TLDR: A new research paper introduces a taxonomy for understanding the integration of Machine Learning (ML) into Robotic Process Automation (RPA), creating “Intelligent RPA.” The taxonomy categorizes this convergence across two meta-characteristics—RPA-ML integration and RPA-ML interaction—and eight dimensions, including architecture, ML capabilities, data basis, intelligence level, technical depth, deployment area, lifecycle phase, and user-robot relation. This framework helps clarify the complex landscape of intelligent automation, aids in evaluating current RPA products, and guides future developments, particularly with the rise of generative AI, by providing a standardized vocabulary for the field.
Robotic Process Automation (RPA) has become a cornerstone for businesses looking to automate repetitive, rule-based tasks, significantly boosting efficiency and cutting costs. However, traditional RPA, with its symbolic nature, faces inherent limitations when dealing with more complex tasks that require human-like cognitive abilities or the processing of unstructured data. This is where the integration of Machine Learning (ML) comes into play, transforming conventional RPA into what is known as Intelligent RPA.
A recent research paper titled “A Nascent Taxonomy of Machine Learning in Intelligent Robotic Process Automation” by Lukas Laakmann, Seyyid A. Ciftci, and Christian Janiesch, delves into this evolving landscape. The authors conducted a comprehensive literature review to explore the intricate connections between RPA and ML, ultimately organizing the joint concept of Intelligent RPA into a structured taxonomy. This taxonomy aims to provide a standardized framework for discussion, evaluation, and selection of RPA solutions that leverage ML, fostering a deeper understanding of the field’s potential and limitations. You can read the full paper here.
Understanding the Taxonomy’s Structure
The proposed taxonomy is built upon two meta-characteristics: RPA-ML integration and RPA-ML interaction. These two broad categories encompass eight distinct dimensions, each with its own set of characteristics, offering a detailed view of how ML augments RPA.
RPA-ML Integration: How ML Connects with RPA
This meta-characteristic focuses on the structural aspects of combining RPA with ML. It includes dimensions like:
- Architecture and Ecosystem: This dimension looks at how ML capabilities are linked with RPA software. Options range from external integration, where users develop and connect their own ML models, to integration platforms that turn RPA software into an open ecosystem for third-party AI providers, and ‘out-of-the-box’ solutions where ML is built directly into the RPA software or added via a provider-controlled store.
- Capabilities of ML: This dimension categorizes the specific ML functionalities utilized. Key capabilities include Computer Vision (for processing images, OCR, and recognizing UI elements), Data Analytics (for classification, pattern recognition, and processing already usable data), and Natural Language Processing (NLP) for understanding and generating human language, crucial for tasks like categorizing inquiries or enabling conversational agents.
- Data Basis for ML: Since ML thrives on data, this dimension highlights the type of data used for learning and applying models. It distinguishes between structured data (e.g., from legacy systems), unstructured data (e.g., documents, emails, audio files), which ML helps unlock for automation, and UI logs (internal system logs used for autonomous learning).
- Intelligence Level: This is a common classification in the literature, differentiating how intelligent the automation artifact is. It ranges from ‘Symbolic’ RPA (traditional, rule-based), to ‘Intelligent’ RPA (incorporating cognitive abilities like processing unstructured data, often referred to as ‘thinking bots’), and ‘Hyperautomation’ (where robots learn adaptively, manage, and improve autonomously, aiming for self-learning and generative capabilities).
- Technical Depth of Integration: This dimension considers the level of technical expertise required for integrating ML. It differentiates between ‘High code’ integration, requiring programming skills, and ‘Low code’ environments, which allow integration through UI-based modeling, making it accessible to power users in business departments.
RPA-ML Interaction: How ML-Enhanced RPA Operates
This meta-characteristic focuses on the augmentation of RPA in use, covering aspects like:
- Deployment Area: This dimension describes where the integrated RPA and ML solutions are applied. It includes ‘Analytics’ (for flexible decision-making), ‘Back office’ (for automating structured and unstructured document processing), and ‘Front office’ (for direct customer interaction, often with conversational agents).
- Lifecycle Phase: ML techniques can be employed at various stages of an RPA robot’s life. This includes ‘Process selection’ (identifying automation candidates), ‘Robot construction’ (autonomously deriving rules and building robots from user observations), ‘Robot execution’ (enhancing task execution with pre-trained models), and ‘Robot improvement’ (adapting to evolving UI interfaces or learning from mistakes).
- User-robot Relation: This dimension focuses on the role of humans during robot execution. It distinguishes between ‘Attended’ RPA (human-triggered, desktop-level interaction), ‘Unattended’ RPA (runs continuously in the background without human intervention), and ‘Hybrid’ approaches (collaborative partnerships between humans and robots, leveraging their respective strengths).
Also Read:
- Navigating Industrial Monitoring: A Look at Rule-Based Versus Data-Driven Systems
- CLIMB: A New Framework for Data-Driven Occupation Taxonomies
Application and Future Outlook
The authors applied their taxonomy to evaluate leading RPA products like UiPath, Automation Anywhere, and Microsoft Power Automate. They found that current practical applications primarily align with the ‘Intelligent’ RPA level, focusing on specialized cognitive capabilities, especially in the execution phase. While these products market themselves towards ‘Hyperautomation’, truly autonomous learning and adaptation are not yet fully realized.
The research highlights that practical reports show ML is predominantly used in the execution phase with a wide range of ML capabilities. Academic concepts, however, lean more towards hyperautomation and self-learning robots. With the advent of generative AI, there’s a noticeable shift towards enhancing the process selection phase, with co-pilot functionalities emerging for specifying or automatically constructing RPA robots. However, human interaction remains crucial.
This nascent taxonomy serves as a valuable tool for researchers and practitioners to systematically categorize and understand the rapidly evolving field of Intelligent RPA, helping to compare products and guide the design of future automation solutions.
