Businesses need a different approach to going from data to decisions
Imagine you have a task to select one photograph for a photo frame in the house or submit a picture for a competition. Would you just take one photograph from your camera or take multiple pictures before selecting one? In the age of digital photography and affordable storage, most likely you will do the latter. Now, consider this analogy in the context of business decision-making.
Most decision-makers consider a range of scenarios, evaluate tradeoffs, and then decide on one scenario to execute. The question is, with a lot of business uncertainties that have forced the decision cycles to condense, are the decision-makers equipped with the right set of tools?
In this article, I present a structured approach to decision-making and the corresponding need for technology.
Decision-making approach as a funnel
Most of you might be familiar with the Sales and Marketing funnel. The top of the funnel is meant to create awareness and source as many prospects as possible. The next step is to qualify where the parties’ interests are aligned. The third step is the most important, which includes human factors and experience. This is followed by the decision point to engage and convert the deal.
The Data-driven decision-making process follows the same methodology as the Sales and Marketing funnel. It starts with simulating, experimenting with lots of scenarios, then qualifying a few of them, followed by human overrides on the output, and finally the decision with ongoing execution.
With the above background, let’s understand the technical interdependencies of each step.
Business Scenarios
By definition, decisions are forward-looking and therefore there is an element of uncertainty. Now if the data for decisions are the output of predictive models, then it compounds the uncertainty for optimum decisions.
Let’s assume the output from an algorithm, which has gone through the rigorous process of experimentation, training, and testing by data scientists is the most accurate and gives the best prediction. Still, the businesses in many operational use cases may not accept the output as is. While the data scientist may have included all the relevant features to predict the outcome, business decisions require more conditions to analyze that are outside the features of the algorithm.
To understand the above assertion better, let’s take an example of a CPG (Consumer Product Goods) company that has a model, which predicts demand with reasonable accuracy for different SKUs, per channel, per location. The decision-makers may not take that output and put it into a downstream system. They may want to run scenarios such as, what is the impact on working capital or cost of borrowing if they carry higher inventory or rebalance service levels for high-margin products.
As you can see, the accuracy of models is just one part; running business scenarios at scale with different levers is a necessary step for decision-making. More importantly, only domain-expert or business users, who have broader knowledge, are best equipped to run a wide range of scenarios.
Qualify
Not all scenarios are qualified for decision-making. Today, there are tools that allow you to run scenarios on a model. For example, you can change input parameters in a spreadsheet and get a different output based on a formula. However, the complexity increases when the ‘formula’ is a business logic with a large number of parameters and it has a dependency on data with high variability.
Hence, there is a need for a system to track all the scenarios. The decision-makers would want to understand the context behind scenarios, such as what inputs went into the model for the output. Automatic tracking helps to easily compare scenarios and evaluate trade-offs. Eventually tracking leads to qualifying a subset of scenarios for informed decisions from a large pool of runs.
Intent
This step is designed for decision-makers to override the output of a model. Most of the analytical systems today are “unidirectional” where read-only pre-computed values are presented to decision-makers as output. This is the right approach when the systems are providing deterministic output. However, when the data is probabilistic, decision-makers may want to change the output based on their experience or human intuition.
Accordingly, a “bidirectional” system is needed that not only generates a new output for every scenario but also captures the intent behind these overrides on the output. This intent of modifications becomes a knowledge base and also a feedback loop to improve the accuracy of models.
Decide
The final step is the decision after all the analyses are done. The output of this step is an action that somebody needs to execute. The interesting aspect to consider is the traceability of a decision.
Let’s go back to the CPG company demand forecasting example. After 3 months of a decision, the company achieved the goal of reducing working capital with low inventory, however, it found that a lot of the products had stockout, which resulted in competition gaining market share.
Without a trackable system that has the context behind decisions, there is no way to quickly trace back what scenarios were run, which of those were qualified, and why a particular decision was made.
Conclusion
Prior to digital photography, people were constrained to only 36 exposure film rolls. They were selective in taking pictures, not to mention the hassle of the manual development of photos. Digital photography has provided unlimited access to take pictures, catalog, and control the best experience.
Similarly in the world of cloud computing, where unlimited scenarios can be run and stored, why are businesses still constrained by “unidirectional” systems that force decision-makers to explore only a few scenarios in spreadsheets or rely on skilled technical experts? Organizations may want to consider designing a transactional system for decision-making.
