9 min read
When it comes to analytics, eliminating your company's "insight gaps" and "execution gaps" is probably impossible, but closing the gap is not. One of the first things to be done is to realistically assess the situation and set out your organization's goal.
With typical academic rigor, Thomas H. Davenport made it very clear to us back in 2006 when he published his article in Harvard Business Review, Competing on Analytics, that analytics are the key to any modern, profitable, and transparent business. If that was true back in 2006, you can only imagine that it will be doubly true now, given the pace of change that we have all experienced over the years since then.
In this first blog of my planned series, I want to explore this topic and talk about how organizations need to define their ambition for the scope and scale of analytics. I don't want to get too hung up on terms here, so in this post I'm going to be using the term "analytics" to cover all disciplines and technologies as performed by a data scientist, including artificial intelligence (AI), machine learning (ML), deep learning (DL), natural language processing (NLP), or descriptive/prescriptive analytics.
While the answer for some businesses may be "yes", the reality of the situation is that we just aren't getting good enough—nor fast enough—to keep up with the demands being placed on us by the growth in data or business needs. Figure 1 illustrates this yawning gap in insight: the difference between the amount of data under management and the amount that can be usefully structured and analyzed to answer key business questions.
Figure 1. Insight gap
Some have argued that the move away from traditional RDBMS-based data warehouse technologies (such as Oracle and Teradata to the Apache Hadoop ecosystem) has done much to change the picture. But while this may be true of one or two specific use cases in many organizations, the majority of the time it was little more than a technology and skills change. In most cases, the adoption of Hadoop has actually exacerbated the "insight gap" problem, as it has allowed significantly more data to come under management without delivering the corresponding increase in the scope and scale of analytics required to make sense of it.
More troubling, perhaps, is that in almost every large enterprise you will still find a massive enterprise data warehouse running alongside the data lake. So far from coming down, costs have actually increased! I'm sure that increase is going to be felt more acutely post COVID-19 when budgets are squeezed further, and existing Cloudera and Hortonworks implementations will be required to migrate to the Cloudera Data Platform (CDP). Not only have Cloudera license costs gone up significantly since they became a de facto monopoly, but there is also bound to be a fairly sizable migration cost and painful effort involved, which can only get in the way of making any more progress on closing the insight gap.
In most cases, we see significant issues around infrastructure provisioning, rigid tooling choices, and the unnecessary number of manual steps in the process. It's typical for it to take months to provision a new data scientist with a suitable environment to start work. I'm certainly not trying to lay the blame on IT departments for this. Many complex steps are required to provision hardware, create and secure the cluster, provision the data, apply appropriate permissions on the data (as well as corporate tooling such as ETL and data governance), provision the corporate data science stack, and coordinate and manage patching.
Analytics can be used in many settings in a business, but they create greater leverage when applied directly to business processes through highly automated machine learning models. Our goal should be to apply analytical decision making to optimize each step of every business process by embedding analytics into the applications, as Figure 2 shows. In this way, we will both differentiate our business by serving customers better and drive down costs. This really should be our North Star.
Figure 2. Embedding analytics into applications
That brings us nicely to Figure 3, which illustrates the "execution gap". That is, the gap between an organization's ability to analyze and make sense of data and its ability to apply it in an operational setting. You will have seen this expressed more recently as the "operationalization problem" (often abbreviated to O16n).
Figure 3. Execution gap
You might well expect industries with a longer history of leveraging analytics to be much further ahead when it comes to integrating them into their applications, but in my experience, this is normally the exception and not the rule. Good examples include telecommunications with customer churn management and financial services with risk and fraud management, but these use cases are almost always one-off, artisanal in nature, and rarely extended beyond the boundaries of the specific department involved.
A lot of things need to happen in order to operationalize a typical machine learning model and put it into a production context. These will often include rebuilding the data pipeline and recreating the model, packaging into a microservice, managing control groups for A:B testing, determining alerting levels for model drift, applying changes to business intelligence (BI) dashboards, and incorporating the model into the process in the application.
Given that these many challenges touch on data architecture, data science, application development, testing, operations, and platform, it's not difficult to see how organization and collaboration are important challenges we must tackle if we are to address the "execution gap". Many companies are starting to experiment with combining roles such as DataOps with MLOps or MLOps with DevOps to overcome these challenges. To have a marked effect, attention also needs to be paid to the supporting tools and infrastructure.
Eliminating your company's "insight gaps" and "execution gaps" is probably impossible, but closing the gap is not. One of the first things to be done is to realistically assess the situation and set out your organization's goal. Do you plan to just update your infrastructure so you can more organically grow analytical capacity? Perhaps you plan to do more by hiring some additional data scientists and the infrastructure to support them, and thus grow analytics in a more inorganic fashion.
As goals go, that might be sufficient. To truly make a dent, I really believe we need to aim high by setting a target of exponential growth in analytical capacity and capability, as Figure 4 illustrates. That way, even if you fail to hit the exponential target, we will have driven value from each of our investments along the way and the business will be in much better shape to react to future market dynamics. That's the power of analytics when they're used to drive the business.
If your goal is for exponential growth, then automating the process is key, as is thinking about how to remove waste and defects from the system. In short, we need to industrialize the process. That's a topic I'm going to be tackling in future blog posts.
Figure 4. Set a target of exponential growth
This is the first in a series of blog posts focused on the industrialization of data analytics. Don't miss the second blog discussing the advantages of shifting right to create business value in your information management portfolio.
This blog post was originally published on HPE's Community: AI Insights.