As the adoption of generative artificial intelligence (AI) accelerates, companies are shifting their focus from "which AI model to use" to "how to effectively integrate AI into their operations." However, despite recognizing the importance of operational design, many attempts to apply the same design across all industries persist.
For instance, approaches that have proven effective in finance are often directly applied to manufacturing, or data structures designed for retail are used in public institutions. The results are typically similar: while the technology is implemented, it fails to function effectively in real-world operations.
According to global research, the failure rate of AI-driven digital transformation (AX) can reach as high as 70-85%, with significant disparities in AI adoption rates across industries, such as tech and finance (80-90%) compared to manufacturing and agriculture (20-50%). Through various AX projects across different sectors, it has become clear that even the same design principles require different applications depending on the industry, as the work structure, data formats, regulatory environments, and the entities utilizing AI vary.
One common mistake in financial AX projects is prioritizing technology implementation over authority design. For example, automating contract reviews or implementing customer service AI often leads to issues during the operational phase due to unclear access rights and responsibilities, necessitating redesigns. Additionally, given the prevalence of scanned contracts and image-based financial and legal documents, it is essential to validate the parsing accuracy of document types and their conversion into AI-compatible formats beforehand.
In public institutions, the accuracy of complaint processing and the consistency of administrative procedures are paramount. Therefore, AI must be designed not only to generate responses but also to indicate the regulations and documents that support those responses, while managing response histories. Furthermore, having personnel review AI results and make final judgments is crucial for enhancing acceptance in the field.
A frequent error in retail AX is attaching AI before preparing the data. If product code systems differ across sources or if returns and exchanges are not accurately reflected in sales data, no matter how advanced the AI model, the results cannot be trusted. Most AI systems that are rejected in the field have skipped this critical step.
In a project involving an e-commerce-based retailer that Redbrick participated in, the first step before applying AI was data integration. Product data, demand data, and regional sales data were dispersed across different systems, and connecting them into a single pipeline was a prerequisite. On top of that, the processes for product recommendations and promotional decision-making were redesigned. This comprehensive approach transformed a task that previously required multiple people several days into one that a single person could complete in a day.
When initiating AX projects, Redbrick's first step is to understand the work structure, data environment, and regulatory context of the industry. Even when implementing the same functions, the designs for financial institutions and manufacturers must differ. Decisions on which areas to prioritize, how to structure data, and where to begin system integration must be tailored to the industry context.
Redbrick's Forward Deployed Engineers (FDE) are deployed on-site from the project's outset to conduct interviews with stakeholders, analyze actual documents and data samples, and observe workflows to diagnose points of repetitive tasks, bottlenecks, and errors. They then evaluate potential tasks based on work frequency, time requirements, data preparation, expected benefits, and security and regulatory risks to establish industry-specific priorities.
Pilots are designed to select a single task that can validate results within 4-8 weeks, establishing clear KPIs for processing time reduction, accuracy, and real-world usage rates. The FDE plays a crucial role in bridging the communication gap between operational staff and development teams, coordinating data refinement, authority structures, system integration, and validation processes. This approach creates an AX structure that not only implements technology but also ensures its integration and expansion in actual operations.
In this process, a full-scale transition is not attempted all at once. Pilots are first executed in areas with high effectiveness and low risk, confirming trust in the field before expanding. This sequence is critical in increasing the actual usage rate of AI post-implementation.
Many companies are already realizing that the success of AX projects depends more on how they are designed than on which AI model is used. However, this design must also vary by industry. Ignoring these differences and attempting to apply a single validated method across all sectors often results in technology being implemented but failing to operate effectively in practice. Successful AX projects do not begin with a good AI model; they start with a deep understanding of the industry and its specific context.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.