Operational analytics not only predict what the next best action is, but also cause the action to happen without human intervention.
In the future, just having predictive models that suggest what might be done won’t be enough to stay ahead of the competition. Instead, smart organizations are driving analytics to an even deeper level within business processes—to make real-time operational decisions, on a daily basis.
These operational analytics are embedded, prescriptive, automated, and run at scale to directly drive business decisions. They not only predict what the next best action is, but also cause the action to happen without human intervention. That may sound radical at first, but it really isn’t. In fact, it is simply allowing analytics to follow the same evolution that manufacturing went through during the industrial revolution.
Centuries ago everything was manufactured by hand. If you needed a hammer, for example, someone would manually produce one for you. While manually manufacturing every item on demand allows for precise customization, it doesn’t allow for scale or consistency. The industrial revolution enabled the mass production of hammers with consistent quality and lower cost. Certainly, some customization and personal touches were lost. But the advantages of mass production outweigh those losses in most cases. It remains possible to purchase custom made items when the expense is deemed appropriate, but this usually only makes sense in special situations such as when the purchaser desires a one-of-a-kind piece.
The same revolution is happening in analytics. Historically, predictive analytics have been very much an artisanal, customized endeavor. Every model was painstakingly built by an analytics professional like me who put extreme care, precision, and customization into the creation of the model. This led to very powerful, highly-optimized models that were used to predict all sorts of things. However, the cost of such efforts only makes sense for high-value business problems and decisions. What about the myriad lower value decisions that businesses face each day? Is there no way to apply predictive analytics more broadly?
Operational analytics recognize the need to deploy predictive analytics more broadly, but at a different price point. An assembly line requires giving up customization and beauty in order to achieve an inexpensive, consistent product. So, too, operational analytics require forgoing some analytical power and customization in order to create analytics processes that can increase results in situations where a fully custom predictive model just doesn’t make sense. In these cases, it is better to have a very good model that can actually be deployed to drive value than it is to have no model at all because only an optimal model will be accepted.
Let me illustrate the difference with a common example. One popular use of predictive models is to identify the likelihood that a given customer will buy a specific product or respond to a given offer. An organization might have highly robust, customized models in place for its top 10-20 products or offers. However, it isn’t cost effective to build models in the traditional way for products or offers that are far down the popularity list. By leveraging the learnings from those 10-20 custom models, it is possible to create an automated process that builds a reasonable model for hundreds or thousands of products or offers rather than just the most common ones. This enables predictive analytics to impact the business more deeply.
Operational analytics are already part of our lives today, whether we realize it or not. Banks run automated algorithms to identify potential fraud, websites customize content in real time, and airlines automatically determine how to re-route passengers when weather delays strike while taking into account myriad factors and constraints. All of these analytics happen rapidly and without human intervention. Of course, the analytics processes had to be designed, developed, tested, and deployed by people. But, once they are turned on, the algorithms take control and drive actions. In addition to simply predicting the best move to make or product to suggest, operational analytics processes take it to the next level by actually prescribing what should be done and then causing that action to occur automatically.
The power and impact of embedded, automated, operational analytics is only starting to be realized, as are the challenges that organizations will face as they evolve and implement such processes. For example, operational analytics don’t replace traditional analytics, but rather build upon them. Just as it is still necessary to design, prototype, and test a new product before an assembly line can produce the item at scale, so it is still necessary to design, prototype, and test an analytics process before it can be made operational. Organizations must be proficient with traditional analytics methods before they can evolve to operational analytics. There are no shortcuts.
There are certainly cultural issues to navigate as well. Executives may not be comfortable at first with the prospect of turning over daily decisions to a bunch of algorithms. It will also be necessary to get used to monitoring how an operational analytics process is working by looking at the history of decisions it has made as opposed to approving up front a series of decisions the process is recommending. Pushing through such issues will be a necessary step on the path to success.
The tools, technologies, and methodologies required to build an operational analytics process will also vary somewhat from those used to create traditional batch processes. One driver of these differences is the fact that instead of targeting relatively few (and often strategic) decisions, operational analytics usually target a massive scale of daily, tactical decisions. This makes it necessary to streamline a process so that it can be executed on demand and then take action in the blink of an eye.
Perhaps the hardest part of operational analytics to accept, especially for analytics professionals, is the fact that the goal isn’t to find the best or most powerful predictive model like we’re used to. When it is affordable and the decisions being made are important enough to warrant it, we’ll still put in the effort to find the best model. However, there will be many other cases where using a decent predictive model to improve decision quality is good enough. If an automated process can improve results, then it can be used with confidence. Losing sleep over what additional power could be attained in the process with a lot of customization won’t do any good in situations where it just isn’t possible due to costs and scale to actually pursue that customization.
If your organization hasn’t yet joined the analytics revolution, it is time that it did. Predictive analytics applied in batch to only high value problems will no longer suffice to stay ahead of the competition. It is necessary to evolve to operational analytics processes that are embedded, automated, and prescriptive. Making analytics operational is not optional!
Photo credit: Aeon