Method for developing ERP system based on AI large model

Abstract

The invention belongs to the technical field of artificial intelligence and data management, and particularly relates to a method for developing an ERP (Enterprise Resource Planning) system based on an AI (Artificial Intelligence) large model, which comprises the following steps of: S1, acquiring functions and constraint conditions required by each process in business processes of the ERP system; s2, creating a special AI agent for each process according to functions and constraint conditions required by each process; s3, training and optimizing each AI agent by using historical data by taking a function needing to be completed in each process as a training target; and S4, integrating and connecting the trained and optimized AI agents according to the logic sequence of the business process through a central composer. The problems that an existing ERP system is high in maintenance cost and low in intelligent level are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and data management, and particularly relates to a method for developing an ERP system based on an AI large model. BACKGROUND

[0002] As the core of an integrated management information system, an ERP (Enterprise Resource Planning) system deeply integrates the logistics, capital flow and information flow of an enterprise through a unified data platform, and implements efficient planning, control and decision support for core links such as a supply chain, production, finance and human resources.

[0003] However, the development and construction method of the prior art ERP system is mainly based on pre-defined function modules and rigidly fixed business process coding, which leads to several essential defects that need to be solved in practice: Firstly, the system flexibility and adaptability are seriously insufficient, when the business mode, management process or external compliance requirements change, even the smallest adjustment often needs to go through a tedious code modification, system testing and re-deployment cycle, which is high in use cost; secondly, the human-computer interaction is complex, and the user experience is poor, users must deeply understand the multi-layer menu structure and complex form logic of the system, and must be professionally trained to operate, which not only raises the training cost, reduces the work efficiency, but also leads to the general resistance of users, so that the value of the ERP system is difficult to fully play; finally, the existing ERP system is essentially a passive data recording and transaction processing system, which is focused on data entry, storage and standardized report output, and lacks the ability to deeply understand, intelligently analyze and make forward-looking decisions on massive business data, the system cannot actively discover process bottlenecks, warn operation risks or provide optimization suggestions, and the intelligent level is low.

[0004] At present, attempts are made to integrate peripheral functions such as chat robots into ERP through artificial intelligence technology, in order to improve user experience. However, these improvements are mostly superficial, and AI is only used as a substitute for traditional menu navigation or a simple information query tool, and cannot realize end-to-end adaptation and intelligent reconstruction of business processes, so the core problems of the ERP system such as rigidity, high maintenance cost and low intelligence level are still not solved. SUMMARY

[0005] The application provides a method for developing an ERP system based on an AI large model, which solves the problems of high maintenance cost and low intelligence level of the existing ERP system.

[0006] The basic scheme provided by the application is a method for developing an ERP system based on an AI large model, which comprises: S1: obtaining the functions and constraint conditions required by each process in the business process of the ERP system; S2: create a dedicated AI agent for each process according to the function and constraint condition required by each process; S3: train and optimize each AI agent using historical data with the function required to be completed by each process as the training target; S4: integrate and connect the multiple AI agents that have completed training and optimization according to the logical order of the business process through a central orchestrator.

[0007] Preferably, the constraint condition includes at least one of a business rule constraint, a data format constraint, and a permission constraint.

[0008] Further preferably, S1 includes: S1-1) According to the domain knowledge base combined with expert experience, sort out the end-to-end business process in the ERP system; S1-2) Task deconstruction is performed on each business process respectively to generate a process list and a business process diagram corresponding to each business process; S1-3) Define the function and constraint condition of each process according to the process list and the business process diagram.

[0009] Further preferably, the definition strategy in S1-3 includes a. Function definition strategy: through result-oriented analysis of each process in the process list, determine the function corresponding to the process according to the analysis result; b. Constraint condition definition strategy: parse the logic and data flow in the business process diagram to generate constraint conditions, specifically including: The business rule constraint is set according to the business logic contained in the decision gateway and sequence flow in the business process diagram, combined with the external laws and regulations and internal control rules that the enterprise must follow; The data format constraint is set according to the data flow objects and their attributes between the task nodes in the business process diagram; The permission constraint is set according to the data objects operated by the task nodes in the business process diagram and the actions performed, based on the principle of separation of duties.

[0010] Further preferably, S2 includes: S2-1) Determine the agent framework corresponding to each process according to the function and constraint condition required by each process; S2-2) Create a dedicated AI agent corresponding to each process, and configure the core behavior of the agent based on the function and constraint condition; S2-3) Associate the domain knowledge base to build the context of the agent; S2-4) Register and record the created dedicated AI agent in an agent registration center.

[0011] Further preferably, S2-1 comprises: Based on the flow-based function evaluation of the corresponding decision complexity, specifically according to the information completeness and the size of the strategy space, for the process with high complexity of decision, the framework supporting the cycle reasoning-action paradigm is selected; for the process with low complexity of decision, the linear chain framework is selected; Based on the function and constraint condition evaluation of the corresponding tool dependency of the process, specifically according to the number and type of necessary tools and the strictness of output structure, for the process with high tool dependency, the framework with powerful tool calling and structured output ability is selected; for the process with low tool dependency, the lightweight framework is selected.

[0012] Further preferably, in S2-2, the configuration strategy comprises: Convert the required functions of the process into natural language form as the core instructions written into the system prompt words of the agent; Encode the business rule constraints into explicit logical statements or conditional judgment rules, and integrate the encoded business rule constraints into the system prompt words or decision logic module of the agent; Convert the permission constraints into restrictions on the set of tools and data range that the agent can call, and explicitly declare its operation boundary in its system prompt words; According to the data format constraint, configure a strict input / output parser for the agent.

[0013] Further preferably, S4 comprises: S4-1) The central orchestrator parses the process list and the business process graph to generate a directed graph data structure, wherein the nodes represent agent tasks, parallel gateways or decision gateways, and the edges represent control flow or data flow; S4-2) By querying the agent registration center, identify each AI agent task node in the directed graph data structure, and obtain the network access address of the corresponding AI agent, and associate and save the address with the task node in the process model; S4-3) According to the data format constraint defined in S1, define and configure the data transfer rules between agent nodes, and set the data mapping relationship for each connection line; S4-4) At the decision gateway of the business process graph, deploy the corresponding routing rules according to the conditional judgment statements written by the output data of the upstream AI agent.

[0014] Preferably, it further comprises S5: During the running of the ERP business process, for the nodes that meet the preset triggering conditions, human intervention is introduced, and the corresponding AI agent is continuously optimized according to the feedback data of the human intervention.

[0015] Further preferably, S5 comprises: S5-1) During the operation of the business process, the central orchestrator monitors the execution status of each AI agent node in real time, compares the monitored data with the preset trigger condition, and if the condition is met, immediately suspends the fully automatic process and routes the task to the human intervention station; S5-2) The human expert handles the suspended task on the human intervention station, the expert reviews the AI output, context information, evaluates the AI decision, and forms feedback data; S5-3) According to the feedback data collected periodically, the corresponding AI agent is retrained.

[0016] The principles and advantages of the present application are: 1. By creating a dedicated AI agent for each business process and integrating it by the central orchestrator, when the business mode, process or compliance requirement changes, the agent's training target, constraint condition or orchestration logic can be adjusted to achieve adaptation. Compared with the existing ERP system which relies on pre-defined modules and rigid process coding, business adjustment requires tedious code modification and deployment. The present application does not need to modify the underlying code on a large scale, greatly shortens the adjustment period and reduces the maintenance cost.

[0017] 2. The business logic and execution component are completely decoupled through the agent registration center and the central orchestrator. Each agent is an independent, plug-in service module that can be created and registered at any time for new business needs without modifying the existing system, achieving seamless business expansion. At the same time, the failure or upgrade of a single agent will not affect the operation of the entire system, and the architecture is more robust.

[0018] 3. Each AI agent is trained based on historical data, with the goal of completing the process function, capable of deep understanding and intelligent analysis of massive business data, actively discovering process bottlenecks, warning operational risks and providing optimization suggestions, upgrading from passive processing to "active decision support", fully realizing the value of the ERP system; and through the combination of the agent and the central orchestrator, AI is deeply integrated into each business process link and end-to-end integration is achieved, fundamentally restructuring the operation logic of the ERP system and solving the core problem of traditional ERP rigidity and low intelligence.

[0019] 4. Human intervention and correction at key nodes are automatically collected by the system and used for continuous optimization of the agent, enabling the system to continuously learn from real business feedback, becoming more intelligent and accurate with use, ultimately forming a life form that can grow together with the enterprise, rather than a rigid software tool. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0021] Further details are described below by way of specific embodiments: The specific implementation process is as follows: referring to Figure 1 A method for developing an ERP system based on an AI large model, comprising: S1: obtaining the functions and constraint conditions required by each process in the business process of the ERP system; The constraint conditions include at least one of business rule constraints, data format constraints, and permission constraints; Comprise: S1-1) According to the domain knowledge base combined with expert experience, comb the end-to-end business process in the ERP system; Specifically, based on the domain knowledge base and the cooperation of domain experts (such as financial, supply chain, sales department managers), the core end-to-end business process of the enterprise is combed; for example, "purchase to payment", "order to cash", "production planning and execution", "expense reimbursement", etc.

[0022] S1-2) Task decomposition is performed on each business process respectively, and a process list and a business process diagram corresponding to each business process are generated; For example: in the process of purchasing to payment, the process list includes 1. Demand application submission; 2. Budget automatic verification; 3. Supplier intelligent recommendation; 4. Purchase order generation; 5. Contract key clause review; 6. Invoice automatic identification and information extraction; 7. Three single (order, warehouse entry, invoice) automatic matching; 7. Payment execution; The business process diagram can use BPMN or UML activity diagram, etc. The starting point, end point, all activity nodes, decision gateway and flow direction of each process are visually described; S1-3) Define the functions and constraint conditions of each process according to the process list and the business process diagram; The definition strategy includes a. Function definition strategy: through result-oriented analysis of each process in the process list, the function corresponding to the process is determined according to the analysis result; b. Constraint condition definition strategy: analyze the logic and data flow in the business process diagram to generate constraint conditions, specifically including: The business rule constraint is set according to the business logic contained in the decision gateway and the sequence flow in the business process diagram, combined with the external laws and regulations and internal control rules that the enterprise must follow; for example, after the "budget automatic verification" activity, the business process diagram has a decision gateway, leading to two paths: one pointing to "supplier recommendation", the condition being [budget sufficient]; the other pointing to process end or exception handling, the condition being [budget insufficient], and the rules involved are?; therefore, the business rule constraint set is: "IF budget verification result == 'budget sufficient' THEN continue the process to the supplier intelligent recommendation task; ELSE terminate the process and notify the applicant"; The data format constraint is set according to the data flow objects and their attributes between the task nodes in the business process diagram; for example, in the "purchase to payment" business process diagram, there is a data flow between the "invoice automatic identification and information extraction" task and the "three single automatic matching" task, which carries the structured information of the invoice. In order to ensure the accuracy of data transmission, the data format constraint set accordingly is: the output of the "invoice automatic identification and information extraction" task must be a standardized JSON object containing "supplier unified credit code" (string type), "invoice number" (string type), and "amount excluding tax" (numeric type); The permission constraint is set according to the data objects operated by the task nodes in the business process diagram and the actions performed, based on the principle of separation of duties; for example, in the "purchase to payment" business process diagram, the "supplier intelligent recommendation" task needs to read the supplier master data, while the "supplier master data maintenance" task is responsible for creating and modifying the data. In order to follow the principle of separation of duties, the permission constraint set accordingly is: the "supplier intelligent recommendation" agent is only granted the permission to read: supplier master data, but not the permission to create: supplier master data or modify: supplier master data.

[0023] S2: Create an AI agent for each process according to the required functions and constraint conditions of each process; S2-1) Determine the agent framework corresponding to each process according to the required functions and constraint conditions of each process; The strategy is to evaluate the decision complexity of the process based on the function of the process, specifically according to the information completeness and the size of the strategy space; for processes with high decision complexity, select a framework that supports the loop reasoning-action paradigm; for processes with low decision complexity, select a linear chain framework; Evaluate the tool dependency of the process based on the function and constraint conditions, specifically according to the number and type of necessary tools and the strictness of the output structure; for processes with high tool dependency, select a framework with strong tool invocation and structured output capability; for processes with low tool dependency, you can select a lighter framework.

[0024] In this embodiment, for tasks requiring complex reasoning, dynamic programming, or multi-step decision-making, a framework based on the ReAct (reasoning-action) paradigm or similar ideas is preferred. This type of framework structures the reasoning process of a large model and interacts with external tools or the environment through a "think-act-observe" cycle, making it particularly suitable for scenarios that require adjusting strategies based on feedback. For example, in "purchase-to-pay," the "supplier intelligent recommendation" process needs to weigh multiple dynamic factors such as price, delivery time, and quality, and may require multiple rounds of information retrieval and comparison, thus employing the ReAct (reasoning-action) paradigm. For tasks requiring efficient and stable calls to external APIs or processing of structured data, frameworks like LangChain and LlamaIndex, which focus on toolchain integration and data connectivity, should be prioritized. These frameworks offer powerful tool invocation capabilities and context management, making them suitable for process-oriented and automated processing. For example, in the "procurement to payment" process, the core of the "automatic budget verification" and "automatic matching of three documents" processes involves querying the database and performing deterministic logical judgments, which utilizes LlamaIndex.

[0025] For tasks centered on information retrieval and question answering, and involving a large knowledge base, frameworks with built-in Advanced Search Enhancement Generation (RAG) optimization chains should be prioritized. These frameworks focus on quickly locating relevant information from massive amounts of enterprise knowledge and generating summary answers.

[0026] S2-2) Create dedicated AI agents for each process and configure the core behaviors of the agents based on functions and constraints; Configuration strategies include: The required functions of the process are converted into natural language and written into the system prompts of the intelligent agent as core instructions; Encode business rule constraints into explicit IF-THEN-ELSE logical statements or conditional judgment rules, and integrate them into the system prompts or decision logic modules of the agent as "hard rules" that it must follow during reasoning; Transform permission constraints into restrictions on the toolset and data range that the agent can access, and explicitly state the boundaries of its operations in its system prompts, such as "You are not authorized to perform payment operations, you can only generate payment suggestions," etc. Based on data format constraints, configure strict input / output parsers for the agent, such as using a Pydantic model or JSONSchema, to force it to accept inputs that conform to the specifications and generate strictly structured and correctly typed outputs, ensuring seamless integration with upstream and downstream tasks.

[0027] S2-3) Associate the domain knowledge base to construct the agent context; Specifically includes: According to the function corresponding to the AI agent, the required knowledge range of the AI agent is determined from the pre-built enterprise domain knowledge base, for example, all contract templates and laws and regulations are associated with the "contract review agent"; The knowledge documents determined are sliced, vectorized, and stored in a vector database. By configuring the RAG interface of the AI agent, it can first retrieve the most relevant knowledge fragments from the vector database when performing the process task. When the agent is executed, the knowledge fragments retrieved from the associated domain knowledge base serve as the key context, combined with the user's original query and system prompt words, to form a complete input submitted to the underlying large model to generate professional output that meets the enterprise's specific knowledge.

[0028] It also includes S2-4) Register and record the created special AI agent in an agent registration center; Specifically record its metadata, including: business process, function target, version number, service endpoint, etc. so that the central orchestrator in step S4 can accurately discover, schedule and manage these agents.

[0029] S3: Train and optimize each AI agent using historical data with the function to be completed as the training target; Specifically, collect historical data of each process, and train and optimize the AI agent using the corresponding historical data; The training strategy includes: For AI agents using the ReAct framework, use process supervision fine-tuning during training and environment interaction reinforcement learning during optimization; For AI agents using the LangChain / LlamaIndex framework, use tool invocation precision fine-tuning during training, specifically using (natural language instructions, correct tool invocation sequence and parameters) data pairs for fine-tuning, and use process efficiency optimization during optimization; For AI agents using the advanced RAG framework, use retriever-generator joint fine-tuning during training, specifically fine-tuning the generation model and optimizing the retriever, using (user questions, relevant knowledge fragments, ideal answers) data pairs to train the retrieval model to find more relevant information and train the generation model to better utilize the retrieved context.

[0030] S4: Integrate and connect multiple AI agents that have completed training and optimization according to the logical order of the business process through a central orchestrator.

[0031] S4-1) The central orchestrator parses the process list and the business process diagram to generate a directed graph data structure, wherein the nodes represent agent tasks, parallel gateways or decision gateways, and the edges represent control flow or data flow; the specific parsing logic is that the orchestrator identifies the start event, end event, sequence flow, decision gateway (exclusive / inclusive), parallel gateway and other elements in the diagram, constructs the internal execution model of the process, and forms a directed graph data structure.

[0032] S4-2) By querying the agent registration center, each AI agent task node in the directed graph data structure is identified, and the network access address (API endpoint) of the corresponding AI agent is obtained, and the address is associated and saved with the task node in the process model; S4-3) According to the data format constraint defined in S1, the data transmission rules between agent nodes are defined and configured, and the data mapping relationship is set for each connection line, for example, the "invoice_number" field in the output of the "invoice recognition agent" is mapped to the "target_invoice_num" field in the input of the "three-in-one matching agent"; S4-4) At the decision gateway of the business process diagram, the corresponding routing rules are deployed according to the conditional judgment statements written based on the output data of the upstream AI agent; for example, at the gateway after the "budget verification" node, the rule is deployed: IF budget verification result.Output.approval_result==True THEN flow to "supplier recommendation" node; ELSE flow to "process end" node; Further comprising S4-5) Before executing the entire process, start an integration verification process; for example, use test data to run the process from the beginning, verify whether each agent can be correctly called, whether the data can be accurately transmitted according to the configuration, and whether the routing logic is correct. After verification, deploy the integrated business process online so that it can be triggered to execute by external events or user requests.

[0033] Embodiment 2 Further comprising S5: During the running of the ERP business process, for nodes that meet the preset triggering conditions, human intervention is introduced, and the corresponding AI agent is continuously optimized according to the feedback data of human intervention.

[0034] S5-1) During the running of the business process, the central orchestrator monitors the execution state of each AI agent node in real time, compares the monitored data (such as output content, confidence, processing time) with the preset triggering conditions, and if the conditions are met, the fully automatic process is immediately suspended and the task is routed to the human intervention station; S5-2) Human experts (e.g. financial director, purchasing manager) handle the suspended tasks at the human intervention station, the experts review the AI's output, context information, evaluate the AI's decision, form feedback data; S5-3) According to the feedback data collected regularly (e.g. weekly or monthly), the corresponding AI agent is retrained.

[0035] The above is only an embodiment of the present application, and the common knowledge of the specific structure and characteristics in the scheme is not described in detail here. The ordinary skilled person in the art knows all the ordinary technical knowledge in the field of the application before the application date or the priority date, can know all the prior art in the field, and has the ability to apply conventional experimental means before that date. The ordinary skilled person in the art can perfect and implement the present scheme under the guidance of the present application, combined with their own ability. Some typical known structures or known methods should not be an obstacle for the ordinary skilled person in the art to implement the present application. It should be noted that for those skilled in the art, without departing from the structure of the present application, a number of modifications and improvements can be made, which should also be considered as the protection scope of the present application. These will not affect the effect and practicality of the patent. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode in the specification can be used to explain the content of the claims.

Claims

1.A method of developing an ERP system based on an AI large model, characterized by, Comprise: S1: obtain the function and constraint condition required by each process in the business process of ERP system; S2: create a dedicated AI agent for each process according to the function and constraint condition required by each process; S3: take the function required to be completed by each process as the training target, and train and optimize each AI agent by using historical data; S4: integrate and connect the multiple AI agents that have completed training and optimization according to the logical order of the business process through a central orchestrator. 2.The method of developing an ERP system based on an AI large model according to claim 1, characterized in that: The constraint condition comprises at least one of a business rule constraint, a data format constraint and a permission constraint. 3.The method of developing an ERP system based on an AI large model according to claim 2, characterized in that: S1 comprises: S1-1) According to the domain knowledge base combined with expert experience, the end-to-end business process in the ERP system is sorted out; S1-2) The task of each business process is deconstructed respectively to generate the process list and business process diagram corresponding to each business process; S1-3) Define the function and constraint condition of each process according to the process list and business process diagram. 4.The method of developing an ERP system based on an AI large model according to claim 3, characterized in that: The definition strategy in S1-3 comprises a. Function definition strategy: through result-oriented analysis of each process in the process list, the function corresponding to the process is determined according to the analysis result; b. Constraint condition definition strategy: the logic and data flow in the business process diagram are parsed to generate constraint conditions, which specifically include: The business rule constraint is set according to the business logic contained in the decision gateway and sequence flow in the business process diagram, combined with the external laws and regulations and internal control rules that the enterprise must follow; The data format constraint is set according to the data flow objects and their attributes between the task nodes in the business process diagram; The permission constraint is set according to the data objects operated by the task nodes in the business process diagram and the actions executed, based on the principle of separation of duties. 5.The method of developing an ERP system based on an AI large model according to claim 3, characterized in that: S2 comprises: S2-1) Determine the agent framework corresponding to each process according to the function and constraint condition required by each process; S2-2) Create a dedicated AI agent corresponding to each process, and configure the core behavior of the agent based on the function and constraint condition; S2-3) Associate the domain knowledge base to build the context of the agent; S2-4) Register and record the created dedicated AI agent in an agent registration center. 6.The method of developing an ERP system based on an AI large model according to claim 5, characterized in that: S2-1 comprises: Based on the function evaluation of the process, the decision complexity is evaluated, specifically according to the information completeness and strategy space size; for processes with high decision complexity, a framework supporting loop reasoning-action paradigm is selected; for processes with low decision complexity, a linear chain framework is selected; Based on the function and constraint condition, the tool dependency of the process is evaluated, specifically according to the number and type of necessary tools and the strictness of output structure; for processes with high tool dependency, a framework with strong tool calling and structured output capability is selected; for processes with low tool dependency, a lightweight framework is selected. 7.The method of developing an ERP system based on an AI large model according to claim 6, characterized in that: In S2-2, the configuration strategy comprises: Convert the required function of the process into natural language form as the core instruction written into the system prompt word of the agent; Encode the business rule constraint into an explicit logic sentence or conditional judgment rule, and integrate the encoded business rule constraint into the system prompt word or decision logic module of the agent; Convert the permission constraints into the restrictions on the set of callable tools and the range of data accessible to the AI agents, and explicitly declare the operation boundaries in the system prompt words; According to the data format constraints, configure the strict input / output parser for the agent. 8.The method of developing an ERP system based on an AI large model according to claim 5, characterized in that: S4 includes: S4-1) The central orchestrator parses the process list and business process diagram to generate a directed graph data structure, wherein the nodes represent agent tasks, parallel gateways or decision gateways, and the edges represent control flow or data flow; S4-2) By querying the agent registration center, identify each AI agent task node in the directed graph data structure, and obtain the network access address of the corresponding AI agent, and associate and save the address with the task node in the process model; S4-3) According to the data format constraints defined in S1, define and configure the data transfer rules between agent nodes, and set the data mapping relationship for each connection line; S4-4) At the decision gateway of the business process diagram, deploy the corresponding routing rules according to the conditional judgment statements written based on the output data of the upstream AI agent. 9.The method of developing an ERP system based on an AI large model according to claim 1, characterized in that: Also includes S5: During the running of the ERP business process, for the nodes that meet the preset trigger conditions, introduce human intervention, and continuously optimize the corresponding AI agent according to the feedback data of human intervention. 10.The method of developing an ERP system based on an AI large model according to claim 9, characterized in that: S5 includes: S5-1) During the running of the business process, the central orchestrator monitors the execution state of each AI agent node in real time, compares the monitored data with the preset trigger conditions, and if the conditions are met, immediately suspends the fully automatic process and routes the task to the human intervention station; S5-2) Human experts handle the suspended tasks on the human intervention station, experts review the AI output, context information, evaluate the AI decision, and form feedback data; S5-3) According to the feedback data collected regularly, it is used to retrain the corresponding AI agent.

Images