Skip to main content

Authors: Raymond Chen, Operations Data Analyst; Dr. Pat Blachly, Data Science Consultant; Dr. Patrick Weinkam, Head of Data Science


Figure 1. The data flow of messages from Transport Companies, RunBuggy operations and the large shipper’s operations. The data flows from RunBuggy Winton software and Slack into our RunBot AI system 

RunBuggy has built the most advanced digital marketplace for automotive transportation. The digital marketplace depends on a combination of software, best in class operations team, and an integrated AI system. When these pieces are put together, we dramatically simplify the shipping process for large organizations and reduce what should be an expensive months-long onboarding process into a relatively cheap process that only takes weeks. In this study, we show how RunBuggy systems and people work together to improve the shipping processes and delivery time for a new, large shipping company. 

This particular process involves three groups of people and three software systems (Figure 1). The shipper places orders in the RunBuggy marketplace using a software called Winton and coordinates with RunBuggy operations in Slack. The orders are claimed by transportation companies in the Winton software and coordinate with RunBuggy operations in Winton. All of this data flows into RunBot, the RunBuggy AI system built on generative AI and other machine learning models. The AI system is used to track order progress both on the individual level but also in aggregate, which allows trends and insights to impact business decisions. 

Unclaimed Events Caused Most Shipping Delays 

During the first weeks of the new shipper partnership, we identified orders that were being delivered more slowly than expected. The data did not suggest any issues with pricing or transporter availability, rather we saw a high number of unclaimed events for orders with slow delivery times compared to orders with fast delivery times (80% increase for long haul moves and 30% increase for short haul moves). These unclaimed events happen when a driver has been assigned to an order but has to cancel because the order cannot be progressed. We sought to investigate what caused the unclaimed events by first evaluating reasons selected in Winton via a dropdown menu. We quickly realized that those reasons were over simplified or didn’t capture the complexity of issues that arise during transport. Therefore, we applied large language models (LLM) to interpret unstructured conversations between RunBuggy operations and the shipper’s operations or the transport company.  

LLM’s Identify the Main Causes for Unclaimed Events 

Figure 2. Results from the unclaimed reason study from unstructured conversations. The main reasons that caused delay were unavailable vehicles and pickup/drop-off date conflicts. Secondary issues were order cancelations, pickup/drop-off address issues, and the sixth most common cause was a driver issue.

The LLM’s suggest the shipping delays were mostly due to unorganized processes within the large shipping company. We compared the frequency of unclaimed reasons for slowly delivered orders to the frequency of unclaimed reasons for quickly delivered orders. In turn, we can identify what likely caused the delay (Figure 2). The main reasons shipping delays were unavailable vehicles and pickup/drop-off date conflicts. Other slightly less common issues were order cancelations from the shipper, pickup/drop-off address issues, and the sixth most common cause was due to the driver. In summary, the top five issues were because the vehicle and/or shipping information were not correct, which are all part of the shipper’s internal processes. The sixth major issue was due to the transporter or driver. This imbalance is not unexpected when there are new locations or processes within a large organization. However, the data suggest room for improvement for the shipping company’s internal processes. On the bright side, RunBuggy can identify the specific issues at hand so they can be corrected in the future.  

Organized Data is Necessary for AI Systems 

Figure 3. Table shows how data from a software drop down menu (Unclaimed Reason left column) rarely captures the reality of the situation. The “Grouped Standardized Unclaim Reasons” column is derived from using an LLM to interpret conversations. This process is only 70 % accurate when using unstructured Slack posts but over 90 % accurate when using conversations organized into Winton task management. 

While the AI system helped solve the problem, the most important contribution is data transparency from the RunBuggy operations process and the Winton software. However, not all data are created equal and some data can be completely useless for AI applications. Figure 3 shows how data from a dropdown menu is wrong roughly 75% of the time – dropdown menus force a distracted user to select from a short list that is too simple to handle the complexities that arise during shipping. Furthermore, we can analyze the accuracy of the LLM on unstructured Slack data as compared to more organized data in the Winton system.  

The Winton software includes a task management system that is used to progress orders through delivery and is the framework through which we can dramatically improve shipping times, control costs, and reduce the carbon footprint of trucking companies []. When using organized data from a task management system, we see accuracy improve to over 90% as compared to only 70% for disorganized slack conversations. Furthermore, disorganized data (such as that found in Slack) requires many times more effort to handle compared to organized data, and in the worst case, cannot be used for automated software solutions at all. 


In detail, automotive shipping is quite complicated due to many issues that are unique to this market – one cannot simply carry a car to someone’s doorstep like traditional freight as vehicles are large, need protection, require insurance and regulatory approval. Most large organizations underestimate this fact. In this study, we identified specific issues that delayed shipments during the onboarding of a new, large shipper and this information can hopefully help that shipper optimize their internal processes. During the onboarding of new customers, we commonly see other issues arise, not just for the shipper, but also for transporters and parties at the pickup or drop-off location. For example, new shipping routes necessitate an increase in transport companies, some of which hire new drivers or are new companies themselves. Because the RunBuggy system allows full transparency of transporter quality, we are able to leverage a ratings system to improve the transporter pool []. We also apply rating systems to businesses at the pickup/drop-off sites. Many companies lack robust software and transparency needed to uncover data patterns and will not effectively leverage AI tools. In conclusion, the RunBuggy digital marketplace uses unprecedented amounts of data to provide an optimized shipping solution in weeks and enables efficiencies not previously possible in automotive shipping. 

Description of the Approach  

For this case, the Data Science team wanted to understand the issues with this specific large shipper’s shipments. How long these moves took, reasons why they may have been delayed, and even why they were unclaimed. Utilizing the current data we had on the shipper’s VINS, we filtered out a subset based on a particular period and further divided the data into different cohorts based on the distance and pickup times. What we found using this method led us to take a closer look at the unclaimed reasons for each of these shipper’s VINS.   

Our current ability to draw conclusions about reasons for unclaimed orders very limited in that we utilize a limited list of reasons in the software that do not describe all of the challenges. To get around this issue, we turned to Slack to extract any unclaimed information they had on these VINS. Slack data is very unstructured and to make sense of this unstructured data, we turned to Large Language Models (GPT by OpenAI) to make sense of such data. By utilizing prompt engineering, we were able to feed it Slack messages and have it output a structured dataset containing any VINS mentioned as well as an unclaimed reason it believes to be associated with such VIN.   

Through the use of AI Systems such as ChatGPT, Large Language Models, and prompt engineering, we were able to structure data from Slack to provide greater context on the shippers moves, particularly their unclaimed reasons.   


Distance Cohort: <= 500 miles = Short Haul
Distance Cohort: >= 1000 miles = Long Haul
Pickup Time Cohort: Pickup Time Speed < 3 = Short Pickup Time
Pickup Time Cohort: Pickup Time Speed > 7 = Long Pickup Time
Pickup Time Speed: Difference between the Driver Pickup Date and Vehicle Available Date        

To learn more about RunBuggy’s Data Science and Artificial Intelligence initiatives, please visit