffing to forecasted demand by day part, reducing overstaffing in slow periods and understaffing during peaks without requiring reactive same-day adjustments. |
| Customer Experience | Reduce wait times by ensuring adequate staffing and food preparation capacity is in place before demand arrives, not in response to it. |
| Food Waste Reduction | Calibrate prep quantities to anticipated demand by time of day, reducing the risk of over-production during slow periods. |
| Location-Aware Planning | Account for store-specific demand patterns driven by proximity to offices, event venues, or high-traffic retail-enabling each location to plan against its own demand profile rather than a network average. |
| Manager Decision Support | Surface actionable intelligence through natural language, reducing reliance on manual reporting and enabling frontline leaders to act on forecast data directly within their workflow. |
Stretch Goal: Proactive Teams Agent for Labor Hour Prediction
As a stretch deliverable, the Smartbridge team built and successfully activated a Microsoft Teams Agent capable of predicting labor hours from sales inputs. The agent was running live on dummy data by the end of the hackathon, with integration into the Lakehouse environment as the defined next step. The model accepts store number, date, DayPart, and sales as inputs, applies ML-trained coefficients derived from date-related features and historical patterns, and outputs a predicted labor hours figure. Smart notifications are delivered via Teams to store managers based on forecast outputs and labor alerts — closing the loop from forecast intelligence to frontline action without requiring managers to navigate a separate reporting interface.
Two future-state architecture options were designed during the hackathon for scaling the proactive agent: Option 1 uses Microsoft Copilot Studio to build a Copilot-native agent with Teams integration; Option 2 leverages Microsoft Fabric’s native Data Agent capability combined with Azure OpenAI to generate SQL, explain data, and trigger email notifications to store managers via the Microsoft Graph API Mail.Send permission. Both options preserve the existing OneLake data foundation and require no additional data infrastructure investment.
From Forecast to Action
The QSR and food service industry operates on tight margins, high transaction volumes, and a customer experience that is acutely sensitive to operational rhythm. When staffing and preparation align with demand, the result is fast service, minimal waste, and an engaged workforce. When they don’t, the consequences — long wait times, food waste, unnecessary labor cost, and associate strain — accumulate quickly.
What Smartbridge demonstrated in this engagement is that the path from high-level, unreliable daily forecasts to granular, trusted, day-part-level intelligence is shorter than many organizations assume. Built in a single day using Microsoft Fabric, machine learning, and a conversational AI interface, the prototype achieved forecasting accuracy and delivered that accuracy through a user experience that requires no technical expertise to operate.
The addition of a proactive alerting agent extends the value further still, shifting the paradigm from “a tool that answers questions” to “a system that monitors operations and surfaces what needs attention.” For a store manager focused on the floor, that distinction matters enormously.
For restaurant and QSR operators evaluating their forecasting capabilities, the relevant question is no longer whether machine learning can improve demand visibility. It can, at meaningful accuracy levels, using data that most mature operators already possess. The question is how quickly that capability can be operationalized, and what it will take to connect forecast intelligence to the frontline decisions that determine whether a shift runs smoothly or not.
Smartbridge’s hackathon prototype offers a direct, demonstrated answer to that question and a clear foundation for what a production-grade implementation can achieve. To learn more about how this can be applied at your organization, book some time on our calendars and let’s start a conversation.
