Algorithmic Retailing

Are Businesses Being Presumptuous in Demanding Greater AI Accuracy?

 
August 10, 2020

Artificial intelligence’s (AI’s) trailblazing signature characteristics such as speed, scalability, and transparency has led to its widespread adoption across every industry and every facet of retail. However, because of its ‘demi-god’ status, the most common expectation in deploying AI revolves around the notion of accuracy; i.e., does it perform the task accurately enough to be useful? Is accuracy a ‘vanity’ parameter to measure AI’s success? Do businesses have unrealistic expectations from AI?

Let’s explore the primary reasons that are driving businesses to demand greater accuracy:

  • Abundance of data: Enterprises collect colossal data through their internal ERP systems, sensor data deployed at stores, DC and telematics, etc. as well as external data from market intelligence partners. Deriving correlations among these data elements not only improves the accuracy of the models dramatically but also facilitates better explainability.  
  • Availability of open source algorithms: Revolutionary algorithms available as open source can deliver amazing predictability very quickly. Institutionalizing of deep learning algorithms deliver great results and reveal in-depth insights.
  • Hardware innovations: These innovations are not only allowing algorithms to run faster but have the power to run multiple algorithms at the same time, ensembling to deliver consistent results and predictions. For example, large retailers ran forecasting algorithms for upto 10 hours during weekends. Today, the same payload can be run in less than an hour, enabling retailers to run forecasts even twice a day bringing greater agility and capability to make/adjust decisions. Retailers are also able to generate intelligence for a given context and infuse the insights into the business decisions.

Should 100% Accuracy Be the Golden Standard?

Based on our experience, we observe a close correlation between enterprise maturity and expectations of AI accuracy. Based on their AI maturity, enterprises can be broadly categorized into Onlookers, Empiricists, and Exploiters. 

Onlookers constantly lookout for ideas and explore options for each context, but are unable to move forward due to various constraints. They look at AI as a generalized solution for automating their standard operating procedures where they expect 100% accuracy, for which many generalized tools (such as RPA and IPA) are available in the market.   

Empiricists are quite convinced about why they should attempt different machine learning (ML) models and a few initiatives take shape and deliver benefits. Though they look at 60-70% accuracy, eventually they want to increase accuracy to add more business value. For example, a consumer goods company measuring its sales performance at the end of every quarter could not make any course corrections midway due to insufficient data. However, with an early warning indicator system, the company can now consolidate its sales at the end of each week/month and use it as a projection for the succeeding months. This helps them to accurately forecast and make tactical and operational decisions, and appropriate adjustments to meet the target. For such use cases, a full-fledged model is not required and it is acceptable to have a model that gives 70-80% accurate results.   

The most prolific AI adopters are Exploiters; they enjoy an early mover advantage by virtue of having built foundational capabilities—enterprise data lake, tools and frameworks, and talent—and their primary focus is on building competitive advantage. They focus on customer-facing applications of AI and continuously invest in AI/ML technologies in the pursuit of greater accuracy. For example, a leading European grocery retailer having a current forecasting accuracy of 98% wants to go further to keep their forecasting accuracy as close to reality as possible. They want to drive towards 100% accuracy.   

In summary, there is a close correlation between enterprise maturity and accuracy expectations (Figure 1). As enterprises mature in their AI journey, they are less likely to take a narrow view on accuracy but are willing to make investments that can give them early indicators of emerging situations and fine tune their AI models to over time.

Figure 1: Enterprise maturity and AI accuracy expectations.

Expand

Conclusion

Most successful AI programs embark on sound understanding of the business problem and the goal of AI interventions, both domain and technical aspects; good access to enriched data; necessary automation in place for model development and deployment; high performing, scalable models; tools for accelerated model iterations; and complimentary in-house talent (domain and technology) augmented with external experts. By and large, the maturity of AI within any industry has been subject to massive disruption brought about by incremental innovations. However, quantum computing is poised to completely change the dynamics of AI. With results likely to be more accurate and faster, niche retail areas such as fashion, electronics, and jewelry will be presented with new opportunities for growth.

Nagarajan Karuppiah heads the ML and AI CoE (Retail). He has been with TCS for over 20 years and has held various leadership positions across retail accounts. An active member of IEEE, he has published white papers on diverse topics from enterprise data warehouses to decision sciences. Nagarajan is currently focusing on building ML solutions, NLP for personalization and recommender systems for retailers. He can be reached at LinkedIn:  https://www.linkedin.com/in/nagarajank/?originalSubdomain=in