CarMax takes gen AI for spin

One company not afraid to take generative AI for a test drive: CarMax.

The used-car retailer behemoth has been kicking the tires of the new tech to bolster its value proposition: helping people find cars.

The input and the output. A CarMax search optimization team employed the neural network-based language model GPT-3.5 to aggregate a comprehensive set of reviews for requested makes and models.

The company’s SEO staff tasked GPT with digesting online evals, sent out 30 days after a purchase. The model then presented a paragraph summary—one that’s reviewed by a human writer and then posted on the site.

A request like, “Summarize reviews CarMax customers left for a 2022 Jeep Liberty” led to a long list of individual opinions, ranging from the detailed (“perfect height to step in. Don’t have to step up or bend knees to get in car”) to the straightforward (“I love my Jeep”), to competing ideas about mileage (“awesome” and “great” to some, “not great” to at least one).

GPT synthesized the survey information and produced output that, after a human review, ended up on the site: The 2012 Jeep Liberty is a car that has been well-received by customers. Customers like the car’s design, the car’s handling, and the car’s performance. Customers also like the car’s storage and the car’s fuel efficiency.

“Manually, a content writer would have to go find those 5,000 different reviews on the different types of make and model, With ChatGPT, we’re able to actually do that in an automated way,” Shamim Mohammad, CarMax CIO, told IT Brew.

But companies aren’t necessarily ready to trust AI to make one-paragraph decisions based on a superhighway full of data.

What’s the holdup? An MIT Technology Review of 1,000 executives recently found that while all 96% of respondents see generative AI as impacting their business, only 9% have actually found a way to use it.

The holdup in implementation, according to the findings, is largely based on a need to better understand regulatory requirements—and their own data.

“Overall, our poll respondents and interviewees are nearly unanimous in recognizing the potential for AI to revolutionize the enterprise. When we examine the speed with which they are deploying the technology within their own organizations, however, we see that their enthusiasm is tempered by substantial challenges and concerns, particularly around risk, and by uncertainty about the technology’s ultimate effects,” concluded one section of the MIT report.

The review cited three main reasons why companies have yet to implement generative AI: a lack of understanding the risks, a need to better understand the data, and a need to develop use cases.

“With generative AI, it’s there. The capabilities and the features are there. It’s just the operational mechanics behind it, which are still quite rudimentary,” said Nicola Morini Bianzino, global chief client technology officer at EY.

With active generative AI users at just 9%, prompt-specific projects like the SEO test at CarMax are often limited affairs, not company-wide mandates.

CarMax’s efforts were an experimental, even perhaps rudimentary idea from one internal team testing and learning rather than a step-by-step demand from the top, said Mohammad. The CIO’s confidence in the company’s data-governance policies helped his org overcome any hesitancy to move forward with Gen AI.

According to Tapio Christiansen, senior manager of PR and executive communications at CarMax, the company has a C-suite-led governance committee, responsible for data privacy and compliance.

“We have a lot of training going on with the company guardrails and things like that so that people know what they can do or cannot do. And then, at the same time, I’m not very strict…We want to empower the users and our engineers,” Mohammad said.

The SEO side of the house is just one small percentage of the many CarMax teams considering the use of the tech, according to the CIO, who didn’t want to share specifics about new prompt-y implementations.

“All these teams have their missions, and they’re looking for how do we deliver those missions. And those teams are coming up with different ways that AI and gen AI especially is part of what they’re working on,” said Mohammad.