Introduction
An AI Agent is an intelligent agent based on a Large Language Model (LLM), capable of simulating human thought and behavior, thereby enabling autonomous decision-making and task execution. It can be understood in simple terms as an intelligent entity without a physical presence but capable of executing corresponding operations on computer software and hardware based on user commands. In November 2022, with the release of ChatGPT, large language models have been elevated to a level where they can provide practical assistance to people in their work and daily lives. This has triggered an explosive development in intelligent AI agents that rely on large language models as their cognitive foundation. As an experimental project launched in March 2023, AutoGPT garnered over 14,900 stars on GitHub within just five months of its release[1], setting a record for the fastest-growing open-source project in GitHub history.
What is an AI Agent
In the era of large models, an AI Agent comprises the following components: LLM (Large Language Model, serving as the core controller and providing core capabilities), planning abilities, memory, and tools. When a user request is input to the AI agent, it is initially processed by the Large Language Model. Through additional designed planning mechanisms, the task is decomposed into subtasks for better execution. Additionally, based on the importance of the questions, it is transformed into short-term or long-term memory, storing essential information for future use.
Once the AI agent obtains a detailed execution plan from the large language model, it can invoke external libraries such as search engines, databases, code executors, etc., to perform the decomposed tasks and return the results. As illustrated in Figure 1 below, an AI agent primarily consists of three components: planning, memory, and tools. The agent utilizes various planning mechanisms to efficiently break down and complete tasks, while transforming information into short-term and long-term memory. Memory assists in planning, and after planning is completed, external tools are called upon to execute a series of operations to achieve the user’s original intent.
Pros and Cons of AI Agents
AI Agents have the potential to revolutionize the way the software industry interacts and business models. The rapid advancement of large language models has provided a groundbreaking technological solution at the core of AI Agents. This development addresses the limitations of past reinforcement learning approaches, which were constrained by deep learning frameworks, allowing agents to acquire skills but lacking in the ability to acquire transferable understanding and knowledge. These shortcomings also left them deficient in background thinking.
Through mechanisms involving memory, planning, and the utilization of external libraries, large language models have transformed from mere advisors who could provide opinions on paper to service personnel capable of providing practical and personalized assistance to users. As depicted in Figure 2, the vision of a single individual leading a team of over a dozen AI Agents for collaborative and highly efficient automated office tasks is no longer just a fantasy.
While AI Agent applications are flourishing, there are still several challenges that need to be addressed. Foremost among these challenges is the issue of cost. Currently, most AI Agent applications require the use of GPT-4 interfaces as the underlying large language model for executing operations. However, invoking GPT-4 interfaces can be costly, and data privacy may not be adequately guaranteed. Therefore, there is an expectation for more powerful and large-scale models to be released and open-sourced to address this problem.
Another point of concern is the stability of current AI Agents. These agents may not exhibit consistent performance, and they tend to perform better and with higher accuracy when executing a single task. When trained to perform two or more different tasks simultaneously, the effectiveness and accuracy of the agent tend to decrease significantly. To handle multiple functions, it becomes necessary to design a separate agent for each function and manage them accordingly.
Potential Use Cases of AI Agents in the Financial Sector
While AI Agents are currently more commonly used in fields such as gaming and software development, they can also play a significant role in the financial sector. Within companies, AI Agents can assist in tasks related to email processing, meeting scheduling, meeting reminders, and more. They can automate tasks such as summarizing emails, drafting responses, and scheduling appointments, thereby reducing the time spent by staff on routine activities. One of the advantages of AI Agents is their ability to personalize and adapt to the progress of each individual, providing tailored assistance to employees on an individual basis.
In financial firms, AI Agents can be utilized to retrieve data from external sources, create charts, write reports, and run models, all of which can be easily delegated to the Agent. This can greatly save time that would otherwise be spent on tedious and lengthy tasks.
In the investment field, as illustrated in Figure 3, AI Agents can analyze market trends for stocks and futures by aggregating real-time web-based news and sentiment data, combined with price curves. They assist traders in making decisions by providing summaries and pushing critical information to traders. They serve as co-pilots or second brains for traders, aiding in the execution of trades. This enables traders to better anticipate potential market risks and reduce losses.
In the financial services industry, AI Agents can provide the best possible service to customers by offering a personalized and human-like interactive experience. According to Forbes magazine statistics, companies and organizations spend over US$350 billion annually on customer service, with suboptimal results. [2] Statistics from companies that have adopted AI Agents for customer service in 2021 show that after implementing intelligent customer service, the cost per customer query per round has decreased from US$5 to US$12 in the era of human customer service to just US$1. [3] Tech giant IBM estimates that adopting AI customer service has saved them approximately one billion US dollars in costs related to recruitment and productivity losses. [3] As depicted in Figure 4, the use of intelligent customer service enables round-the-clock assistance, addressing customer needs promptly, and enhancing user engagement. As technology continues to advance, the presentation of AI Agents can evolve from single-text interactions to various forms of interaction, such as voice, visuals, video, and digital interfaces, further enhancing the customer experience.
Future Outlook and Challenges of AI Agents in Enterprises
With the development and maturation of large language model technology, database technology, and toolkits, the application of AI Agents in the financial vertical domain is expected to grow rapidly. They will assist human decision-making and enhance operational efficiency. It is estimated that the adoption of conversational AI and intelligent solutions is poised to generate over 1.2 trillion US dollars in revenue in the next three years. [3] Numerous consulting and investment firms, both domestically and internationally, such as Deloitte and Ernst & Young, have entered the AI service industry and hold optimistic views on related investment directions. With an unlimited number of specialized AI Agents in their respective roles, enterprises can achieve the highest-quality enhancements at the lowest cost. Through the automation of repetitive tasks and processes, they can improve operational efficiency, productivity, and enhance customer experiences.
However, until reliable open-source large language models are widely available, data privacy and security issues remain critical in AI Agent operations. Building stable, trustworthy, and secure AI Agents will be the next goal for researchers.
Reference
[1] Significant-Gravitas/AutoGPT. GitHub Repository.
[3] Livevox. How Do AI Virtual Agents Lower Customer Service Costs?
[6] Deloitte US. Using AI to Improve Customer Experiences in Finance.
[7] Building Your Intelligent Financial Assistant AI Agent from Scratch.
[8] AI-Powered Customer Service: 10 Real-World Examples.
The work described in this article was supported by InnoHK initiative, The Government of the HKSAR, and Laboratory for AI-Powered Financial Technologies (AIFT).
(AIFT strives but cannot guarantee the accuracy and reliability of the content, and will not be responsible for any loss or damage caused by any inaccuracy or omission.)
