Skip to content

AI Success, But Not Business Success

In their book, “Mining Your Own Business,” Jeff Deal and Gerhard Pilcher, COO and CEO of Elder Research respectively, describe what I’ll call “The Case of the Climbing Churn.” Churn is when a subscriber cancels or fails to renew a service or subscription. A successful predictive model for identifying likely churners was deployed for a mobile phone service provider, and call center agents began reaching out to them to encourage renewals. Unfortunately, the churn rate rose!

Anyone attempting to reach another party on the phone knows that the most likely outcome is failure to answer the phone. Investigation revealed that in these many cases of non-reply, the agents were leaving voicemail messages for the customers. These voicemails, instead of generating renewals, were alerting subscribers that their contracts were about to expire, effectively letting them know that they could now switch carriers without penalty.

Fixing the problem was easy, requiring only a degree of business sense: agents were told not to leave voicemails. Churn then dropped, as predicted by the model, since many customers found the renewal offers when they could be explained, appealing. The churn reduction from using the model the right way for just one month more than paid for its development; after that, it was all profit.

Next, we turn to a couple of cases where strategic actions or inactions, unrelated to AI, dampened the business trajectory of a couple of famous businesses that were built on data and machine learning.


Pandora is an internet music radio service based fully on predictive algorithms and data. It allows users to build customized “stations” that play music similar to a song or artist that they have specified. When it started, Pandora used a nearest-neighbor style clustering/classification process called the Music Genome Project to locate new songs or artists like a user-specified song or artist.

Pandora was the brainchild of Tim Westergren, who worked as a musician and a nanny when he graduated from Stanford in the 1980s. Together with Nolan Gasser, who was studying medieval music, he developed a “matching engine” by entering data about a song’s characteristics into a spreadsheet.

In simplified terms, the process worked roughly as follows for songs:

Pandora established hundreds of variables on which a song can be measured on a scale from 0 to 5. Four such variables from the beginning of the list are:

●Acid Rock Qualities
●Accordion Playing
●Acousti – Lectric Sonority
●Acousti -Synthetic Sonority 

Pandora paid musicians to rate tens of thousands of songs on each of these attributes. This step represented a significant investment and provided a basis for defining highly individualized preferences as a user gave a thumbs up or thumbs down while listening. Over time, Pandora developed the ability to deliver songs that matched the taste of each user. A single user might build up multiple stations around different song clusters. Clearly, this is a more sophisticated approach than selecting music on the basis of which “genre” it belongs to.

Note the role of domain knowledge in this machine learning process. The variables were tested and selected by the project leaders, and the measurements were made by human experts. Yet, this human role was the Achilles heel of Pandora: it was a costly bottleneck, obstructing the flow of new songs into the system.

As the industry matured, music streaming services later came to omit this step of pre-labeling songs, and to rely on machine learning algorithms that get input only from users. Collaborative filtering, for example, recommends songs that are liked by other people who share your tastes (enjoy the same songs). Deep learning networks (which were not practically available at Pandora’s inception) can take the sound waves in songs and derive features that can then be used to predict user choices.

Pandora was a pioneer in licensed music streaming but was later eclipsed by Spotify and Apple Music. The key competitive differentiator between Pandora and its rivals, though, was not a difference in algorithms. In fact, it had nothing to do with AI. Rather, it was the nature of the product being sold. Pandora was designed to be “personalized radio.” It did not let you play songs on-demand or build up a library of downloaded music. Spotify and Apple Music both offer those features to users, which gave them a leg up in the marketplace. They now claim more than half the global market, with Pandora reduced to 2%.


To be fair, Pandora’s future was hobbled by the past from which it emerged. The commercial music business had successfully fought off challenges from internet platforms like Napster, where users could widely redistribute songs without paying royalties. Pandora decided to create a legitimate streaming path, but its business model could not afford to pay artists royalties akin to those paid by record companies. Its mission was, therefore, circumscribed from the beginning to avoid legal challenges from the music industry. In establishing itself as an early market leader, Pandora softened up the music industry to the point where it accepted the inevitability of streaming, opening the way for competitors to offer more to the customer.


The internet is famous for disrupting existing businesses, and, in 2004, the home sale business represented one of the biggest targets. Over 3 million realtors in the U.S. enjoyed a cartel-like protection from competition in the form of “ethics” codes that dictated adherence to a strict commission structure. Promulgated and promoted by realtor organizations, the codes effectively assured a commission in the neighborhood of 6%. The industry was extremely disaggregated, with no brokerage company accounting for more than 3% of the realtor agents.

In 2004 Zillow arrived with an internet platform that allowed homeowners and prospective purchasers to see the estimated price of virtually any house they were interested in. This information was previously the province of licensed realtors through the Multiple Listing Service. Zillow went public in 2011. Its statistical models on which price estimates were based did not require information that was hard to find, and, indeed, relied heavily on assessed values of properties, which were publicly available. The mechanics and effort required to obtain these data constituted the bulk of the modeling effort. Nevertheless, even if less than 100% accurate, the estimates attracted consumer attention and, once they became ubiquitous, the Zillow platform became an attractive place for realtors to advertise. As the platform became more dominant and widely used, the need for realtors to be seen on Zillow increased, and the advertising premium that Zillow could command grew. Zillow’s strategy was to accept the role of independent realtors but capture more and more of the commission in the form of ad fees.

Zillow’s position was challenged by another internet entrant, Redfin, which provided a similar platform that enabled consumers to view house prices. Unlike Zillow, Redfin did not eschew the realtor role itself -in fact, it started business as a real estate brokerage. Redfin sells homes directly to consumers via its own agents, posing more of a challenge to the established industry. By offering this more traditional sales service, a service unrelated to predictive algorithms, Redfin began to catch up to Zillow. The two are now approximately equal in market.


Zillow, whose stock price was flagging in the several years prior to 2020, has been revitalized by the strong housing market that followed the end of pandemic lockdowns. The company is now doing well, conventional realty brokerages continue to advertise with it, and it remains an open question whether a “data + advertising” strategy (Zillow) or a “data + sales force” strategy (Redfin) will prevail. Or, perhaps, a third competitor with a new strategy will emerge: traditional independent realtors have continued as a strong force and remain a target for disruption.

Note: In its consulting engagements, Elder Research is known for developing analytics strategies only in the context of a broader business strategy. Read more in Leading a Data Analytics Initiative, an ebook extract from Mining your Own Business.