TDWI | Training & Research | Business Intelligence, Analytics, Big Data, Data Warehousing

Analytics Meets Decision Making: What Does Automation Mean for the Future of Data Scientists?

There are many analytics products on the market that claim you merely need to load your data and out will come actionable insights. That's almost certainly wrong.

• By Joel Shapiro
• October 31, 2016

As analytics becomes increasingly sophisticated, it's tempting to buy in to the idea that analytics can be automated to solve business problems. Automating data collection, storage, retrieval, analysis, and action can work quite well in specific contexts (such as self-driving cars). However, automation falls short for solving most important business issues, which tend to have unique and idiosyncratic goals, processes, and contexts.

There are many analytics products on the market today that claim you need merely to load your data and out will come actionable insights. Although alluring, that's almost certainly wrong.

If your goal is only to create a predictive model, and action is entirely determined by that model, then automation can work well. For instance, every time you swipe your credit card, your credit card company compares your purchase to a range of variables based on your spending patterns, including the vendor, the amount, the location, and the time of day.

That comparison helps the credit card company flag suspicious activity, and it works because the action to deny the transaction is wholly based on the prediction and one underlying rule: if the predicted probability of fraud is greater than an accepted level, deny the transaction; if not, accept it.

This type of binary decision is not unlike a thermostat, which determines its action -- turn on or turn off the furnace -- based on the settings and the incoming data of temperature readings. If the temperature in a room or building is too low, the thermostat tells the furnace to turn on.

Business Strategy Is Not a Thermostat

Driving business strategy -- what do we need to do to please customers or to grow or expand the business -- is not a decision that relies exclusively on a a predictive model. Business strategies rarely fit into neat, binary predictions, and they almost never have easy-to-implement solutions like an on/off switch.

That's why, fundamentally, analytics is a managerial and leadership problem -- not a data science and IT problem. Managerial processes are difficult to automate because the connection from analytics to action can mean very different things across different contexts. Context is key: managerial actions typically require context-specific judgment and content expertise, which are lacking in analytics alone.

Analytics, first and foremost, is about identifying important problems that analytics can help solve. The next step is determining how results from analytics can inform action.

For instance, it's probably a safe assumption that people (like me) who use the Starbucks mobile app to purchase their favorite beverages and food items spend more money at Starbucks than people who don't have the app. Assuming that's correct, then the Starbucks marketing department could conclude that all Starbucks needs to do to increase sales is get all of its customers to use the mobile app.

That logic, however, might be faulty. It's quite likely that people who use the Starbucks app are already loyal customers and spend more than others regardless of the app.

Business leaders often overlook the fact that good decisions with analytics require a deep knowledge of business context as well as an understanding of how the data was generated. If Starbucks ran a true experiment, randomly assigning some customers to get the app and some others to a control group that didn't, then concluding that the app increases spending might be absolutely correct. In the absence of a true randomized, controlled experiment, however, you can't draw any conclusions about whether the app increases spending.

I don't imagine that Starbucks will be automating their marketing decisions anytime soon, but this hypothetical example makes clear the shortcomings of automation. Even when the data looks identical, different data-generation processes often lead to very different conclusions.

Automated analytics doesn't easily distinguish among data-generation processes, and it can't capture nuanced business context. When we gloss over these critical details, we come to wrong conclusions and make bad decisions. That's an outcome that neither data scientists nor business strategists can afford.

Bridging Data Science and Business Strategy

What does this all mean for the future of data scientists? Successful businesses will need to build clear bridges between data science and business strategy so the former can inform the latter.

We can't think of data scientists simply as the technical experts who build good models and then send those results to the business experts. That approach has already proven to be problematic -- plus, data scientists who have only technical skills will be the most easily replaced with analytics automation.

To succeed with analytics, business leaders must be conversant (or even fluent) in data analytics and they must ensure that their data scientists are immersed in the business context. Data scientists must strive to become more than technical experts and must develop as much business expertise as possible.

Analytics automation can do an exceptional job of finding trends and building models, but it can't solve complex business problems. That still requires deep knowledge of the business context and enough knowledge of data science to understand the role of trends and models in making business decisions.

Over time, automation might reduce the demand for technical-only experts, but the need for good data scientists who are skilled at translating data science to business problems will grow. That's great news for data scientists who are willing to embrace it.