How to improve data quality to empower business decision-making process and business strategy agility in the AI age

Introduction

Data Quality and Availability should always be the initial point of consideration prior to any machine learning model implementation (Dhar, 2013). Accuracy and reliability depend heavily upon quality data input into an ML model’s prediction engine (Hastie et al., 2009; Serhani et al., 2021). Yet, many businesses struggle with incomplete datasets, data silos, and issues like consistency deduplication completeness, which make developing accurate predictive models challenging (Fan and Geerts, 2022). Remembering that poor Data quality negatively influences innovation competency and creates what has been classified as the “Cost of Poor Data Quality (CPDQ)” (Ghasemaghaei and Calic, 2019; Radziwil, 2020).

To overcome data quality challenges, companies should implement robust data governance and management policies that establish rules and guidelines for handling information accurately, completely, and organized. Organizations can then maintain data integrity while increasing output quality with such policies (Khatri and Brown, 2010).

Recognizing potential conflicts between Data Quality Management and Security can also be worthwhile. Due to data sets’ large sizes, heterogeneity, legitimacy of sources, flexible read/write access requirements of Security systems, and implementation of Data Quality Management could pose potential conflicts with each other. Although Data quality differs considerably from security, both require accuracy, completeness, and consistency with integrity properties such as accuracy; an IT organizational strategy must focus both on Data quality challenges while simultaneously developing effective security systems (Talha et al., 2019).

Solutions to guarantee data quality: The funnel

• Data Pre-processing or Cleansing:

Data cleansing is the critical first step in creating machine learning models. Data cleansing entails eliminating errors or inconsistencies from data to make it reliable for analysis; normalizing brings it all into a standard format to make comparison easier; integration brings in data from various sources in ways that make sense for analysis; finally, data fusion represents merging multiple sources into one coherent analysis (Brownlee, 2020; McCarthy et al., 2022).

• Data-as-a-Service (DaaS):

Recent efforts and proposals attempting to ensure data quality from raw sources for Machine Learning and Artificial Intelligence have resulted in the concept of Data-as-a-Service (DaaS), where users receive data without knowing its source, hence requiring continuous Data Quality Management processes using Machine learning models for quality management (Azimi and Pahl, 2021).

• Explainable AI: A Possible Means for Increasing Trust in Advice:

As for the utilization of the data, other factors must be considered that can create issues in building trust due to the lack of transparency and possible misinterpretation of the data, which can undermine the ML and AI adoption related to the decision-making process (Tiwari, 2023).

To tackle the interpretability challenge, businesses can adopt Explainable AI (XAI) techniques (Dhar, 2012). XAI provides clear and understandable explanations, not only giving advice but also clearly explaining the reasoning behind it, making trust higher because there is a better understanding of how elaborations arrived there. Considering the sophistication of the AI models, the need for interpretable and trustworthy AI grows. Consolidated techniques are available for XAI: LIME and SHAP are two methods used to explain machine learning models. LIME presents single predictions by a simple model, making it quick and easy to understand but less reliable. SHAP is more accurate in supplying explanations for predictions. However, it works on global interpretations and is considered ideal for understanding patterns and variable relationships.

To leverage XAI effectively, it is necessary to develop internal capabilities for interpreting and communicating AI-driven insights. Training employees, managers, and executives to comprehend the AI models’ outputs and interpret the explanations will empower decision-makers to act confidently on the recommendations.

There is a compelling proposal from De Bruijn et al. (2022) elaborated for Governmental institutions about the potential strategies to consider to validate the XAI. They are compelling to be transposed to business and impact the Data quality funnel as they consider data quality management from a different angle: the algorithms.

With the aim of sensemaking of the Data quality funnel, we have elaborated on the strategies of De Bruijn et al. (2022). Three levels are proposed as in Figure 1:

1. Operational Challenges do not Intervene Directly on Algorithmics: The proposal is to challenge them following diverse approaches, which can require intermediate levels to execute controls and eventually return to the original algorithm to be tuned or changed.

OPEN IN VIEWER

Figure 1.

Challenging levels strategies to implement and use XAI.

Research in this area is still in its infancy regarding findings and their application in business and public sectors. We propose a specific hierarchical structure that highlights roles with different levels of responsibility and accountability. Our proposed model requires integration between AI experts, managers, and executives. These responsibilities are diverse and different before and after the outcomes of AI’s decision-making processes.

Figures 2–4 present a visualization of the possible roles in a hypothetical classic hierarchical structure. We have elaborated on the findings from the previous studies (Clarke, 2019; De Bruijn et al., 2022; DeMott, 2013; Hofbauer and Maier, 2021; Juho, 2018, 2019; Sanderson et al., 2022; Schäfer et al., 2022).OPEN IN VIEWER

OPEN IN VIEWER

Figure 3.

Existing roles at management levels for AI governance (ethics, governance, Privacy, Risks).

OPEN IN VIEWER

Figure 4.

Existing roles at operational levels for AI governance (ethics, governance, Privacy, Risks).

The funnel is also shown in Figure 5.OPEN IN VIEWER

Cloud computing and distributed systems

Companies should use cloud computing with machine learning platforms that can better manage massive amounts of data (Lenarduzzi et al., 2021; Sadlier and Baksh, 2022). Leaders must focus on solutions that can support the company’s agility and ensure the capacity to evaluate quick answers to adapt to changing markets, consumer requirements, and trends (Liu et al., 2018). Then, the IT department and data teams have to determine the risks and possibilities to improve machine learning and resource usage.

Moderating factors for implementing data quality funnel for decision-making

• Data culture and organizational culture: integrating Data culture and Leadership role expected.

Establishing a data culture within an organizational culture is of vital importance in creating successful business strategies, particularly considering that start-ups rely heavily on data from day one (Davenport and Bean, 2018).

Establishing a data-driven culture requires transactional Leadership and clear communications that emphasize the benefits of data-driven decision-making and its outcomes.

Resistance to new technology, inadequate focus on providing actionable analysis, rigid organizational structures, and skill deficits in non-data-centric departments can be expected. Still, without a data-driven culture, moving towards ML and predictive analysis will fail (Storm and Borgman, 2020).

• Trust in ML and AI outcomes

Machine learning and AI often come under fire for failing to take security into account in their designs; training data could be falsified, predictions altered, and models used to access sensitive information within training data sets accessed, all due to insufficient security considerations being integrated into its algorithms’ designs and utilization (Arrieta et al., 2020). Thus, activities to make ML and AI trustworthy are to be in place, even if, so far, the adoption in practice is low. With widespread applications, new solutions will take care of this issue (Eshete, 2021; Serban et al., 2021).

• Organizational Technical Resource Capabilities:

Incorporating ML and AI into businesses requires technical and organizational capabilities, including assessment frameworks. One of the elements to consider for choosing an ML provider, if it is external, is understanding their model design, debugging, and maintenance across diverse applications and impacts on various stakeholders (Yang et al., 2022). This is related to the engineering phase, while processes to define a robust production and deployment of ML is another related area that is part of the technical organization resources capabilities (Lavin and Renard, 2020).

To enhance resource capabilities, it is essential to handle new data sets. When creating models for managing these data sets, resources such as data knowledge, statistics, and internal services are optimized to establish shared feature spaces that can reduce development lead times. Also, they can mitigate risks associated with inadequate controls in previous models.

In conclusion

We firmly believe that using the power of Machine Learning (ML) and AI as decision-makers can be a transformative step for businesses. However, companies must address critical points about data and, as antecedents, companies’ capabilities, including Leadership, Data culture, new roles to challenge AI results and algorithmics related to the decision-making process, ethical principles, and inclusivity are needed, empowering technical resources capabilities.

Following the Data Quality for Decision-Making Processes Funnel, companies can gain significant competitive advantages for sustainable growth using ML and AI.