Bridge the Tech-Business Gap to Harness the Power of AI

While AI is often celebrated as "the new electricity," powering everything from marketing to finance, its successful implementation requires understanding not just technical capabilities and limitations, but most crucially, business fit.

“Even the best AI can fail, if the business context is not right,” said Professor Jiaming Xu,  of Duke University’s Fuqua School of Business.

In a live talk on Fuqua’s LinkedIn page, Professor Xu provided insights on when AI works well in business, where it typically falls short, and what future business leaders need to understand to use it effectively. 

Xu examined the case studies of Zillow and Netflix, two eminent examples illustrating how AI models, even powerful ones, can succeed or fail depending on business context and implementation strategy.

When AI forecasting falls short: Zillow's iBuyer program

Founded in 2006 to bring transparency to real estate, Zillow developed a sophisticated AI system called "Zestimate" to evaluate home values, Xu said. 

After years of refining this technology – including the $1 million Zillow Prize competition to crowdsource improvements to its home valuation algorithm – in 2018 the company launched "Zillow Offers," an iBuyer program where Zillow would purchase homes directly from homeowners based on AI-generated valuations, then resell them for profit.

“They used a very sophisticated deep learning model that took all information like home details, geographical data, tax assessments, and market trends,” Xu said. ”It would take ‘multi modal’ inputs like video, image or text, to come up with a nationwide value for any property.”

Zillow was so confident in this new line of business that its CEO once said, ‘Can you imagine if Netflix just ignores streaming?” Xu noted.

However, this venture ended abruptly in November 2021, when Zillow announced it was quitting the iBuyer business, causing its stock to plummet 23% in a single day.

The problem wasn't Zillow's AI house price estimate, but rather a business model problem, Xu explained.

While Zillow's AI excelled at estimating current home values, he said, the iBuyer business required accurately predicting future values, often six months or more ahead. "But there are just too many uncertainties to build an AI model to predict future prices without economic indicators," Xu said.

Beyond prediction limitations, the business model faced additional challenges: maintaining large inventories of illiquid assets, managing property repairs, and dealing with information asymmetry – where homeowners possessing more information about the actual condition of their properties would gladly accept offers they perceived as overpriced, leaving Zillow with overvalued inventory which is more difficult to resell.

In general, Xu said, it is very difficult to estimate the future value of the home, because there are just too many uncertainties about market conditions.

"Even world-class AI models may fail if paired with high-risk business models," Xu said. "Predictive power on the current value of a house doesn't guarantee predictive success six months from now."

Adapting AI to evolving business needs: Netflix's content strategy

In contrast, Netflix demonstrates successful AI adaptation to changing business contexts, Xu said. Originally a DVD rental service, Netflix transitioned to streaming in 2007 and needed algorithms to recommend content to subscribers. Like Zillow – but more than a decade earlier – Netflix ran the $1 million Netflix Prize competition in 2006 to improve its recommendation engine, resulting in sophisticated algorithms to predict user ratings.

Interestingly, Netflix didn’t use the winning algorithm, Xu noted. Its formula was basically identifying users with similar rating patterns and recommend movies they liked — “if people with similar tastes to you tend to like these movies, then you might like them too,” Xu said. But by then, Netflix's business was already evolving to original content production, which led to a new AI approach. Rather than simply predicting ratings, Netflix began analyzing viewing time data to evaluate content value, he said.

This approach led to data-driven decisions like investing $100 million in "House of Cards" in 2013. "Netflix was confident because their data showed that both Kevin Spacey and director David Fincher were popular with subscribers," Xu explained. This AI-informed decision helped propel Netflix's growth, making it one of the best-performing tech stocks of the past two decades, he said.

" The key difference from Zillow is that Netflix didn't sit there with their original algorithm. They adapted their AI to their evolving business strategy," Xu said.

How to lead successful AI initiatives

For business leaders looking to implement AI, Xu identified several critical factors:

• Understand AI's capabilities and limitations: Recognize what AI can and cannot predict in your specific context.

“AI cannot really predict the future price of a house, but in the Netflix case, AI can predict how much time this user is going to spend on a movie,” he said.

• Evaluate data quality and quantity: The success of AI models depends on having abundant, high-quality data relevant to the specific business problem.

“You're not going to have abundant, high-quality data in every business context,” Xu said. “If you work in a retail shop, you may only have limited customer data, and you need to re-evaluate whether AI is going to be useful for you.”

• Manage AI risks: Don't blindly trust algorithms to solve all problems; identify and mitigate potential failures.
• Bridge the tech-business gap: Business innovation is needed to effectively drive AI implementation.

"There's a lot of discussion about AI technology, but relatively less focus on the business innovation needed to drive AI success," Xu said.