Machine Learning Deployments Need to “Change”

These days, my email inbox is stuffed with messages about products promising to transform my life through artificial intelligence (AI). From a growing number of AI “research” firms to a seemingly endless array of newly emerging application vendors touting their software’s revolutionary ability to leverage large language models (LLMs) to instantly solve virtually every use case that I could possibly think of, the hype for generative AI is simply off the hook.

Given the current economic challenges for much of the technology industry, it would be natural for the customer success (CS) management community to aggressively adopt these types of tools in hopes to scale delivery capacity in the presence of headcount limitations. However, at the recent San Francisco Customer Success Meetup, few of the 250+ attendees had actually begun to do anything with generative AI. Intrigued, I asked several people why this was the case. The most insightful answer that I received was from a longtime CS leader who simply stated, “I have not been an early adopter when it comes to AI. Most of what I’ve seen so far has been marginally helpful, but hasn’t been enough to push me into changing workflows.”  

Upon hearing this explanation, something suddenly clicked for me. CS teams are tasked with helping their customers figure out how to achieve the tangible outcomes – lower costs, higher productivity, increased responsiveness – that their salespeople have promised that “the software” will deliver. From experience, CS teams are aware that most of these promised outcomes require more work on the customer’s part than the vendor (or at least their salespeople) will ever admit and that these additional organizational costs are usually significant. Thus, CS professionals bring a heavy dose of skepticism to any claims of “plug-and-play” re$ult$.

Eric Siegel believes that we should be equally skeptical of claims made by machine learning (ML) software vendors and consultants. In his new book, “The AI Playbook: Mastering the Rare Art of Machine Learning Deployment” (MIT Press, 2024), Siegel points out that “the vendors selling ML software tools aren’t inclined to advertise that their products do not themselves perform operational change. They may be slow to explain that ML software takes on only limited – albeit central – technical portions of an end-to-end ML project. These vendors are incentivized to keep the focus on their technical products rather than the enterprise process …”

In his new book, it is clear that Siegel himself has gone through a significant change in perspective in the 10+ years since his first book, “Predictive Analytics: The Power to Predict Who Will Click, Buy, Lie, or Die” (Wiley, 2013), was published. A former computer science professor, an experienced analyst and consultant, and founder of the Predictive Analytics World conference series, Siegel is still a huge fanboy for predictive analytics and machine learning (in his new book, he asserts that “ML is the world’s most important technology” and “prediction runs the world”). However, given that most machine learning projects – some estimate up to 85% – fail to deliver actual business value, Siegel’s new book explores some of the root causes of these failures while proposing a methodology that he calls bizML to address these challenges. The steps in the bizML approach are as follows:

1. Value: Establish the deployment goal.
2. Target: Establish the prediction goal.
3. Performance: Establish the evaluation metrics.
4. Fuel: Prepare the data.
5. Algorithm: Train the model.
6. Launch: Deploy the model.

From my perspective, there is a lot to like here. One key insight from this methodology is that the deployment is at the center of the entire process and, in particular, that the ML decisions are made with operational implementation in mind. Failure to do this shows up in different parts of the process. With regard to the “Target” step, I have too often seen ML models make predictions at the wrong level of detail for implementation (e.g., aggregated monthly forecasts for environments with daily or hourly decisions at a regional or product level). Similarly, in the chapter on “Performance,” Siegel points out that traditional metrics around forecast accuracy are often deceptive as well as disconnected from the business goals of the deployment.

Beyond introducing the bizML framework and its deployment-focused mindset, Siegel’s book also strives to explicitly educate business professionals on some of the fundamental concepts associated with probability and prediction: “This is driver’s education, not auto-mechanic school. In order to drive, you do need extensive know-how, familiarity with core fundamentals such as navigation, acceleration, momentum, friction, and collisions. You must become intimately acquainted with how a car interacts with the world and how you control it.” As someone who has been teaching data, modeling and analysis skills to MBA students for 20+ years, this was music to my ears.

There are many familiar ideas in this book. Siegel cites several relevant references, most notably, James Taylor’s 2019 book entitled “Digital Decisioning: Using Decision Management to Deliver Business Impact from AI” (Meghan-Kiffer Press, 2019). While reading, I was often reminded of George Roumeliotis’ 2015 presentation “Design Thinking for Data Scientists” at the O’Reilly Strata Conference. More recently, some of these same themes are evident in Iavor Bojinov’s article “Keep Your AI Projects on Track” (Harvard Business Review, November/December 2023).

Still, I believe Siegel’s book makes a major contribution toward helping business people and machine learning professionals collaborate more effectively. My most significant complaint is that (subtly but stubbornly) Siegel still places machine learning at the center of the universe. For example, when telling the story of Jack Levis’ arduous but ultimately successful attempt to optimize package assignments and vehicle routing at UPS, he focuses primarily on the machine learning models associated with predicting package arrivals while paying relatively little attention to ORION, the Edelman Award-winning delivery routing optimization system that leveraged these predictions to drive massive cost savings. 

The reality is that ML models are usually just one component of what my longtime colleague Steve Alter has defined as “Work Systems,” which include people, processes, technologies, models and strategies. In today’s world, what Siegel refers to as “Machine Learning deployment” actually involves significant changes to these increasingly complex Work Systems. And very few data scientists have near enough training – or interest – in helping facilitate these types of changes or in how to anticipate some of the inevitable organizational challenges. From my perspective, one of the major contributions of the bizML framework is shining a light on the importance of organizational change.

In addition, Siegel explicitly points out that “sustaining a model’s viability moving forward takes maintenance, monitoring, and vigilance … these things require upkeep.” And, as my friends in the world of customer success intuitively understand, the same is true of any LLM-based technologies that become embedded into operational workflows.

On the plus side, it seems clear that the demand for customer success management professionals at companies with solutions based on machine learning and/or LLMs will be robust for the foreseeable future. As Siegel astutely points out, “an ML project is a consulting gig, not a technology install.” 

And don’t let any internal consultant or outside software vendor tell you differently.