Human or Machine: The Most Important Question in Analytics

Arguably the most important questions in analytics these days is, “Who (or what) is going to make the decision?” There are two fundamental answers: a human or a machine. How the question is answered has all sorts of implications for what kind of people will do the analysis, what kinds of tools will be used, the process for the analysis, and so forth.

The import of this question has rarely been discussed, although Michael Li wrote an excellent article about it on Data Informed last summer. His focus was on what kind of data scientist you need. He correctly pointed out that data scientists who work on machine-made decisions are typically very different from those who prepare analyses for human decision-makers.

Much of what has been written and said about analytics assumes that the results will be created for and delivered to a human audience. If the primary decision-maker is a human, that human requires a lot of sensitive care and feeding. The analysis should be only as technical as the decision maker, and generally should be packaged as a story or a visual display. The analyst should be prepared to “socialize” the results – describing why the analysis was performed, persuading the decision maker that the results are valid and relevant, and answering any questions in non-technical terms. Relationship issues between analyst and decision maker are key; as Karl Kempf, an Intel Fellow and head of a decision engineering group at the company once told me, “If you want your analyses to have an impact, it’s not about the math, it’s about the relationships.”

From a technology standpoint, human decision makers as an audience for analytics are behind the fast rise of visual analytics. Relatively new analytics software companies like Tableau, Qlik, and Domo are primarily oriented to creating visuals for humans. Older analytics firms like SAS, IBM, SAP, and TIBCO also are emphasizing visual analytics. Companies like Narrative Science and Automated Insights are creating automated text narratives – again for human decision makers to consume.

Of course, not all humans are alike in their information orientations. An analyst working with a human decider needs to learn his or her preferences for visuals vs. text, the level of statistical analysis he or she finds comfortable, and how much interest there might be in underlying data assumptions (usually pretty low). And perhaps the most salient decision-maker characteristic is whether or not he or she has any real interest in analytics as a guide to the decision. Many executives prefer to use intuition or experience whenever they can get away with them. Analysts need to decide whether their time and efforts are worth the trouble before they even start doing quantitative analysis.

More and more, however, it’s not humans who are the recipients and decision makers of data and analysis, it’s machines. Machines are making all or most of the decisions in areas like programmatic advertising, search engine optimization, credit approval, insurance underwriting, Internet of Things applications, and many more. Machines are necessary for these jobs because there is a vast amount of data involved, and the results have to be so granular that many different models are involved. The decisions also need to be made in real time, and humans can’t react at that pace.

Perhaps obviously, there are far fewer human factors to consider when a machine is making the call. Machines don’t like visuals and they don’t need stories. They want data, ideally in structured forms like rows and columns so that it can be easily digested and analyzed. They don’t get tired or cranky, and they don’t have intuition to substitute for analysis. They do benefit from some human oversight, but it’s at a much higher level than with human decision analytics. Somebody just needs to make sure that they are still doing what they were intended to do and that they haven’t gotten off track.

Machine-driven decision making involves technologies like machine learning, neural networks and deep learning, and business rules. It also often needs something to move the work along, like complex event processing or event stream processing. There are vendors that supply these tools but, in most cases, the analysts (perhaps “data scientists” is a better term for these people) work with open-source tools. They feel they need to build systems themselves rather than employing packaged solutions. Much of their time is spent “munging” data – getting it into shape so it can be analyzed. Some of the vendors I mentioned earlier – SAS, IBM SPSS, and TIBCO Spotfire come to mind again – offer some software for machine-based decisions, but their history as human-based analytics firms makes it difficult to change focus.

The people who do machine-oriented analytics work fit the classic data scientist profile. Perhaps in the future we should refer to people who do analytics for human consumption as “decision scientists,” and those who work primarily with machines as “data scientists.” That might limit the confusion that surrounds these terms today. Decision scientists would specialize in working with human decision makers and all the social, psychological, and political challenges that implies. Data scientists would have to work with human managers as well, but less frequently and at a higher level.

The most important thing, however, is to be clear about which type of decision maker one is working with. That will drive many aspects of the analytics project. If you expect that a machine is going to make the decision and a human ends up doing it, or vice versa, the project probably will not go well. And if you have no idea who’s going to make the decision, you are really in trouble.

Tom Davenport, the author of several best-selling management books on analytics and big data, is the President’s Distinguished Professor of Information Technology and Management at Babson College, a Fellow of the MIT Initiative on the Digital Economy, co-founder of the International Institute for Analytics, and an independent senior adviser to Deloitte Analytics.

* This article was originally published by DataInformed on March 2, 2016.