Eat Mor Chikin, and fast? The story of Chick-fil-A’s multimodal data analysis and optimization

By Data Duets on
Image courtesy of Imago / Pubity Illustration - pubity.com

Introduction to the business case

The Wall Street Journal article reveals how Chick-fil-A is using innovative data collection and analysis methods to optimize its drive-through operations. The company has developed a "Film Studies" unit that combines drone footage with security camera data to create comprehensive "game films" of its restaurant operations. This multimodal data collection approach, inspired by NFL game analysis, allows the company to model traffic patterns, identify operational bottlenecks, and analyze service efficiency in drive-through operations. Read the article here.

The data-centric insights have led to significant operational improvements, including the development of new restaurant designs with elevated kitchens and multiple drive-through lanes capable of serving 700 cars per hour. The analysis also helped optimize staffing patterns, identify gaps in Wi-Fi coverage, and improve order processing workflows. One location reported a 50% increase in drive-through transactions after implementing the changes.

In this post, we'll answer the following question:

What's interesting about this case, and what does it say about the future of data collection, analysis, and data-centric optimization?

Academic's take

This case is really about data-centric optimization, so I'll leave most of the discussion to the Director, whose background is in management science and operations research. What I find most fascinating about this case is the novelty and multimodal nature of the data collection. With recent advances in deep learning, object recognition is now making its way into retail. This opens up a new venue for experimentation. This is interesting. Previously unthinkable data for a fast food chain can now be collected using drones, as detailed in the case, and combined with in-store camera data and data from transaction processing systems. The story does not go into the details of the modeling, but it does point out how data from different modalities complemented each other to solve a decades-old optimization problem: how to get customers through the drive-through quickly and keep them happy.

Chick-fil-A's combination of transactional data with video and other data to model and optimize the cars going through the restaurant's drive-through reminds me of the growing interest in multimodal applications. While the case does not hint at the methods used, I'd like to add a note on multimodal AI.

Multimodal AI: The future of data collection and analysis?

Multimodal AI describes machine learning models that are capable of processing and integrating information from multiple data modalities. These data modalities can include text, images, audio, video, and other forms of input. In contrast to traditional models, which are typically designed to process a single type of data, multimodal AI combines and analyzes different forms of data input to achieve a more comprehensive understanding of the problem.

Liang et al. (2022) describe three characteristics of multimodal AI: heterogeneity, connections, and interactions. Heterogeneity refers to the different properties, structures, and representations of modalities. A textual description of an event will differ fundamentally in quality, structure, and representation from a video of the same event. Connectivity refers to the complementary information shared between different modalities, which can be reflected in similarity scores. Finally, interactions refer to how different modalities interact when brought together.

With that as a side note, I will leave the discussion of the optimization problem to the next section.

Director's cut 

In my opinion, the Chick-fil-A case and success story illustrates how to approach process optimization. There are several lessons to be learned. The first is relatively obvious: Optimization efforts generally require a focus on the entire system rather than its individual parts. The second is somewhat hidden between the lines: systems are complex, so optimization efforts must be coupled with experimentation. Let's take a closer look at each lesson.

Lesson one: Focus on the whole system, not the parts

Chick-fil-A realized early on that eliminating one bottleneck would only create another. When the ordering process is optimized, the kitchen becomes overloaded. When the kitchen capacity is improved (along with any layout changes to accommodate larger vent hoods), the teams that handle the bagging and delivery of orders are overwhelmed. If the bagging and delivery operations and stations are optimized, customers eating in their cars in the parking lot at that time would leave little room for walk-in customers. Walk-in demand would decrease, and the overall productivity of the system would decrease again.

Process optimization, when done right, requires a focus on the entire system rather than its individual parts. Optimizing individual operations independently fails to consider downstream effects and often results in suboptimal overall performance. A holistic (systems) approach ensures that improvements are consistent throughout the process and do not maximize the performance of one operation at the expense of another.

In Chick-fil-A's case, data collection efforts are not limited to drive-through lanes. The team not only monitors what's happening outside the restaurant, but also in the kitchens to understand the impact on the overall system. It appears that solutions are not siloed, and the downstream effects are also being analyzed.

Lesson two: Couple optimization with experimentation

Chick-fil-A pays for its franchisees' remodeling costs to experiment with different design ideas. Why does Chick-fil-A need to experiment in addition to all this data? Wouldn't simulations help determine which design would produce the best results?

Simulation models are only as good as their underlying assumptions. They may fail to mimic behaviors that may emerge over time and affect actual system performance, especially if the process has manual components. Humans are extremely creative and adaptive, organically finding better and more efficient ways to achieve the same goal. When bottlenecks occur, frontline workers may find workarounds. Experiments would reveal how people adapt and interact with processes to overcome system limitations. These adaptations are difficult to predict in simulations.

There is an additional benefit to using experimentation as part of the optimization process. Each franchisee has local insight into their unique market characteristics and customer base. As a result, the effectiveness of solutions and customization approaches may vary from market to market. By encouraging franchisees to test different designs, Chick-fil-A taps into the collective intelligence of its entire network and gathers a greater variety of ideas.

This is a great case of how process optimization, when coupled with experimentation, can provide a more complete understanding of system behavior and reveal insights that simulations alone cannot capture.

References

  • Liang, P. P., Zadeh, A., & Morency, L. P. (2022). Foundations and trends in multimodal machine learning: Principles, challenges, and open questions. arXiv preprint arXiv:2209.03430.

Podcast-style discussion of the article

The raw/unedited podcast discussion produced by NotebookLM (proceed with caution):

Other popular articles

How to (and not to) log transform zero

By on
Image
Image courtesy of the authors: Survey results of the papers with log zero in the American Economic Review Academic's take Log transformation is widely used in linear models for several reasons: Making data "behave" or conform to parametric assumptions, calculating elasticity, etc. The figure above shows that nearly 40% of the empirical papers in a selected journal used a log specification and 36% had the problem of the log of zero. When an outcome variable naturally has zeros, however, log transformation is tricky. In most cases, the solution is to instinctively add a positive constant to each value in the outcome variable. One popular idea is to add 1 so that raw zeros remain as log-transformed zeros. Another is to add a very small constant, especially if the scale of the variable is small. Well, the bad news is that these are all arbitrary choices that bias the resulting estimates. If a model is correlational, a small bias due to the transformation may not be a big con...
Read more

What if parallel trends are not so parallel?

By on
Image
Image courtesy of github.com/asheshrambachan/honestdid Academic's take In difference-in-differences models, parallel trends are often treated as a make-or-break assumption, and the assumption is not even testable. Why not? One misconception is that comparing the treated and control units before treatment is a test of the assumption. In reality, the assumption is not limited to the pretreatment period but for the entire counterfactual. Comparing trends in the pretreatment period is only a plausibility test. That is, "parallel pretreatment trends" and "parallel counterfactual trends" are not the same. We can observe the former, but we need the latter. So, what we have is not what we want, as is usually the case. Even testing for parallel pretreatment trends alone is tricky because of potential power issues (evidence of absence is not the absence of evidence!) and other reasons, including a potential sensitivity to the choice of functional form / transformations ...
Read more

Causal inference is not about methods

By on
Image
Image courtesy of Eleanor Murray - epiellie.com Solo post: Academic's take Causal modeling is becoming more popular after some notorious failures of overly optimistic reliance on black-box predictive models to solve business problems (e.g., Zillow's iBuying). This is great. We are also increasingly seeing the introduction of a new method that "solves" causal inference. This is not so good because it misdirects attention. Causal inference has more to do with data and assumptions than it does with methods. No method can "solve" the causal inference problem (although it can help by reducing bias). If anything, regression is one of the most common methods for causal inference, so regression is an effective causal inference method when all else is in order . This is different from predictive modeling, where brute force bias reduction using the most complex method may succeed. Price elasticity of demand problem Simply put, we want to know how demand will change ...
Read more