Sequoia"The first act of Generative AI, the Era of Intelligent Reasoning Opens"

“Two years after the generative AI revolution, the field of research has developed from ‘fast thinking’-the pre-trained response of rapid response-to'slow thinking‘-the in-depth thinking of reasoning. This transformation is spawning a number of brand new smart applications”

Author of this article: Chris, Mixdesk AI Research Center

introduction

In 2024, the concentration of AI agent AI customer service in various fields will explode, and the value of generative AI in the enterprise application layer will rapidly expand. Global AI application layer companies will gradually occupy the market by providing scenario service tools with autonomous reasoning capabilities. In this issue, we will build onCustomer service scenario, The combination of AI Agent definition, working principle, and technology evolutionMixdesk AI Agent R&D practice discusses the development trend of AI Agent AI customer service.

AI Agent What is it? How does it work?

Definition and working principle

AI Agent, That is, the artificial intelligence body. In layman's terms, it is a highly anthropomorphic computing program. With the help of a “big model” brain, it can fully understand what is happening in its environment, think about problems by itself, make decisions, and perform tasks autonomously. Complete tasks. From the technical point of view, the AI Agent infrastructure is composed of 4 key parts: Planning, Memory, tools, and action.

Planning

LLM can give agents"Planning”The thinking mode of dismantling complex tasks, splitting them into multiple steps, thinking and solving step by step, and using technical tools such as RAG to make the output more accurate.

Memory Memory

Similar to humans“memory"Mechanism, the agent has short-term memory and long-term memory. The contextual memory of a single session will be stored briefly for multiple rounds of sessions and will be emptied after the task is completed; long-term memories such as user characteristic information, business information, etc. are usually stored and quickly retrieved in vector databases.

Tools Tools

The agent has the ability to use a variety of“Tools”The ability. For example, by calling the APIs of different application modules of the software system, the specified business information and the operating permissions to perform the business are obtained. By calling external plug-in tools, we can obtain capabilities that were not available in the original LLM.

Action action

The agent performs specific tasks based on planning and memory.“Action”, Complete specific tasks.

The relationship between AI Agent and large model and RAG

LLM large model、RAG Retrieval enhancement often appears with AI agents. What kind of connection is there between them in actual scene applications?

LLM can be seen as the “thinking center” of the agent. The agent uses the reasoning power of LLM to refine complex problems into multiple small problems and determine the order in which they are solved, that is, which problems need to be prioritized. Subsequently, the AI agent will call LLM, RAG technology, or external tools in this order to solve these small problems one by one. In this process, RAG can realize the addition of local knowledge base, real-time data, etc. to enhance the retrieval and generation capabilities of large models, and improve the quality of information query and generation.

Customer service AI Agent Application Value and Technical Difficulties

The same is the dialogue, why is the AI Agent so different?

Before the widespread application of large-model technology, customer service automation mainly relied onChatbot/Chatbot To achieve. Chatbots work in accordance with a pre-written dialogue process, usually repeating fixed replies, and AI Agent can perform intelligent reasoning.

with AI Agent The experience of a conversation is more like communicating with a smart customer service staff, while a conversation with a chatbot is like choosing answers from a menu of options. Traditional chatbots need to create a list of frequently asked questions and prepare scripted answers to each question. This process is time-consuming and difficult to scale. In contrast，AI Agent is like a digital employee with great potential, Can quickly access existing information sources, quickly learn and provide personalized solutions to help customers solve problems.

For enterprises, what is the immediate application value of customer service AI Agent?

In the field of marketing and customer service, the application of AI Agent is rapidly innovating the traditional model. The AI agent with mature technology can efficiently perform standardized and repetitive tasks in marketing and customer service scenarios, and handle some complex scenarios through human-computer collaboration.

Smarter pre-sales reception and lead acquisition：Online 7X24 hours a day, intent recognition, emotional analysis, fast and accurate response, can highly restore real pre-sales personnel, and improve the efficiency of pre-sales dialogue lead acquisition.

More efficient after-sales service and customer support：Automate standard tasks, intelligent routing, and human-computer collaboration to improve service efficiency, liberate more manpower, and enhance the customer experience.

Barrier-free overseas marketing and customer service：Automatic translation and switching of multiple languages, cross-time zone services, directly provide enterprises with the most practical tools for overseas business and global customer acquisition.

Technical Difficulties of Customer service AI Agent

With a large model, why are there so many kinds of agents derived? This is because the real world is complex, and every scene in a corporate event is also complex, and the real scene requires a lot ofSpecific application reasoning, These inferences cannot be effectively incorporated into a common model. This is the AI application layer mentioned in the article “The First Act of Generative AI: The Opening of the Era of Agent Reasoning”, which is constantly customizing the "cognitive architecture" of specific scenarios.

Taking the field of customer service as an example, whether it is the construction of AI agents in pre-sales or after-sales scenarios, it is not as simple as adding a user interface to the basic AI model. This is a very complex development system.，Facing different AI models is like facing a versatile team. Each member has their own strengths and needs to be integrated through an efficient coordination method, which is the routing mechanism.Other than that，Special databases are also used to store and manage information (RAG vector/graph database) to ensure data security and compliance. Most importantly, agents need to imitate the way humans think and help solve complex problems.

The technological Evolution of Mixdesk-the Birth of AI Agent

The content of this chapter comes from the transcript of the interview with the person in charge of the Mixdesk AI Research Center.

According to the general law of product development, product evolution usually follows three stages：

Initial stage：At this stage, the core functions of the product can be realized to meet the basic needs of users.

Maturity stage: With the continuous enrichment and improvement of functions, the product has begun to support a higher level of user interaction, including manual intervention and adjustment to improve the user experience.

Innovation stage：At this advanced stage, the product realizes independent learning and adaptation through intelligent technology to provide more personalized and efficient services.

Mixdesk product lines follow this evolution path, keeping up with cutting-edge technologies, gradually transitioning from the initial basic function realization to feature-rich and user-friendly optimization, and now embracing AI across the board, redefining scenario requirements, and promoting the next level of product development.

The first generation: the RASA Era

Rasa Open Source, An open source machine learning framework born in 2019. In early 2022, Rasa 3. With the release of x, this machine learning framework was quickly widely used in the field of intelligent customer service. Mixdesk also launched a commercial product based on Rasa customer service robot at that time, covering service pre-sales marketing and after-sales Q&A scenarios.

AI customer service at this stage mainly depends onIntent recognition and entity extraction technology, Able to handle a structured dialogue process. But it is not smart enough to handle complex and multiple rounds of conversations. For example, it is difficult to accurately understand the context, cannot handle complex user inquiries, and lacks flexibility when dealing with questions that are not within the preset range. When it is necessary to optimize recognition and adjust speech skills, it is necessary to supplement the corpus and retrain.

The second generation: SOP + LLM Agent

OpenAI The advent of the GPT-3 model marks the beginning of a new era. Based on this generation of technology, Mixdesk has developed a set of controllable and adjustable scenario-based SOP for the field of customer service, integrating the capabilities of large-scale language models. Since April 2024, the product form has changed from a single agent to multi-agent collaboration and Workflow management, to the current Agent combination tool + SOP + Agent Workflow integrated solution.

In the journey of technological exploration, we conducted extensive research and testing on a number of mainstream large-model platforms, including Dify、Coze, Alibaba Cloud Bailian, Volcano Cloud Ark, FastGPT, Flowise, RagFlow, hoping to find the most suitable To B solution for customer service scenarios. At present, Mixdesk has established in-depth cooperative relations with leading manufacturers in the industry, including Azure OpenAI, Tongyi Qianwen, Doubao, Claude, DeepSeek, Llama and so on. The speed of technical updates is faster than we expected. In 30 days in September, we launched 3 consecutive versions of the RAG solution. Before the 2.0 version was fully developed, we found a more advanced solution, so we decisively decided to abandon the 2.0 version and turn directly to the development of the 3.0 version in order to achieve the rapid evolution of technology.

So far, the AI functions of all Mixdesk product lines are open to new and old customers for trial.，Mixdesk AI Agent Capabilities have already played a role in some customers' first-line business scenarios. Old users can upgrade to the latest AI version in place, and new users can experience our new version of AI customer service functions for free, and feel the convenience and efficiency brought about by technological innovation.

The future of AI Agent

In the construction of Open AI, Agent is a stage of the development of the AI system：

It has evolved with the development of AI technology. What can be seen in the future is that AI Agent will integrate natural language processing, computer vision and speech recognition technologies.Multimodal interactionTo achieve breakthroughs and provide more natural and efficient interaction. Regarding the evolution route of AI Agent, we cite a chart to show：

Temporarily unable to display this content outside the Lark documentation

In the higher development stage of the agent in the future, it is able to self-organize, self-make decisions, and self-schedule to complete complex tasks. This level of ability will be most directly reflected in the customer service scenario.

summary

The skill groups in traditional manual customer service: pre-sales consultation, product recommendation, transaction promotion, order processing, after-sales service, complaint management, technical support, feedback collection, etc. Correspond to different agent capabilities. In all these links, AI can self-organize agents with different skills to complete the reception work 7*24 hours a day.This should be the ideal state of intelligent customer service for all customers, and this is alsoMixdeskThe way forward, “We are not sure how far this future is, we only know that every step forward, we are closer to this future. ”

Maybe you want to see