Although AI technology has been available for a couple of years now, 85% of companies still struggle in scaling their efforts (Accenture, 2018) and 70% fall short in providing business value with AI (BCG, 2019). That is especially the case, when AI use cases need to be implemented in processes of an organization that require the involvement of business stakeholders. However, companies that are outperforming their respective industry in creating business value with AI solved the problem focusing on AI Transformation and spend more than 50% of their analytics budget on it (McKinsey, 2021).
While the technology is clearly an enabling factor, cultural transformation/change management is a crucial one for value creation. The transformation challenges I have been facing are for example that people are not used to work with data on a daily basis and are trained to interpret data to take actions, might have wrong expectations of AI, do not trust an analytics solution or fear that AI could reduce their control and power of decision-making. This article is addressing some of those points and proposing an analytical journey (Figure 1) that can help to facilitate AI Transformation by developing trust in AI and facilitating the feeling of involvement and ownership of an analytics solution. It is important to walk the whole journey, as only the last step of it is the one, which creates business value.
The Analytical Journey for AI Transformation
The analytical journey for AI Transformation can be broken down in 4 steps: Descriptive Analytics, Causal Analytics, Predictive Analytics and Prescriptive Analytics.
1. Descriptive Analytics
Descriptive Analytics helps to understand the past and provides the context for an analytical problem. It is important for the analyst/data scientist to align with business stakeholders on Key Performance Indicators (KPIs) to validate previous KPI developments and define the problem at hand, including the measurement of context factors (e.g. a target group, economic development, product perception) and actions that can be taken to increase the outcome (e.g. TV spending, pricing & promotions, product development). This step is crucial in order to move forward in the journey. There will not be any trust in the results of causal analytics if there is no alignment on KPIs, business drivers and their measurement. It also helps data scientists to understand the business challenge and business objectives of the stakeholders. Descriptive analytics can be done through insightful visualizations with charts or dashboards.
For example, a data scientist might want to understand the company’s sales volume over the past 2-3 years, across regions and in comparison to competitive companies. Descriptive analytics might give insights in and will tell a data scientist if a decrease in sales is a company, a regional or an industry wide challenge. Hence, descriptive analytics is important to create hypotheses together and to define possible actions with the business stakeholder for causal analytics. It is the most important step to create stakeholder involvement for the final solution.
2. Causal Analytics
Causal Analytics takes the hypotheses defined during the descriptive analytics step and provides insights on why something has happened. This step is important to explain the reasoning for past developments and a requirement to evaluate future actions. This can be done through statistical/econometric models (e.g. with Regression Analyses) or machine learning algorithms (e.g. Decision Trees), which are applied to understand the relationships between predictor and response variables. That step is insightful to business stakeholders to better understand what actions drove business outcomes in the past and can than be used for the predictive analytics step.
For example, a data scientist might want to understand why the sales volume of a company is less than average in a specific area, for a distinct customer segment or over recent months to find the drivers of that development (e.g. less investment in media or pricing above competitor products) and to identify actions on how to influence that development (e.g. own pricing, marketing). In order to be actionable the final analytics solution has to provide detailed answers on where and how to act to drive sales.
3. Predictive Analytics
Predictive Analytics takes into account the provided historical data used during causal analyses and delivers projections of possible future outcomes. That step takes advantage of specific algorithms to learn patterns from the past to forecast future outcomes. Several statistical or machine learning algorithms can help to accurately predict the future (e.g. Regression Analyses, Decision Trees, Neural Networks).
For example, a data scientist might want to predict the sales volume over the next 6-12. The prediction should include the actions a business stakeholder took (e.g. increased budget by 5%, shifted budget from TV to social media advertisement) to predict future sales outcomes. However, to evaluate the impact of possible future actions, plan resources and evaluate their outcomes, it is necessary to predict different scenarios (e.g. increase budget by 5%vs. 10% or shift budget from TV to social media advertisement by 10%, 20%, 30%) to recommend the “best” actions to take.
4. Prescriptive Analytics
The step of Prescriptive Analytics is building on top of the algorithm, which was built during the predictive analytics step. The automation and optimization of predicting the outcome of different actions is call prescriptive analytics. While the previous step is focused on predicting the future based on past trends and actions, prescriptive analytics provides insights on what to do to achieve a specific outcome (e.g. optimize Media Return on Investment). It recommends specific actions (e.g. decrease price by 5% in region A) and provides insights on their outcomes (e.g. price decrease will decrease sales by 2% but increase profit by 5%).
For example, the predictive model forecasts a significant dip in sales volume over the next 12 months. Causal analyses suggests that this might happen because of a planned reduction of media spend, which accounts for a significant incremental impact on sales volume. The specification of a desired sales volume outcome can than be used to evaluate the best actions (e.g. defining the total media budget and a possible budget allocation across media channels) to reach that sales target.
All four steps are important to build a successful AI solution: while the first two steps create trust, the next two create opportunity. I will show how that could look like for a sales use case. It is (in my experience) not possible to skip a step to create business value within an organization.
AI in Upsell-/Cross Sell Recommendation
A sales-representative has a challenging job! He/she has quotas to fulfill by interacting with customers and trying to sell a specific product or product mix. While that sounds pretty operational, there is a strategic element to it, which is crucial in order to be successful. In detail, the sales-representative has to:
- evaluate the sales opportunity of a (potential) customer,
- identify relevant actions that could increase sales volume,
- understand the customer’s needs to use those actions in order to increase sales volume.
Let us discuss that use case on a detailed example. A sales-representative for a consumer product company is selling his/her product portfolio to the supermarkets in his/her territory. The sales-representative is visiting 6-8 customers every day and trying to up-/cross-sell his/her products, trying to avoid out-of-stock situations for the company’s products, providing sales incentives to get more shelf space and using promotion material to make consumers aware of the company’s products sold in a store. The question, however, what is the sales opportunity for each store and what actions need to be taken to achieve that sales opportunity?
Descriptive analytics (usually based on the insights of Customer Relationship Management systems) will tell us more about the sales-representatives’ customers. Figure 2 is giving an exemplary overview of all the customers of a sales-representative. It provides information on sales revenue, store location and characteristics as well as actions taken by the sales-representative in the past. For example, we might be able to see that the sales-representative had a revenue of $55,000 last year with supermarket ABC by selling 40% of product 1, 40% of product 2 and 20% of product 3. Store characteristics include information on store size, which is 20,000 sq.ft., high competition in the area, high footfall traffic (describing the number of people passing the store) and data regarding the shelf space for the product category within the store. Information of actions taken by the sales-representative include the number of his/her store visits, delivery frequency and volume, promotions, and the average pricing compared to other stores in the territory (see Figure 2).
What we can see in Figure 2 is that there are stores with low and high revenue and that we have a lot of information about the stores and our relationship history. However, we do not know how to help the sales-representative with evaluating a customer’s opportunity and the relevant actions in order to get to a higher sales volume. In order to come up with hypotheses on how to help our sales-representative, let us gain some insights by comparing two stores with each other. Figure 3 is giving a direct comparison of two supermarkets. While supermarket ABC is accounting for $55,000 in revenue last year, supermarket XYZ is accounting for $155,000. The revenue difference is $100,000. The question is, if there is a sales opportunity of $100,000 for supermarket ABC? If yes, which of the actions we have taken in supermarket XYZ compared to ABC can increase our sales by $100,000?
We can see that the sales-representative sold a different product mix to supermarket XYZ, visited the store more often, had more frequent deliveries, gave a better pricing and had a lot more promotions. It would be too easy to assume we just need to catch up on all those actions to improve sales. So the question is, given the characteristics and location of the store, are we able to increase sales based on possible actions and which actions would have the biggest impact on sales?
In order to better understand the real opportunities, we need to estimate the impact of sales drivers by investigating their causal relationship on sales. Therefore, we can use machine learning models to evaluate the impact of individual drivers on sales. In detail, we use all information on store characteristics and historic actions to predict sales and to attribute sales to its causal drivers. Figure 4 (left side) is showing the overall impact of the different drivers on sales (grouped in non-influenceable and influenceable drivers). While non-influenceable drivers (context factors) account for the majority of the variance in sales between the stores, influenceable drivers can be used to drive a significant share of sales.
Both groups of drivers are important to help sales-representatives with his/her tasks, we described at the beginning of the chapter. First, it is possible to calculate a potential opportunity for each store considering non-influenceable store and location characteristics as well as influenceable actions taken by the sales-representatives. Figure 4 (right side) is showing the current revenue of a customer on the x-axis and the predicted possible revenue on the y-axis. Every customer above the 45-degree dotted line is providing an additional revenue opportunity (size of bubble is related to a store’s opportunity). However, the question is how to realize that opportunity?
While Figure 4 (left side) is giving an overall evaluation of sales driver importance, it is not providing insights on how to achieve the predicted sales opportunity for an individual store. The chart, however, is giving a good understanding of the general impact of environmental/situational factors and actions that contribute to the differences in sales between stores.
To better understand what has to be done in order to drive revenue and leverage the opportunity for a specific store, we need detailed insights on a store’s environment/situation and scenario analyses and insights (recommendations/prescription) on what actions to take. Figure 5 is providing insights on the sales opportunity optimization for supermarket ABC. The results show the difference of supermarket ABC compared to the average supermarket in its region. Supermarket ABC has a lower sales because it has a lower square footage, a higher competition in the area, a lower footfall and a smaller shelf space allocated to the product category than the average store in the region. Given those characteristics, the store should perform $66,000 below average. However, the lack of promotions, a higher pricing, less sales-representative visits and a lower delivery frequency account for -$33,000 in sales compared to the average supermarket.
Those very specific recommendations make it now possible for the sales-representative to plan a store visits one day ahead by ranking stores based on their sales opportunity and taking specific actions to improve sales on the most granular basis.
The value creation of this use case is especially challenging as it requires a high level of change management as it deeply affects the day-to-day working process of sales-representatives. However, the selected use case was just exemplary for the analytics journey, which can be applied to any use case out there. All analytical journey steps (Descriptive Analytics, Causal Analytics, Predictive Analytics and Prescriptive Analytics) are built on top of each other to eventually result in a business solution a sales-representative is hopefully going to trust and apply to his/her day-to-day work.
While there are use cases that require limited focus on cultural transformation/change management (e.g. mostly process-oriented), there are a lot of use cases that deeply affect the day-to-day job of business people and require their involvement and their trust in the provided solution to drive business value. The analytical journey approach can provide that from an analytical point of view. Furthermore, the analytical journey allows data scientists to better understand the context and the details of a business challenge, provides the opportunity to create hypotheses and possible actions to take from a business perspective, helps to identify relevant data and define the type of algorithms to use.