Shrestha, Y. R., Krishna, V., & Von Krogh, G. (2021). Augmenting organizational decision-making with deep learning algorithms: Principles, promises, and challenges. Journal of Business Research, 123, 588–603. https://doi.org/10.1016/j.jbusres.2020.09.068
Summary
Organizations have long struggled with processing vast amounts of information to make effective decisions. Over the decades, scholars have looked at how technology—first through traditional information technology and later through artificial intelligence (AI)—can enhance decision-making. In recent years, advances in deep learning (DL), a subset of AI, have shown promise in augmenting decision-making processes. Shrestha et al. (2021) explore how DL algorithms can support and transform decision-making within organizations by automating and improving various steps of the decision process, while also discussing the promises and challenges that come with this transformation.
Deep Learning–Augmented Decision-Making (DLADM)
The authors introduce the concept of Deep Learning–Augmented Decision-Making (DLADM) as an approach where outcomes from DL algorithms—such as predictions or insights derived from unstructured data—are integrated into the decision-making process. Unlike earlier technologies that required heavy manual input (such as feature engineering), DL algorithms are designed to automatically extract and learn relevant patterns from data, including images, text, and videos. This makes them particularly effective in processing the “big data” that modern organizations generate.
Historical Context and Evolution
Decision-making in organizations has historically been influenced by theories like bounded rationality, which posits that managers operate under constraints in information and processing ability. Early IT systems helped alleviate some of these constraints by improving communication and data storage. However, the explosion of unstructured data in today’s digital world requires more sophisticated tools. DL algorithms have emerged as a solution by not only speeding up data processing but also by offering more nuanced insights through techniques such as image recognition and natural language processing (NLP).
How DLADM Works
The article explains that DLADM can be integrated into organizational decision-making through a multi-stage process. These stages can be broadly divided into three parts:
- Data Stage:
In this first phase, organizations identify the decision-making problem and gather data. DL algorithms are particularly useful here because they can automatically process unstructured data (such as customer reviews, social media posts, or product images) without the need for manual feature engineering. This stage is crucial for converting raw data into a form that is useful for analysis. - Learning Stage:
Once data is preprocessed, the next step is to select and train a deep learning model. The authors provide a tutorial-style overview of DL architectures, including neural networks (NNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). These models are chosen based on the type of data and the decision task at hand. For example, CNNs are particularly effective for image classification tasks, while RNNs are better for analyzing sequential data such as text. The learning stage also involves fine-tuning model parameters, choosing loss functions, and optimizing performance through techniques like transfer learning—where pre-trained models are adapted to specific tasks, thereby reducing training time and data requirements. - Evaluation Stage:
In the final stage, the performance of the trained DL model is evaluated using test data. The goal is to determine how well the model generalizes to new, unseen data. This evaluation process is iterative; if the model does not perform to expectations, adjustments are made in earlier stages. The authors stress that this iterative cycle is fundamental, as it enables continuous improvement of the model and, by extension, the decision-making process.
Case Studies in DLADM Applications
The paper illustrates DLADM with two detailed case studies drawn from real-world datasets:
- Case Study 1: Fashion Image Classification
In this example, the authors use the Fashion-MNIST dataset—a collection of grayscale images of apparel—to demonstrate how DL (specifically, CNNs) can classify product images into different categories. The study compares traditional machine learning (ML) methods with DL approaches. The results show that DL models (e.g., a variant of VGGNet and transfer learning with ResNet18) significantly outperform traditional ML algorithms in terms of accuracy (with accuracies reaching over 92–94% compared to around 78–85% for traditional methods). For a fashion retailer, such insights can drive better product recommendations, inform design decisions, and optimize marketing strategies by accurately identifying trends and customer preferences from visual data. - Case Study 2: Textual Sentiment Analysis
In the second case, the focus shifts to textual data using a movie review dataset from Rotten Tomatoes. The goal is to classify the sentiment of review sentences as positive or negative. Here, the authors compare traditional ML approaches (like logistic regression and SVM) with DL methods such as a multilayer perceptron (MLP) and fine-tuning a pre-trained BERT model. Again, the DL approaches achieve higher accuracy (around 90%) compared to traditional methods (around 76–79%). This type of sentiment analysis can empower managers to quickly understand customer feedback and market sentiment without manually sifting through thousands of reviews.
Promises and Challenges of DLADM
DLADM holds considerable promise for augmenting decision-making by providing:
- Enhanced Information Processing: DL algorithms can digest massive volumes of unstructured data, uncovering patterns that human analysts might miss.
- Improved Prediction Accuracy: By learning from historical data, DL models can provide accurate forecasts, which can be crucial in marketing, finance, HR, and operations.
- Faster Decision Cycles: With automated data processing and continuous learning, organizations can respond to changes more quickly, which is essential in today’s dynamic business environment.
However, the article also highlights several challenges:
- Economic Costs: Cutting-edge DL models require significant investments in data collection, annotation, and computing infrastructure. Models like BERT require massive computational resources, which can be costly.
- Organizational Challenges: Implementing DLADM demands new skill sets in data science and changes in organizational processes. Managers need to be aware of the risks of algorithmic bias, errors, and the opacity of deep learning models. These “black-box” models can be difficult to interpret, making it challenging to justify decisions to stakeholders.
- Ethical and Governance Concerns: The potential for bias in DL algorithms can lead to unfair outcomes, and increased data collection may raise privacy and security issues. Firms must set up robust governance frameworks to monitor and mitigate these risks.
- Need for Human Oversight: Despite the advances in DL, human judgment remains critical. Managers must balance the insights provided by DL models with domain expertise and ethical considerations.
Integration into Organizational Decision-Making
The authors propose that for organizations to fully benefit from DLADM, they must adopt an iterative approach that integrates DL outputs into existing decision-making processes. This involves:
- Establishing clear business questions and aligning DL projects with organizational goals.
- Iteratively refining data collection and preprocessing strategies.
- Selecting appropriate DL architectures based on the type of decision problem.
- Continuously evaluating model performance and updating models as new data becomes available.
- Ensuring transparency and accountability in how DL-driven decisions are made, which helps build trust among employees and stakeholders.
Ultimately, the paper argues that while DLADM can significantly enhance decision-making by augmenting human analytical capabilities, its success depends on the organization’s ability to manage its economic, technical, and ethical challenges. Managers must be proactive in building the necessary capabilities and governance structures to harness the benefits of deep learning while mitigating its risks.
10 Practical Insights for Business Owners and Managers
- Embrace DLADM as a Decision Tool:
Use deep learning models to augment, not replace, human decision-making. They can sift through massive unstructured data, helping you uncover trends and insights that would otherwise be missed. - Leverage Transfer Learning:
Adopt pre-trained models (like ResNet or BERT) to reduce training time and data costs. This approach allows your organization to quickly deploy effective models without needing extensive in-house expertise. - Integrate Iterative Feedback:
Build processes where decision-making is continually refined based on model outputs. Use feedback loops to improve data collection, model selection, and evaluation strategies. - Invest in Data Infrastructure:
Ensure you have robust data collection and processing systems in place. High-quality, annotated data is crucial for training effective DL models and gaining reliable insights. - Foster a Data-Driven Culture:
Train your team to understand and interpret DL outputs. Encourage collaboration between data scientists and domain experts so that DL insights are effectively integrated into strategic decisions. - Prepare for High Accuracy and Fast Responses:
Use DL algorithms to enhance predictions—whether it’s classifying product images or analyzing customer sentiment—so that you can quickly respond to market changes and customer feedback. - Plan for Economic and Technical Costs:
Be aware of the investments required for state-of-the-art DL, including computing power and specialized personnel. Consider using cloud-based solutions or AutoML tools to manage costs. - Address Ethical and Bias Concerns:
Develop governance frameworks to monitor for bias and ensure transparency in your DL models. This helps maintain trust among employees, customers, and stakeholders. - Balance Automation with Human Oversight:
Although DL models offer high accuracy, human judgment is essential to interpret results and make final decisions. Establish “human-in-the-loop” systems to review and adjust model recommendations. - Align DL Initiatives with Business Goals:
Ensure that every DL project addresses a clear business need—whether it’s improving customer targeting, optimizing operations, or enhancing product development. This alignment ensures that technology investments translate into real competitive advantage.
Closing Thoughts
Shrestha et al. (2021) illustrate that deep learning can significantly augment organizational decision-making by automating the processing of complex, unstructured data and delivering insights that drive faster, more informed decisions. However, realizing these benefits requires addressing both the technical and organizational challenges inherent in DL adoption. By investing in data infrastructure, fostering a collaborative data-driven culture, and implementing strong governance frameworks, managers can harness the power of deep learning to secure a competitive edge while mitigating risks.