Too many organizations still view data science a technical skill, acquired simply by hiring a few data scientists. By doing so, they are not building the critical organizational muscle necessary to leverage predictive models as an enduring competitive advantage. Today’s leading companies didn’t get there by accident. They got there by putting models at the center of their business. They release innovative products powered by models, such as Netflix’s recommendation algorithm or Amazon’s Alexa assistant. What’s less obvious is how these companies are weaving models into their business operations to drive constant, iterative improvement. Whether it’s fraud detection, supply chain optimization, or employee retention, model-driven companies are leveraging the power and scale of models in ways often invisible to the end customer or competitors.
This transformation is not limited to the technology sector. It is akin to what began 30 years ago with the emergence of software development as a key organizational discipline. At first, there were only a few “software” companies. Software teams were relegated to the dark corners of the enterprise. Now, software engineering is a defining organizational capability across industries. Those companies who did not fully embrace the change have ceased to be relevant, if they exist at all. The same dynamic is happening with data science. Companies that embrace it as a discipline will become leaders, and those that do not will cease to exist.
According to a recent survey of data science practitioners and managers, 90% of companies view data science as a key source of innovation, but less than 30% have more than five models driving production business processes. The reality is most organizations still view data science as a technical practice of talented individuals, while model-driven companies view data science as an organizational discipline.
Watch for These Obstacles
Based on our experience working with hundreds of data science organizations, we believe the root causes of this disconnect generally fall into four clusters.
- Static infrastructure – Data science teams require far more flexibility and agility with infrastructure than their BI predecessors. IT departments that equip data science as a side-project will likely see turnover and/or the creation of “shadow IT” silos on the desktop of every data scientist.
- Knowledge silos – Data science teams spend too much of their time re-doing work that someone else has already done, leading to frustration and (again) turnover. Companies that view data science as an experiment often have single tracks of work happening in independent swimlanes, limiting cross-fertilization and improvement in their collective abilities.
- Iteration friction – Teams that view data science as an experiment over-emphasize the value of any one project. Model-driven organizations recognize the value of iteration velocity across all projects. They quickly and effectively prioritize work, ensure that promising results are actually operationalized, and incorporate feedback to inform new research.
- Model liability – Teams that view data science as an experiment often fail to anticipate and mitigate the potential risks that surface once a model is in production. Consider the case of Knight Capital, where an errant trading model lost hundreds of millions of dollars in a few hours. Even in non-regulated industries, organizations cannot claim ignorance of the impact of experimental models.
Advice for Transitioning Data Science from Experiment to Business Lynchpin
Organizations need to shift their mindset and investment strategy to cement data science as a business lynchpin. Simply hiring more data scientists and buying more tools will not be enough. The three core elements for organizational transformation ring true here and should be central to every strategy aimed at increasing the business value of data science.
- People – Executives need to bring data science closer to the business. This applies to both hiring, where stakeholder empathy must be prioritized, and to organizational design, where teams cannot be isolated in ivory towers without business accountability. Further, focusing on maximizing collaboration, both among data scientists and between data science and other business stakeholders, has been identified as the top contributor to successful data science programs.
- Process – Data science cannot be treated as an experiment. However, the actual data science process is much more experimental than most executives realize, particularly compared to the traditional software development process. To build this organizational capability, companies must acknowledge the realities of probabilistic research. Some projects won’t yield fruit. Iteration is key as models inevitably drift over time.
- Technology – Smart technology choices can facilitate best practices and increase data scientists’ productivity. Data science technology is not an end in and of itself. By focusing on tool agility rather than hoping for any single silver bullet, teams will position themselves to benefit from the ongoing innovation in the space.
Companies that invest in the people, process, and technology to create an organizational capability will increasingly separate themselves from those who dabble in side experiments. Although most companies believe this, they are stymied by a consistent set of barriers. By addressing the root causes holding organizations back from becoming model-driven, these companies not only survive, they can thrive.
Sign up for the free insideBIGDATA newsletter.