Dr. Charles Liu
Business Analytics and Statistics
The University of Tennessee, KnoxvilleSelling a car with AI: I sold my old car to CarMax using their online instant offer. It was super accurate and convenient!Since I would be traveling for a while, I wanted to trade in my car. After a decade of driving my car, it was time for an upgrade anyway. So, I tried CarMax, a company known for its instant price estimates. I entered a few pieces of information about my car on the CarMax website. In a couple of minutes, I received an email offer with a highly competitive price. With this instant online offer, I visited the local CarMax shop, and the in-person appraisal confirmed that the online estimate was 100% accurate. Amazing! I could drop off my car anytime within one week and receive a check. Convenient!
How CarMax does it: They use "predictive modeling" (aka supervised learning) which is basically fancy AI. You give them info about your car, and their model spits out a price.How can CarMax estimate prices for used cars? Remember, I had to provide some information to obtain the price estimate. The information I provided, and any associated information CarMax can match with the VIN (Vehicle Identification Number), are considered inputs for the predictive modeling problem. The potential resale price is the output the model aims to estimate.
Now, how to derive the output given the output? This is essentially a predictive modeling or supervised learning problem, which is the core of modern Artificial Intelligence systems. The supervised learning requires training data, consisting of a comprehensive and representative collection of input and output pairs. Once we have collected the training data, a supervised machine learning algorithm can be trained to predict, for instance, prices for used cars. There are many such algorithms developed by machine learning researchers, and we can use computing experiments to select the one most likely to yield optimal predictions given the particular application.
AI needs good data: For the AI to work well, it needs tons of "comprehensive" (covers everything) and "representative" (like what real customers do) training data.These two adjectives I used to describe the training data are important. By “comprehensive”, the training data must encompass diverse instances of the problem to be solved. For example, if the training data includes price records for only a specific car brand or model, the trained model will only be effective for the covered cars, or those similar enough to the covered ones. By “representative”, the training data must closely resemble typical user inquiries. If sedan trade-ins are common among Carmax customers, then sports cars should not disproportionately dominate the training data.
Too much or too little info = bad predictions: If the AI doesn't have enough info, it's "underfitting." If it has too much irrelevant info, it's "overfitting." Both lead to bad results.Can one predict prices for used cars without training data? Sure, but that would not be a machine learning or AI approach. The essence of machine learning-based AI systems lies in their ability to derive input-output associations without explicit knowledge of the theoretical connections or mechanisms. Instead, we hope the algorithm can discover the association links automatically. To this end, the input must contain enough information to support the training process. Otherwise, without sufficient data, the generated predictions cannot be accurate. Conversely, the input should not rely on overly specific and fine-grained information that is scarce for making new predictions. For instance, a model designed to select job candidates might get fixated on irrelevant details like font choices in the resumes, rejecting qualified candidates based on superficial aspects of the resumes rather than job skills of the candidates. Both issues, insufficient information and excessive information, contribute to underfitting and overfitting in AI systems, respectively.
Human smarts still matter: Even with AI, human "domain knowledge" (expert understanding) is super important.The automatic discovery of input-output associations from vast training datasets is undeniably beneficial. However, a crucial question arises: what role does human expert knowledge play when available? Is domain-specific expert knowledge still relevant in the age of AI? If so, how can such domain knowledge be integrated into the training of machine learning models? Pondering these questions means you’ve stumbled upon a vein of potential business success. These are vibrant research areas within the machine learning and artificial intelligence communities. Companies like Amazon, Google, and Microsoft exemplify how the integration of machine learning with specific domain knowledge and problem contexts can lead to market dominance.
CarMax's secret sauce: They dominate because of their huge network, real-time demand info, and deep AI expertise. They can move cars around the country to match buyers and sellers perfectly.To make sure I was getting a good deal, I went to another dealer with the CarMax offer. Not surprisingly, the dealer offered a significantly lower price than that from CarMax. The dealer admitted that “nowadays it’s hard to compete with CarMax”. Why? He pointed out another advantage of CarMax: its large network of locations, sellers, and buyers. This network allows CarMax to leverage real-time demand data when determining optimal purchase prices. For instance, if a customer in California is seeking a particular car model, CarMax can acquire that vehicle in Tennessee at a competitive price, still factoring in the transportation cost to the California buyer. This strategic approach is unique to CarMax, as conventional local dealers lack the benefits of large-scale historical training data and access to immediate information from a large network.
Everyone wins with efficiency: CarMax's AI can make both sellers and buyers happy. More efficiency means better prices and selections for everyone.We have discussed several strategies that CarMax can use to buy cars from sellers like me. How about players at the other end of the game, i.e., buyers who want to buy cars from CarMax? Obviously, CarMax can leverage similar strategies to improve the efficiency of its selling process so that the buyers will be satisfied with the cars and prices meeting their expectations.
Now you may wonder, since CarMax works with both sellers and buyers, when it adopts some AI models to make sellers happier, will buyers be hurt in some way, e.g., having to pay more? While it can be difficult to fairly assess the benefits received by different customers, in theory, it is possible that all customers benefit from a more efficient platform. No one was hurt and everyone won! Consider the following example. Buyer #1, living in Knoxville, is trying to buy a relatively cheaper vehicle. However, the best he can find locally is a used BMW, which is nice but slightly too expensive. Without access to a broader platform, he decides to buy the BMW with an expensive financing option. Meanwhile, Buyer #2, living in Nashville, is trying to buy an entry-level luxury car. The BMW on the Knoxville market is ideal, but she is not aware of the opportunity. So, she eventually chooses a car from a local seller. She saves some money but is not really satisfied with the car. Now consider what CarMax can help them by transporting the two vehicles. Buyer #1 does not have to borrow money since the car from Nashville is much cheaper than the BMW. Buyer #1 is happy to pay more and drive a BMW. Since both buyers are satisfied with the purchases, both transaction prices would be higher, so both sellers would also benefit from the platform services.
Other businesses using this: Think Google Ads, Uber, and Airbnb – they all use data-driven AI for their success.CarMax is not the first one. There have been many successful business models built on data-intensive predictive optimizations. For example, using machine learning, optimization, and data-driven analytics, Google can optimize online advertising, Uber can match drivers and passengers, and Airbnb can help us find a nice vacation place.
Reference:
https://volweb.utk.edu/~buad220/unlocking-business-value-with-ai.htm