Data-Driven Decision is more than Actionable Insights
Generally accepted popular opinion for an enterprise leadership is to take decisions from data, which is often referred to as being ‘data-driven’. In my recent conversation with over 10 Chief Information Officers and Chief Data Officers, this goal is more aspirational, than realistic.
While enterprises are investing in a lot of technologies and hiring skilled resources, to realize the returns or access ROI, one must understand, “Are these investments being made for ‘visibility’ or ‘better decisions’ from data?”
In this article, I present the difference between a data-driven approach for ‘visibility’ and ‘better decisions’, and how taking a structured approach to decision-making can lead to the desired return.
How Visibility and Data-Driven Decisions Differ?
Visibility is historical insights from data. It gives enough information to take actions as far as environment and states are more or less fixed. For example, an annual planning for a retail product. The outcome of visibility is deterministic and decision-makers can run sensitivity analysis to further refine the actions.
For actionable insights, traditionally analytical tools are used that can slice-dice data, show trends, correlate cause and effect and more. In the last decade, these insights are obtained using advanced technologies including AI, NLP to answer questions, rather than just inferring from dashboards. Regardless, the visibility is from a fixed state of data and often it leads to more questions.
For example, you can ask Alexa or Siri “What is the trending song for the day?” and expect a reasonably accurate answer.
However, can Alexa or any conversational AI tool decide an outcome if you ask, “What channel should I focus first to sell the new product based on my 4 constraints?”
The fundamental premise of ‘decisions’ is it is forward-looking. While decision-making relies on historical data or past experience, it is an iterative process. By nature there are many considerations for decisions and a range of options to evaluate. Hence, there is the need for predictive models where outcomes change with each scenario. Since the output is probabilistic, inputs and data from previous actions are important to continuously refine for better decisions.
One other important consideration is the objective of decision — short term vs. long term. While historical data or a predictive model can provide an outcome that is beneficial in the short term, it may not be the most optimized outcome for the long term.
Consider the classic strategy of Kohl’s partnering with Amazon as a place to return online orders. Historical data might have given a different result to avoid working with the competitor; however post pandemic Kohl’s would have found that foot traffic to store is more for returning Amazon orders than attracting people with a 30% off coupon.
This leads to framework of decision-making where one has to balance the outcome from historical data and explore different options to maximize long-term benefit.
Exploration vs. Exploitation
A well known framework of decision-making — ‘Exploitation vs. Exploration’ — helps to strategize the levers for data-driven decisions. Here ‘Exploitation’ relies heavily on historical data and ‘Exploration’ considers different options to maximize the objectives. Focussing heavily on exploitation might lead to missing out on untapped opportunities. For example, “Are there new trends happening in the market (such as sustainability) that you can take advantage of when you introduce a new product that was not applicable to your other products?”
Let’s go back to the above example of selling a new product through a channel to maximize profit. A predictive model with generated data forms the basis of the analysis. This may include the historical data of a similar product from the past. However, since the conditions might have changed from past to present, historical data have to be supplemented with an algorithm for simulations. The input to the simulations such as constraints, weightage, are drivers to explore different outcomes. Domain experts who have the context to the problem normally defines the inputs to these simulations.
How does the traditional vs. ideal data-driven approach differ?
Most organizations currently have analytical tools that provide a bottoms-up approach for data-driven decisions. They have various technologies to blend data, create a sophisticated data model and present the enriched data in a report. Since by design, the data flows one way, decision-makers download the reports into spreadsheets and run various scenarios to evaluate different outcomes.
A more practical approach to decision-making is top down that starts with what decision to consider. Then gather the relevant data (historical or generated), build a predictive model, explore the different outcomes and finally, optimize for long term. This collaborative approach aligns the objectives of all stakeholders
Key Takeaways:
- Decisions intuitively have two components — learning from the past and predicting for the future. Insights provide data to the first part and humans or systems act on the second part. In order to make ‘Higher Quality Decisions’, one must explore different options before deciding on the best optimized outcome.
- Domain experts’ input is necessary to evaluate options for decisions. A ‘Decision-first’ approach helps considering all alternatives and understanding the tradeoffs. It also provides a framework to iterate and continuously improve on objectives.
- Human intelligence combined with machine throughput on predictions makes Data-driven Decisions as strategic assets for the organization.
References
Dinin, Aaron “What if You Could Turn Your Biggest Competitor Into Your Competitive Advantage? ” Medium Dec 2, 2021
Exploration vs. Exploitation in Coupon Personalization, by Aliaa Atwi, Devavrat Shah, MIT Department of Electrical Engineering and Computer Science, Jan 2018 http://oastats.mit.edu/handle/1721.1/115729
Kor, Peyman “The Downside of Data-Driven Decision Making” May 27 2022
Acknowledgements:
Prof. Devavrat Shah, MIT Department of Electrical Engineering and Computer Science. Director of Statistics and Data Science Center at MIT.
Anshuman Lall, PhD, Founder and CEO, Predmatic AI
