Guest Speaker｜Hu Yichuan (PhD), Co-founder & CTO of Laiye Technology
The content has been simplified. For the full video transcript and presentation materials from the expert, please scan the QR code.
RPA (Robotic Process Automation) is, in simple terms, the automation of tasks such as software and system operations, data processing, etc., by simulating keyboard and mouse operations. It is suitable for repetitive tasks with clear logic and rules.
01
Why Has RPA Developed Rapidly in the Past Five Years?
1、Non-intrusive，RPA does not rely on other systems’ open interfaces and can achieve automation based on existing interfaces and screens. It can perform any work on the computer interface non-intrusively.
2、Easy to Use，RPA is essentially a low-code UI automation and system integration tool. Its widespread adoption today is due to the fact that non-technical business personnel can use RPA to automate repetitive tasks.
3、Accurate and Stable，RPA can operate 24/7 without interruption, and its success rate is very high.
02
What Problems Does RPA Solve?
The above image typically illustrates the current state of digital construction in most enterprises. Over the past decade or two, almost every enterprise, especially large ones, has been undergoing digitalization, implementing many systems, hoping to run traditional businesses digitally and accumulate data. Although this goal has been achieved, the years of construction have resulted in many systems within the enterprise.
Statistics show that large enterprises commonly use hundreds of business systems internally. This can lead to a problem, for example, systems A and B in the image may have been built at different times and by different vendors, with no connection in between. However, business often requires connecting these two systems, and that task falls to humans.
For instance, when systems A and B were designed, they did not consider how users would solve specific business tasks, leading to situations where system A does not cover non-standard processes, forcing users to complete non-standard processes outside the system and then import the data back into system A.
System B only considers standard processes, but as business volume increases, the frequency of repetition skyrockets. This results in the need to spend a lot of time every day on repetitive processes in system B.
In the past five years, RPA has been able to scale in large enterprises and organizations because it effectively connects employee systems and data. On this basis, RPA has defined a new software form—digital employees.
Digital employees have become a bridge between employees and systems, employees and data, and systems and data.
1、Operate Complex Systems
2、Process Various Types of Data
3、Interact and Collaborate with Employees
With digital employees, most tasks involving systems and data can theoretically be completed by them, relieving employees from tedious and repetitive work and allowing them to utilize internal knowledge and data more efficiently. This is also a significant reason for the rapid development of this new category of software in the past five years.
As enterprise systems and data continue to grow, how to connect systems and use data more efficiently becomes a new challenge. RPA emerged before the popularity of large models and even AI. Today, with large models, especially large language models, what is the relationship between RPA and large language models?
The image above describes the state of work for white-collar workers in the office. RPA acts as the hands of employees, automating repetitive, rule-based, structured data tasks. However, the development cost is relatively high, as each RPA process is customized and has strong task execution capabilities.For example, as mentioned earlier, it is non-intrusive and can operate any system and data based on interfaces and graphical interfaces.
Today’s large models, especially large language models, resemble the human brain, being probabilistic and better at handling unstructured data, with lower development costs. Trained large models exhibit strong capabilities in language understanding, generation, and content generation, requiring no development, and possess strong reasoning and decision-making abilities, which traditional RPA lacks.
04
Three Stages of Scaling Digital Employees from RPA to AI Agent
From two dimensions, we define the three stages of digital employee implementation. The x-axis represents the scale of digital employees, indicating the scope and number of digital employees that can be covered. The y-axis represents the complexity of the business being processed.
The first stage is RPA, with the keyword being automation. The greatest value of this stage is to replace humans in automatically completing high-frequency, repetitive, process-oriented tasks, such as data scraping, data entry, cross-system operations, etc.
A typical application scenario is in the enterprise back office, such as finance, which is the most typical and suitable business department for implementation, including human resources, legal, IT departments, etc. However, the covered scenarios are relatively limited, and the scale of empowered employees is relatively small. Additionally, since RPA operates based on rules, it must be instructed on how to operate to automate tasks, thus only handling tasks of relatively low business complexity.
The second stage is RPA + AI, with the keyword changing to intelligence. This stage primarily allows RPA to leverage AI capabilities, especially in language understanding and generation, to extend RPA’s boundaries, achieving a transition from automation to intelligence. RPA can handle natural language, text, unstructured data, and generation tasks without being entirely rule-based, but it remains repetitive work, and typical application scenarios can expand to customer service, marketing, etc.
The third stage is AI Agent, with the keyword being human-machine collaboration. This stage further releases model reasoning and multimodal capabilities, equipping each employee with an intelligent assistant available 24/7 in all scenarios. This is the ultimate form and will take a long time to achieve. Once this stage is reached, digital employees can almost fully cover existing business scenarios and automate complex business tasks.
05
Two Major Practical Directions for RPA + AI.
1、Reduce the Development Cost of RPA
With the generative capabilities of large models, especially code generation capabilities, the development cost of RPA can be significantly reduced. For example, empowering developers to generate process code, plugins, and documentation can greatly enhance overall development efficiency.
2、Expand the Application Scenarios of RPA
Originally, RPA could only automate rule-based processes; now it can utilize large models’ understanding and reasoning capabilities to perform tasks based on probabilistic models, such as processing unstructured documents and data analysis. This can expand the application scenarios of RPA.
Empowering RPA with large models primarily involves code generation, document processing, data analysis, and other areas requiring large model capabilities.
The leap from RPA + AI to AI Agent is significant. Our platform can enable enterprises to build their own AI Agents today, with a design concept divided into three layers.
The first layer is to achieve understanding and interaction of dialogue, which corresponds to the interaction between digital employees and employees, collaborating on tasks.
The second layer is to perform task orchestration and execution. After obtaining tasks through interaction, it involves decomposing, orchestrating, and executing tasks.
The third layer will utilize several tools, such as using tools, processing documents, querying enterprise knowledge, data analysis, content generation (text, code), etc.
Digital employees essentially require multiple skills, and the design concept of the AI Agent platform specifically includes the following four aspects.
1、To aggregate various skills required by the AI Agent through an integrated platform, enabling more efficient development of digital employees.
2、Task orchestration based on objectives is a significant difference between AI Agent digital employees and RPA digital employees. Traditional RPA digital employees essentially lack the concept of task orchestration, or they simply follow rules, where humans instruct them on what processes and steps to follow. However, in AI Agents, this must rely on models.
3、Achieving human-machine collaboration through natural language interaction is key to AI Agents completing complex tasks. Today, using AI Agents in enterprises does not require them to complete complex tasks 100%; instead, it allows employees to collaborate with AI to accomplish complex tasks, greatly enhancing efficiency.
4、Normal employees need to go through processes such as selection, use, training, and retention. The corresponding design, development, usage, and evaluation for digital employees also require methodological and systemic support.
The image above is an overview of the AI Agent digital employee platform architecture. Based on the previous design concepts, it is divided into four layers.
Model Layer, which connects various closed-source and open-source large models, supporting model fine-tuning to meet different enterprises’ large model integration needs.
Capability Layer, which includes the capabilities required for digital employees to execute tasks, covering UI automation and system integration. This includes intelligent document processing, intelligent Q&A, intelligent forms, content generation, etc. The capability layer continuously expands horizontally, with the aforementioned capabilities existing as atomic capabilities that can be utilized and orchestrated by AI Agents.
Management Layer, which requires managing the entire lifecycle of digital employees, including demand management, development process management, sharing of digital employees and skills, operational management, especially status and performance evaluation.
Interaction Layer, which integrates a conversational user interface supporting multimodal and multi-channel interactions.
Many enterprises face a challenge today: while large models are indeed beneficial, how can they be implemented within the enterprise? We believe this architecture provides the “last mile” for implementation, helping enterprises better deploy large models through the model layer, capability layer, management layer, and interaction layer, while also assisting enterprises in developing their own AI Agent digital employees based on large models.
For digital employees to be scaled and truly transition from RPA to AI Agent, what organizational upgrades are necessary? The key is that the scaling of digital employees also requires corresponding organizational upgrades.
The image above illustrates the organizational upgrade of the digital employee excellence center achieved through consulting, design, and construction in actual client service processes.
On the left is the training center, empowering various business and IT departments from headquarters to branch offices through various online and offline training. The middle consists of the empowerment and promotion team, which includes demand response, development, operation and maintenance testing, digital employee management, etc. On the right is the partner management team, where the development and capabilities of digital employees may be provided by multiple vendors, necessitating a mature partner management mechanism.
This poses new requirements for organizations. In the future, when an enterprise intends to utilize digital employees on a large scale, it must consider not only products, technology, and solutions but also how the corresponding organization should upgrade.
06
Three Application Scenario Case Studies
Case 1: AI Agent Empowers Insurance Agents
The work of insurance agents essentially involves maintaining communication with clients, selling insurance products, and answering client questions. Based on this definition, AI Agent can empower insurance agents in three major scenarios.
The first major scenario is making recommendations. Insurance products are inherently complex, and each agent has different understandings of the products based on their experience and expertise. AI Agent can leverage the capabilities of large models to help agents find insurance products that are more suitable for their clients, mainly utilizing RPA and the large model’s long-text reading and understanding capabilities.
The second scenario is providing specialized Q&A based on insurance products. Through large models, combined with the knowledge base of the enterprise’s insurance products, professional and precise knowledge Q&A can be achieved.
Lastly, AI Agent assists insurance agents in automatically maintaining relationships, conducting regular follow-ups, and even helping employees fill out policy forms. The core work of insurance agents is to maintain good client relationships, and as the number of clients increases, the time cost of maintaining these relationships will rise, potentially degrading the client experience. By combining RPA and AI Agent, proactive inquiries and even proactive task execution can be achieved.
The above scenarios can realize three major business values:
1、Providing clients with more professional and personalized services, enhancing client satisfaction.
2、Increasing insurance order conversion rates and order amounts.
3、Enabling insurance agents to enhance their service capabilities, allowing them to serve more clients.
Case 2: AI Agent Empowers Financial Reimbursement Work
RPA can automate the filling of overseas business trip reimbursement forms. Now, RPA combined with large models can effectively understand the content of these multilingual, unstructured documents and input the corresponding fields into the reimbursement system, even understanding screenshots from mobile phones, with the entire process running in the cloud.
Case 3: AI Agent Covers HR Recruitment Processes
From job postings to resume screening and invitations, to candidate evaluations, AI Agent can achieve human-machine collaborative work, covering almost the entire HR recruitment process.
07
Three Challenges in Implementing AI Agent in Digital Employee Scenarios.
First, the model capabilities need enhancement.
Second, the API ecosystem needs improvement，Whether digital employees can assist employees in executing numerous tasks ultimately depends on the richness of the underlying operating system’s task execution capabilities.
Lastly, and we believe most importantly, the implementation of AI Agents in enterprises must ultimately take the form of human-machine collaboration. This means that many processes need to be optimized or even reconstructed. For example, in the previous recruitment example, what should HR do with and without the Agent’s assistance? The approaches differ significantly. Thus, optimizing and reconstructing business processes becomes extremely important.
This concludes the content of this sharing on the large-scale implementation of AI Agent digital employees. For the full video transcript and presentation materials from the expert, please scan the QR code.
⩓
Long press the QR code to receive the full video transcript and presentation materials
Member of the Technical Frontier Committee of the China Computer Association, Visiting Professor at the School of Artificial Intelligence, Wuhan Engineering University, with a master’s and PhD from Tsinghua University. He has over 15 years of research and entrepreneurial experience in machine learning and artificial intelligence, published more than 20 papers, and holds over 200 patents and patent applications. He co-founded the personalized video recommendation product “What to Watch Tonight”, which was acquired by Baidu.
Note: Click the lower left corner“Read the Original” to receive the full expert transcript and shared materials.
