How can enterprises use LLMs efficiently? Full-stack intelligent AI is the main form of enterprise-level AI

01

—

Last year, I visited a financial shared service center of a large enterprise and saw a group of people bending over and hunching over the one-meter-long workstations. I suddenly remembered a fully automated digital factory ( " dark factory " ) that I visited a few days ago. There , only the equipment was running in an orderly manner, and occasionally some technicians would appear between various equipment. The scene in front of me was in sharp contrast to the automated factory. Modern manufacturing has completely realized the automation of physical labor, but the automation process of mental labor (even simple repetitive mental labor) has been stagnant. The emergence of full-stack AI agents may completely change all this, completely liberate people from simple repetitive mental labor, and greatly reduce the management costs of enterprises.

Limitations of AI Agents in Enterprise Applications

Various AI agents based on production AI for consumers and professionals have emerged in large numbers this year, from OpenAI 's DeepResearch to the " general agent " Mannus , which became popular at the beginning of the year . They focus on quickly collecting and organizing information and writing long articles of 10,000 words. At first glance, they seem to be able to replace humans to carry out work. However, from my own actual use, I found that these agents can often write long articles of 10,000 words but cannot write three sentences of insight, and the output results are variable. The answers given to the same question are different when asked several times. The quality of the article is mainly a blind box card draw, relying on luck. Obviously, such an agent still cannot replace humans to work independently, and at most it is an efficient data sorting assistant. At the end of 2024  , the non-profit organization METR compared the performance of human experts with 13  cutting-edge AI  models (such as the GPT  series and Claude 3.7 ), and found that AI  has a success rate of nearly 100% in tasks that take humans less than 4  minutes , but a success rate of less than 10% in tasks that take more than 4  hours . At present, the stability and reliability of independently working AI agents based on generative AI are far from meeting the requirements of enterprise-level applications.

How can enterprises make full use of AI large language models at this stage ? Full -stack intelligent AI that focuses on global process automation will help enterprises achieve " automation " and " autonomy " of the entire business process with AI as the core , while supporting the collaboration between people and AI to systematically improve the overall efficiency of the enterprise.

Full-stack intelligent agent AI integrates multiple agents into complex workflows, autonomously executes the business processes of complex organizations, and realizes end-to-end closed-loop and automation of business processes. For example:

Supply chain management: Through process automation and optimization, AI can prioritize customer order delivery, rationalize production and sales coordination, and autonomously place orders with suppliers or adjust production plans to maintain optimal inventory levels.

Sales processing: Based on customer inquiries, the agent quotes the customer, applies for price discounts based on customer feedback, generates sales contracts and orders, and notifies the warehouse to ship the goods after the customer confirms and pays.

There is a conceptual distinction between Full Stack Agentic AI and AI Agent . The former is a framework, a broad concept of intelligent systems that " solve problems under limited supervision " , while the latter is a component within the framework, handling tasks and processes with a certain degree of autonomy. Usually an AI agent completes a specific capability, while Full Stack Agentic AI integrates multiple AI agents, each with its own goals and tasks, and works together through "AI orchestration " within the framework of Agentic AI to efficiently and stably achieve the goals of users of Agentic AI systems. AI orchestration technology is the core element of Full Stack Agentic AI .

AI orchestration refers to the coordination and management of information systems, business capability services, and AI agents . It integrates multiple agents that perform specific tasks and connects agents with other business services, data, and artificial intelligence capabilities, so that multiple intelligent business services can work together, give full play to their respective capabilities, and ensure that all agents and business services effectively contribute to more comprehensive and extensive goals.

The specific work of orchestration includes managing the data flow between agents, synchronizing their activities, and optimizing the resource usage of the entire system. The orchestration platform can automate AI workflows, track task completion progress, manage resource usage, monitor data flow and storage, and handle failure events. Through reasonable architectural design, in theory, dozens, hundreds, or even thousands of agents can work together efficiently.

The following figure is an example of AI-based orchestration of sales business. We assume that there are several intelligent agents responsible for tasks such as quotation, order entry, credit control, shipment processing, and inventory control. They are based on an orchestration framework and support orchestration data acquisition and process operation based on continuous dialogue interaction with users, large language models, and other AI models (such as transaction risk control and intelligent decision-making for supply chain inventory optimization):

Without AI orchestration, the entire system may not be able to effectively share information globally, resulting in inefficient order processing, or inventory out-of-stock, delivery errors, etc. Such orchestration can achieve intelligent automation of business processes, and it must have two prerequisites:

First, each business system, namely ERP , CRM , etc., must have sufficient and standardized external service call interfaces, namely APIs , so that different software components and services can communicate with each other.

Second, these intelligent capabilities and business services must be based on a cloud platform, have flexibility, scalability, and the computing power necessary to process large amounts of data and complex artificial intelligence algorithms, and facilitate management, continuous development, and continuous integration.

Full-stack intelligent body AI will be the mainstream form of large language models applied to enterprises. It is the third generation of business process automation technology after workflow + OA technology and process rule engine + RPA technology. Full-stack intelligent body AI will gradually realize the automation of mental labor using large language models. Its impact on enterprises is as significant and far-reaching as the impact of modern automatic assembly line technology on the manufacturing industry.