Real-time budget control and AI dynamic cancellation method and system embedded in business process

Abstract

This invention relates to the field of business-finance integration technology, specifically to a real-time budget control and AI-driven dynamic reconciliation method and system embedded in business processes. The method includes: deploying an AI agent in the business system, triggered when a business instruction is initiated; the AI ​​agent generating a budget allocation decision in real time based on application information and budget strategy, calculating a dynamic allocation amount based on predictions of payment probability and timing; executing dynamic allocation in the budget pool based on the decision; and dynamically adjusting the allocation amount based on event details when a payment or reconciliation event is detected. This invention solves the problems of lagging and rigid traditional budget control by placing the intelligent control point at the business initiation end and achieving flexible dynamic budget allocation and automated reconciliation adjustment. It realizes real-time synchronization and refined intelligent control of budget management and business processes, improving business experience and budget resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of business and finance integration technology, specifically to a real-time budget control and AI dynamic reimbursement method and system embedded in business processes. Background Technology

[0002] In business management, budget control is a core means of ensuring financial compliance and optimizing resource allocation. Traditional budget control models, especially for expenses and procurement, typically function as a "checkpoint" in the post-audit or reimbursement process. Specifically, when business departments incur expenses or make purchases, they usually advance the funds or sign contracts. Only when invoices are obtained and the process enters the financial reimbursement or payment process does the financial system deduct from the budget and conduct compliance checks. This "post-event interception" model has significant drawbacks: First, control is severely delayed. Budget overruns or non-compliance are only discovered after the business has actually occurred. By then, the transaction is a fait accompli, correction costs are high, and budget control often becomes merely a formality, failing to serve its purpose of pre-emptive constraint and planning. Second, the business experience is poor. Business personnel are only informed of insufficient or non-compliant budgets at the end of the process, leading to failed reimbursements and payment delays. This not only affects business progress but also causes inter-departmental conflicts. In addition, traditional budget management is mostly static and rigid "total amount control" or "pre-freezing". Once the application is approved, the corresponding budget amount is frozen in full and rigidly, and cannot be dynamically adjusted according to the actual execution risks of the business (such as cancellation, change, price fluctuations). This leads to low efficiency in the use of budget resources, a large amount of budget is used ineffectively, and affects the flexibility of enterprise operation.

[0003] While some existing technologies incorporate budget checks during the business application process, their control logic is often based on simple balance checks (i.e., "application amount ≤ budget balance"). This rigid, one-size-fits-all control lacks intelligent assessment of business rationality, cannot handle complex expense control policies, and the "freeze all funds upon application" model fails to address resource inefficiency. Another approach is to introduce workflow approvals for human judgment, but this leads to long approval chains, low efficiency, and inconsistent standards. Therefore, how to intelligently and seamlessly advance budget control to the source of business operations and achieve flexible and dynamic budget management throughout the entire business lifecycle, thereby truly realizing business-finance integration and refined control, has become a pressing technical challenge in this field.

[0004] Therefore, existing technologies still need further development. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a real-time budget control and AI dynamic reimbursement method and system embedded in business processes to solve the problems existing in the prior art.

[0006] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides a real-time budget control and AI dynamic reconciliation method embedded in business processes, comprising: S1. Deploy an AI agent for real-time budget control in the business system, and trigger the AI ​​agent when the business system receives an instruction to initiate a business; the AI ​​agent generates a budget allocation decision in real time based on the application information of the business and the associated budget strategy; S2. Based on the budget allocation decision, perform dynamic budget allocation in the budget pool; S3. When a payment or reimbursement event related to the business is detected, the AI ​​model dynamically adjusts the amount of the budget pool occupied based on the details of the event.

[0007] Specifically, the AI ​​agent is triggered when the business system receives an instruction to initiate a business, including: When the business system detects that a user has submitted a purchase request, advance payment request, or contract payment request, it calls the interface corresponding to the AI ​​agent and transmits the request information, including project identifier, budget item, amount, and business type, along with user identity and context information of the current business scenario, to the AI ​​agent. The real-time decision-making of the AI ​​agent includes: The system retrieves the real-time budget balance bound to the project identifier and budget item from the business database, and obtains historical business data associated with the user and the project within a preset time period. Based on a preset authentication model that integrates multiple enterprise expense control rules, the system performs a fusion analysis on the application information, historical business data, and current context information to determine whether the application is compliant and whether it exceeds a preset flexible threshold. When the application is approved, the AI ​​agent generates a decision result containing an approval identifier and a suggested occupancy plan; otherwise, it generates a decision result containing a rejection identifier and a reason.

[0008] Specifically, the AI ​​agent generates budget allocation decisions in real time based on the application information of the business and the associated budget strategy. The specific steps for generating a suggested allocation plan include: Based on the business type, current project stage, and payment conversion rate records of similar past businesses, a machine learning model is invoked to predict the probability and expected time of cancellation, modification, or final payment completion of this application in subsequent business stages. Based on the probability and expected time, a dynamic budget allocation weight between 0 and 1 is calculated. This weight is positively correlated with the predicted final payment probability and negatively correlated with the expected time. Based on the application amount and the dynamic budget allocation weight, a preliminary budget allocation amount is calculated. Combined with preset minimum allocation thresholds for different business types, a final recommended budget allocation amount is determined to form a recommended allocation plan.

[0009] Specifically, based on the aforementioned budget allocation decision, dynamic budget allocation is performed in the budget pool, specifically as follows: Based on the suggested budget allocation in the decision results, the corresponding budget allocation is marked as "dynamically occupied" in the corresponding budget pool. In this state, the budget allocation can still be partially occupied by higher priority business applications, but it is no longer fully included in the available balance for regular approval. The system records the business application identifier, the amount occupied, the dynamic occupancy weight, the occupancy timestamp, and the predicted reimbursement node information for this occupancy, and associates the occupancy record with the business process instance of the business application.

[0010] Specifically, upon detecting a payment or reconciliation event related to the aforementioned business, the AI ​​model dynamically adjusts the allocation to the budget pool based on the details of the event, including: The payment or cancellation events include invoice registration verification, service completion confirmation, contract payment execution, or business application cancellation. When an invoice registration verification event is detected, the AI ​​model obtains the invoice amount, tax rate, and actual supplier information and matches them with the original business application. If there is a price-tax difference, the model automatically calculates the amount of budget to be released or added from the "dynamically occupied" state according to the preset price-tax difference handling strategy, and updates the actual occupied value in the budget pool.

[0011] Specifically, upon detecting a payment or reconciliation event related to the aforementioned business, the AI ​​model dynamically adjusts the allocation to the budget pool based on the details of the event, and also includes: When an event is detected where a business application is actively withdrawn, the AI ​​model determines, based on the process node of the business application, the reason for withdrawal, and the probability of re-initiating after a previous withdrawal, whether to immediately release all dynamically occupied budget quotas or to retain a portion of the quotas in a "suspended" state for a preset period of time to deal with possible emergency re-applications.

[0012] Specifically, upon detecting a payment or reconciliation event related to the aforementioned business, the AI ​​model dynamically adjusts the allocation to the budget pool based on the details of the event, further including: In scenarios where contracts are paid in multiple phases, when an event is detected that a certain phase payment has been completed, the AI ​​model automatically calculates the difference between the amount payable for this phase and the original dynamic occupancy limit based on the payment plan stipulated in the contract and the evaluation information of the current phase completion quality. It then completes the actual write-off of the corresponding amount and generates a new dynamic budget occupancy suggestion based on the updated project progress for the next payment phase, thereby updating the occupancy status in the budget pool.

[0013] Specifically, when the AI ​​agent generates budget allocation decisions in real time, it also performs the following steps: For specific business types, key entities in the application information are identified and compared with a risk knowledge base. If a high-risk entity is identified, an enhanced review process is triggered. This enhanced review process includes automatically retrieving the entity's past transaction dispute records and market price fluctuation data, simulating the impact of different approval decisions on the overall project budget expenditure over a future period, and finally generating a report containing risk warnings and multi-scenario decision suggestions for authorized nodes to review.

[0014] Specifically, the method also includes an online dynamic optimization step for the budget strategy: continuously collecting full-link data of all business applications and their subsequent payment verification to form decision-outcome sample pairs; periodically using new samples to train the budget occupancy weight prediction model and compliance judgment model, and optimizing their parameters; deploying the optimized model parameters to the AI ​​agent in a hot-update manner to realize the adaptive adjustment of the budget control strategy based on actual business feedback, wherein the full-link data is anonymized before being used for training to remove directly personal identity information.

[0015] According to a second aspect of the present invention, a real-time budget control and AI dynamic reimbursement system embedded in business processes is provided, comprising: The deployment module is used to deploy AI agents for real-time budget control in business systems; The triggering module is used to trigger the AI ​​agent when the business system receives an instruction to initiate a business. The decision-making module, built into the AI ​​agent, is used to generate budget allocation decisions in real time based on the application information of the business and the associated budget strategy. The execution module is used to perform dynamic budget allocation in the budget pool based on the budget allocation decision; The adjustment module is used to dynamically adjust the amount of the budget pool occupied by the AI ​​model based on the details of the event when a payment or reconciliation event related to the business is detected.

[0016] Beneficial effects: The real-time budget control and AI dynamic reimbursement method and system embedded in the business process provided by this invention achieves a fundamental transformation from passive, lagging, and rigid traditional budget control to proactive, real-time, and flexible intelligent control by deeply embedding a lightweight AI agent into the beginning of the business system process and introducing a machine learning-based dynamic budget occupancy and reimbursement mechanism. This has yielded many significant benefits.

[0017] First, the core effect of this invention lies in fundamentally shifting budget control points forward and achieving real-time and intelligent control. Through millisecond-level real-time authentication by an AI agent at the moment a business application is submitted, budget and compliance control are seamlessly and imperceptibly embedded at the forefront of the business process, transforming "post-event interception" into "pre-event prevention" and "in-event control." This fundamentally eliminates over-budget and non-compliant business activities, greatly improving the business experience and achieving "seamless management." Simultaneously, the AI ​​agent integrates rules and models, combining historical data with real-time context for comprehensive judgment, achieving a leap from simple balance control to intelligent decision-making based on business rationality.

[0018] Secondly, this invention achieves refined and flexible budget resource management by introducing the concepts of "dynamic budget occupancy weight" and "flexible budget pool." It utilizes machine learning models to predict the payment probability and timing of business applications and calculates differentiated budget occupancy amounts accordingly, changing the traditional extensive model of "full freeze upon application." This ensures that high-risk, long-term applications only occupy a portion of the budget, while highly certain applications can be fully utilized, significantly improving the efficiency and turnover rate of budget resources. When budgets are tight, the system can support more potential business, optimizing resource allocation.

[0019] Third, this invention automates and dynamically manages budget reconciliation. By monitoring key business events such as invoice verification, contract payment, and application cancellation, and utilizing AI models to automatically calculate discrepancies, adjust usage, and complete reconciliation, it achieves precise and real-time synchronization between budget status and business execution status. This not only significantly reduces the manual operations and reconciliation workload of finance personnel and lowers human error, but more importantly, it can automatically handle normal fluctuations in business execution and intelligently release or adjust the budget when business changes occur, enabling budget management to truly follow business flow and achieving a deep closed loop of business-finance integration.

[0020] Fourth, the adaptive optimization capability of this invention ensures the continuous effectiveness of the system. By continuously collecting data across the entire "decision-outcome" chain and periodically retraining the model, the system can learn from actual business feedback, automatically optimizing its predictive accuracy and decision rationality. This enables the budget control strategy to adapt to changes in the enterprise's business model and external environment, possessing long-term evolutionary capabilities.

[0021] Finally, from a system perspective, the modular design allows for flexible integration with various existing business systems (OA, ERP, etc.), facilitating easy implementation. Enhanced review and risk simulation functions improve automation while providing robust data support and scenario analysis for high-risk and complex decisions, assisting managers in making scientific decisions and achieving a perfect combination of automation and human intervention. In summary, this invention brings revolutionary improvements in control effectiveness, resource efficiency, process automation, system intelligence, and implementability. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the real-time budget control and AI dynamic reconciliation method embedded in the business process provided in a specific embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0024] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0025] Please see Figure 1 This invention provides a real-time budget control and AI-driven dynamic reconciliation method embedded in business processes, comprising: S1. Deploy an AI agent for real-time budget control in the business system, and trigger the AI ​​agent when the business system receives an instruction to initiate a business; the AI ​​agent generates a budget allocation decision in real time based on the application information of the business and the associated budget strategy.

[0026] It should be further explained that the real-time budget control and AI dynamic reimbursement method embedded in business processes provided by this invention aims to solve the problems of poor user experience and control lag caused by the "post-event interception" in the reimbursement process of traditional budget control. The core of this method lies in moving the intelligent control point forward to the source of business initiation. First, in the enterprise's OA, ERP, SRM and other business systems, a lightweight AI agent service (for example, in the form of microservices or API middleware) is deployed to achieve seamless integration with the existing business systems. This AI agent is the carrier of budget control rules and intelligent judgment logic.

[0027] When a business user initiates a purchase request, an advance payment, or a contract payment process within the business system, the corresponding functional module (such as the backend interface for the "Submit" button) will, upon receiving the instruction, not immediately proceed to the next stage or perform only simple validation as in traditional processes. Instead, it will simultaneously invoke the aforementioned AI proxy service interface. During this invocation, the business system will package and transmit the complete context information of the application. This information typically includes, but is not limited to: application number, applicant user ID, department, associated project or cost center number, specific budget item code, application amount, business type (such as "equipment purchase," "travel expenses," or "service fee"), supplier information, item description, and the current process instance ID.

[0028] Once triggered, the AI ​​agent's core task is to complete a comprehensive "real-time budget authentication" within a millisecond response time and generate a decision. This decision is not merely a binary "approve" or "reject" judgment, but also includes a "flexible" budget allocation scheme, laying the foundation for subsequent dynamic management. To this end, the AI ​​agent needs to obtain necessary information in real time (typically within 200 milliseconds) from multiple data sources within the enterprise. These data sources include the budget management system (for real-time budget balance and budget version), historical business databases (for past similar applications and execution data), and master data systems (for supplier and material information).

[0029] S2. Based on the budget allocation decision, perform dynamic budget allocation in the budget pool.

[0030] It's important to further clarify that after a decision is generated, the system will operate on the corresponding "budget pool" based on the decision results. Here, the budget pool is a logical concept that corresponds to the total annual budget of a project under a specific budget item. The operation is not a simple full freeze, but rather a "dynamic allocation": based on the suggested solution in the decision, a portion of the budget pool is marked as "dynamically allocated." Budget amounts in this state are somewhere between "available" and "frozen," and the specific allocation rules will be explained in detail later.

[0031] S3. When a payment or reimbursement event related to the business is detected, the AI ​​model dynamically adjusts the amount of the budget pool occupied based on the details of the event.

[0032] It should be further clarified that the business process is not interrupted; the application will continue to the subsequent approval and execution stages. The system will continuously monitor subsequent key events related to this business application, which will trigger dynamic budget reimbursement. Key events include: (1) The associated invoices are registered and verified in the financial system; (2) The relevant services have been confirmed as completed and an acceptance form has been signed; (3) The payment milestones stipulated in the contract are reached and the payment process is triggered; (4) The application is withdrawn by the applicant or approver during the process. When these events are detected, a dedicated AI reimbursement model (which may be part of the aforementioned AI agent or another collaborative service) will be activated. The model will obtain detailed data about the event (such as the actual amount of the invoice and the reason code for withdrawal) and, based on a set of preset rules and learned patterns, automatically calculate what kind of adjustment needs to be made to the original "dynamically occupied" quota—whether it is partial release, full release, additional occupation, or conversion to actual reimbursement. Then, the system will automatically execute this adjustment instruction, update the status of the budget pool, and thus complete a complete cycle of budget occupation and release synchronized with the business flow.

[0033] Understandably, this method seamlessly and imperceptibly embeds budget control points from the post-financial stage to the very beginning of business operations, achieving proactive and real-time control and significantly improving the business experience. By introducing "dynamic occupancy" status and AI-driven reconciliation adjustments, it changes the traditional rigid "one-size-fits-all" budget freeze model, making budget management more flexible, intelligent, and precise. It can flexibly respond to changes in actual business execution (such as price fluctuations, application cancellations, and installment payments), improving the utilization efficiency of budget resources and the flexibility of business operations while effectively controlling costs. This is a key technology for achieving deep integration of business and finance management.

[0034] Specifically, the AI ​​agent is triggered when the business system receives an instruction to initiate a business, including: When the business system detects that a user has submitted a purchase application, expense prepayment application, or contract payment application, it calls the interface corresponding to the AI ​​agent and transmits the application information, including project identifier, budget item, amount, and business type, along with the user's identity and the context information of the current business scenario, to the AI ​​agent. The AI ​​agent's real-time decision-making includes: retrieving the real-time budget balance bound to the project identifier and budget item from the business database, and retrieving historical business data associated with the user and the project within a preset time period; based on a preset authentication model that integrates multiple enterprise expense control rules, performing a fusion analysis on the application information, historical business data, and current context information to determine whether the application is compliant and whether it exceeds a preset flexible threshold; when the determination is approved, the AI ​​agent generates a decision result including an approval identifier and a suggested occupancy plan; otherwise, it generates a decision result including a rejection identifier and a reason.

[0035] It should be further explained that this step defines in detail how the AI ​​agent is precisely triggered and the specific process of its real-time authentication. The triggering mechanism is key to embedding it into business processes.

[0036] (1) Triggering Timing and Information Transmission: Business systems (such as OA and ERP) need to embed calls to the AI ​​agent service at the code level when users perform key actions such as "submit" and "save and submit for review". The transmitted information package needs to be structured, for example, encapsulated in JSON format, and must contain two types of information: first, the core fields of the application form itself, and second, rich contextual information. Contextual information may include: the user's department and job level, the current date and time, the approval opinion of the previous stage associated with the business application (if any), the urgency level of the application, and whether the current operation is a submission from a draft or a temporary submission, etc. This contextual information is crucial for the AI ​​to make reasonable judgments.

[0037] (2) Real-time Data Acquisition: Upon receiving a request, the AI ​​agent executes multiple data queries in parallel. First, it queries the budget management system for the real-time available budget balance using the "project identifier + budget item code" as the key. Here, "real-time" refers to the net value after deducting all approved, frozen, and "dynamically occupied" quotas. Second, it queries the data warehouse or business database for historical behavioral data, with the preferred time period being the past 12 months. Query criteria include: historical application records of the same applicant / project under similar budget items, approval rate, difference between actual payment amount and application amount (execution deviation rate), and common reasons for rejection in historical applications. In addition, it can also query the historical cooperation records and price levels of the application suppliers.

[0038] (3) Authentication Model Analysis: The preset authentication model is a hybrid system combining a rule engine and a lightweight machine learning model. Its analysis and judgment process is hierarchical: a) Compliance hard rule check: First, run a series of Boolean rules, such as "Is this budget item open to this department?", "Does the application amount exceed the single transaction limit?", "Is the supplier in the qualified supplier list?", and "Is the selected material in the procurement catalog?". If any hard rule fails, a rejection decision is generated directly.

[0039] b) Reasonableness Soft Rules and Flexible Threshold Judgment: After passing the hard rules, the reasonableness analysis begins. Here, the concept of a "flexible threshold" is introduced. Enterprises can set different flexible thresholds for different business types and departments. For example, the flexible threshold for regular procurement can be set at 10% of the budget balance, while for urgent R&D procurement it can be set at 20%. The AI ​​model will calculate the proportion of the application amount to the real-time budget balance. If the proportion is lower than the flexible threshold, it tends to be approved quickly. If it exceeds the flexible threshold but does not exceed 100%, a more complex judgment is triggered: a comprehensive score is calculated by combining historical behavioral data (e.g., the applicant's past execution deviation rate is low, indicating accurate estimation), project stage (e.g., the critical implementation period of the project), and market price fluctuations (judged through access to external data services). If the comprehensive score exceeds a certain threshold (e.g., 0.7), it is approved; otherwise, rejection or manual review may be recommended.

[0040] (4) Decision Result Generation: The decision result is a structured data object. For approved results, in addition to the "Approved" identifier, it must include a "Suggested Usage Plan," which includes at least a suggested budget usage amount (which may differ from the requested amount) and the expected usage weight (initial value). For rejected results, in addition to the "Rejected" identifier, it must include a clear reason code and description, such as "Exceeded the department's monthly flexibility threshold by 5%" or "This supplier has a record of contract disputes in the past year," so that the front-end business system can display it to the user.

[0041] Understandably, through the aforementioned refined triggering and authentication processes, budget control is no longer an independent link detached from the business flow, but rather becomes an intelligent, embedded service within the business process. Real-time data acquisition ensures the timeliness of decision-making. The hybrid authentication model, while adhering to compliance standards, introduces flexibility and humanization into management through flexible thresholds and intelligent scoring, avoiding the obstruction of reasonable business needs by mechanical controls. Structured decision results provide clear input for subsequent steps and improve the efficiency and standardization of interaction with business systems.

[0042] Specifically, the AI ​​agent generates budget allocation decisions in real time based on the application information of the business and the associated budget strategy. The specific steps for generating a suggested allocation plan include: Based on the business type, current project stage, and payment conversion rate records of similar past businesses, a machine learning model is invoked to predict the probability and expected time of this application being converted into actual payment. Based on the probability and expected time, a dynamic budget allocation weight between 0 and 1 is calculated. This weight is positively correlated with the predicted final payment probability and negatively correlated with the expected time. Based on the application amount and the dynamic budget allocation weight, a preliminary budget allocation amount is calculated. Combined with preset minimum allocation thresholds for different business types, the final recommended budget allocation amount is determined to form a recommended allocation plan.

[0043] It should be further explained that this step reveals the core algorithmic logic of "dynamic budget allocation," namely, how to "translate" a business application into a flexible and intelligent budget allocation instruction. This is key to achieving refined management of budget resources.

[0044] (1) Payment Probability and Time Prediction: The AI ​​agent internally calls a dedicated prediction microservice, which is supported by a machine learning model. The model's input features mainly include: business type (e.g., encoded as a categorical variable), project stage (e.g., "initiation," "design," "development," "acceptance"), application amount, applicant's job level, application quarter, and statistical features extracted from historical data, such as "payment conversion rate for this business type in the past 6 months" and "average application-to-payment time for this applicant in similar project stages." The model preferably uses a Gradient Boosting Decision Tree (GBDT) algorithm, such as XG Boost or Light GBM, because it has good capture capabilities for tabular data and non-linear relationships, and its training and prediction speeds are relatively fast. The model outputs two predicted values: payment probability. ( This indicates the probability that the application will eventually be actually paid out; the expected timeframe. ( (Usually in days), representing the average time from application submission to the predicted payment date. For example, for a standard equipment purchase order, the model might predict its payment probability. Expected time sky.

[0045] (2) Calculation of dynamic budget occupancy weight: After obtaining the predicted value, calculate the dynamic budget occupancy weight. A preferred weighting formula is as follows: in, It is a dynamic budget allocation weight. It is the payment probability predicted by the model. It is the expected time (in days) predicted by the model. It is the time decay coefficient, a hyperparameter that needs to be tuned. It is a natural constant. This formula reflects the relationship between weights and payout probabilities. Positive correlation (the higher the probability, the more budget should be allocated), with expected time. Negative correlation (the later the payment, the lower the current necessity of the funds; this can be simulated using an exponential decay function to represent the time value of money). Time decay coefficient. The preferred value range is between 0.1 and 0.5, for example, it can be taken as... .choose The rationale is that it generates meaningful weight decay over typical business cycles (30-90 days), avoiding both overly aggressive approaches that lead to insufficient short-term payment requests and overly conservative approaches that result in excessive long-term budget tie-up. For example, when hour, This value is too small, therefore, in practical applications, it will... Normalization or piecewise functions can be used to prevent weights from decaying to zero prematurely. Another more practical formula is: in, It's about time The decay function, for example Or a piecewise linear function. This is a balancing coefficient; for example, a value of 0.7 indicates a greater emphasis on payment probability. Companies can define and adjust this weighting calculation function according to their own business characteristics.

[0046] (3) Recommended budget allocation: Preliminary budget allocation Then, it needs to be combined with the minimum usage threshold. Make corrections. Different business types have different requirements. For example, for high-frequency, low-amount transactions with high payment certainty, such as "travel advance payment," It can be set at 80% of the application amount (or a fixed small amount, such as 500 yuan) to ensure necessary control; for "consulting service procurement", It may be set at 30% of the application amount. The final suggested budget allocation. .this This refers to the credit limit that will be marked as "dynamically used." It may be much smaller than the application amount (for long-term, uncertain applications) or close to the application amount (for short-term, certain applications).

[0047] Understandably, by introducing AI-based predictive dynamic budget allocation weights, this method achieves intelligent and refined budget allocation. It changes the traditional, indiscriminate "full freeze upon application" model, differentiating budget resources based on the actual risk (probability of payment) and the time value of money (expected time) of each transaction. This significantly improves the availability and flexibility of the budget pool, enabling limited budgets to simultaneously support more potential business needs. Especially under budget constraints, it optimizes resource allocation and prevents other urgent business from being unable to proceed due to a few long-term, uncertain, large applications prematurely tying up the entire budget.

[0048] Specifically, based on the aforementioned budget allocation decision, dynamic budget allocation is performed in the budget pool, specifically as follows: Based on the suggested budget allocation in the decision results, the corresponding budget allocation is marked as "dynamically occupied" in the corresponding budget pool. In this state, the budget allocation can still be partially occupied by higher priority business applications, but it is no longer fully included in the available balance for regular approval. The system records the business application identifier, the amount occupied, the dynamic occupancy weight, the occupancy timestamp, and the predicted reimbursement node information for this occupancy, and associates the occupancy record with the business process instance of the business application.

[0049] It should be further explained that this step defines the specific implementation logic and state management rules of "dynamic occupancy" at the system level, which is the key to connecting intelligent decision-making with physical storage.

[0050] (1) Budget Pool Status Flags: The budget pool can be represented as a record in the database, containing fields such as: Budget Pool ID, Project ID, Budget Item, Total Budget, Frozen Amount, Actual Paid Amount, Dynamically Occupied Total Amount, Available Balance, etc. When dynamic occupancy is executed, the system performs the following operations: a) Calculate the new "Dynamic Total Amount": New Dynamic Total Amount = Original Dynamic Total Amount + Proposed Budget Amount ( ).

[0051] b) Update "Available Balance": The calculation logic for available balance is: Available Balance = Total Budget - Frozen Amount - Actual Paid Amount - Dynamically Used Total. However, the key point is that the credit limit in the "Dynamically Used" state (i.e., the total dynamically used amount) is not completely unavailable. When determining whether a subsequent application exceeds the budget, the system can add "Available Balance + Total Dynamically Used Total". The "reusability factor" represents the actual available credit limit. The reusability factor is a parameter configured globally or per budget pool; for example, 0.5 means that half of the dynamically allocated credit limit can be reused in an emergency. This reflects "flexibility."

[0052] c) Create a separate "Dynamic Occupation Record" for this occupancy. This record is a budget occupancy log and contains the following core fields: Occupation Record ID, Associated Business Application Number, Corresponding Budget Pool ID, and Occupancy Amount ( ), dynamic weighting ( ), Occupancy status (initially "Occupied"), Occupancy start time, and predicted write-off time (based on the predicted expected time). The calculation includes a "priority" identifier (which can be automatically assigned based on business type, application amount, applicant's job level, etc.).

[0053] (2) Association and Function of Occupancy Records: This dynamic occupancy record is strongly associated with a specific business process instance (such as a purchase requisition form, PR) through the "associated business application number". This means that the occupancy of the budget and its subsequent release and adjustment can be precisely traced back to the specific business document. The collection of all dynamic occupancy records for the same budget pool constitutes the "flexible occupancy layer" of that budget pool. When a new, higher-priority business application (such as one marked as "urgent" or from a higher-level project) requires budget occupancy, but the current "available balance" is insufficient, the system can scan the lower-priority records in the "flexible occupancy layer" that are "occupied" and attempt to "borrow" a portion of the quota from them. The borrowing rule can be: only the occupied amount can be borrowed. The part of (1-dynamic weight allocation) is due to the weight. High occupancy rates indicate a greater likelihood of eventual payment and should therefore be given more protection.

[0054] Understandably, by clearly defining the "dynamic occupancy" status and its management rules, this method achieves "flexible" budget management at the technical level. It both locks in a portion of the budget through occupancy records to control risk and allows for limited reuse within the occupancy limit, significantly improving the flexibility and efficiency of budget resources. This mechanism is particularly suitable for project-based and matrix-style management enterprises, capable of handling scenarios where multiple projects are running concurrently and budgets require flexible allocation. Simultaneously, detailed occupancy records provide a high-quality data foundation for end-to-end budget traceability, analysis, and continuous optimization of AI models.

[0055] Specifically, upon detecting a payment or reconciliation event related to the aforementioned business, the AI ​​model dynamically adjusts the allocation to the budget pool based on the details of the event, including: The payment or cancellation events include invoice registration verification, service completion confirmation, contract payment execution, or business application cancellation. When an invoice registration verification event is detected, the AI ​​model obtains the invoice amount, tax rate, and actual supplier information and matches them with the original business application. If there is a price-tax difference, the model automatically calculates the amount of budget to be released or added from the "dynamically occupied" state according to the preset price-tax difference handling strategy, and updates the actual occupied value in the budget pool.

[0056] It should be further explained that this step details how the AI ​​model drives the dynamic adjustment of budget allocation in a typical reconciliation event such as invoice verification. This is one of the core components for achieving "dynamic budget reconciliation".

[0057] (1) Event Listening and Data Acquisition: The system listens for the "Invoice Verification Completed" event via the event bus, Database Change Capture (CDC) tool, or by directly integrating with the financial system interface. Upon detecting the event, the reconciliation AI model is triggered. The model retrieves the original business application number associated with the invoice (achieved through the association ID passed between the front-end and back-end systems in the process) and reads the key invoice information: amount excluding tax. The data includes the tax amount, total price including tax, invoice code, and the name and tax number of the actual invoicing supplier. Simultaneously, the model retrieves the dynamic occupancy record corresponding to this business application from the database to obtain the original application amount. The amount of money that has been dynamically used Information such as the original applicant supplier.

[0058] (2) Matching and Difference Detection: The model first performs a consistency check: a) Supplier Consistency: Compare the actual supplier on the invoice with the original supplier in the application (or whether they are on the list of permitted affiliated suppliers). If they are inconsistent, an exception process will be triggered, which may require manual intervention.

[0059] b) Calculation of Price Difference: Assuming consistent supplier relationships, calculate the price-tax difference. Differences can arise from various reasons: changes in purchase quantity, unit price discounts, adjustments to tax calculations, etc. The model calculates the difference value. . This indicates that the invoice amount exceeds the application amount. This indicates that the invoice amount is lower than the application amount.

[0060] (3) Automatic adjustment of budget occupancy: Based on the preset price-tax difference handling strategy, the model automatically determines how to adjust the original dynamic occupancy record. A typical strategy is as follows: ①If (in This is a tolerance threshold, for example, set to 5% or 10%, and If the invoice amount is less, the system will automatically calculate the amount that needs to be released. Then, update the amount occupied in the corresponding dynamic occupancy record to... The status is marked as "partially released." Simultaneously, the "total dynamically allocated amount" in the budget pool is reduced. This makes more budget available.

[0061] ②If ,and If the invoice amount slightly exceeds the limit, the system will automatically calculate the additional amount that needs to be used. Then, update the amount occupied in the dynamic occupancy record to... In the budget pool, increase the "Total Dynamically Used Amount". .

[0062] ③If If the difference exceeds the tolerance range, the system will not adjust automatically. Instead, it will mark the event as a "significant difference," generate an early warning notification, and forward it to the finance personnel or the original approver for manual review. After manual review and confirmation, a significant release or addition of occupancy will be implemented based on the review results, which may trigger a change process for the original business application.

[0063] Understandably, by automatically handling price-tax discrepancies in invoice verification using AI models, this method automates and intelligently processes budget reimbursement, significantly reducing manual operations and communication costs for finance personnel. Preset tolerance thresholds and strategies enable the system to flexibly handle minor fluctuations common in business operations, improving process efficiency. Simultaneously, the systemic interception and manual review mechanisms for significant discrepancies ensure the seriousness and accuracy of budget control, preventing control loopholes that may arise from automated system processing.

[0064] Specifically, upon detecting a payment or reconciliation event related to the aforementioned business, the AI ​​model dynamically adjusts the allocation to the budget pool based on the details of the event, and also includes: When an event is detected where a business application is actively withdrawn, the AI ​​model determines, based on the process node of the business application, the reason for withdrawal, and the probability of re-initiating after a previous withdrawal, whether to immediately release all dynamically occupied budget quotas or to retain a portion of the quotas in a "suspended" state for a preset period of time to deal with possible emergency re-applications.

[0065] It should be further explained that this step provides an intelligent budget release strategy for the common but easily overlooked scenario of "business application withdrawal", which further demonstrates the "intelligence" of dynamic management.

[0066] (1) Event Listening and Context Acquisition: When a user actively withdraws a submitted application (such as a purchase application or expense application) in the business system, the system generates an "Application Withdrawal" event. The reconciliation AI model is triggered and acquires the following context information: the unique identifier of the withdrawn application, the process node at which the application was withdrawn (e.g., "Submitted for departmental approval", "Under approval", "Approved but not executed"), the withdrawal reason filled in by the user (selected from a predefined list, such as "Request cancellation", "Find a better supplier", "Budget adjustment", etc.), the timestamp of the withdrawal operation, and the dynamic occupancy record currently associated with the application.

[0067] (2) AI Model Decision Logic: The model determines the budget release strategy based on rules and predictions. The decision-making process can be broken down as follows: a) First, check the reason for cancellation. If the reason is "permanent cancellation of the requirement" or "project termination," the model directly decides to immediately release all dynamically allocated quotas. The system updates the status of the corresponding quota record to "released," clears the allocated amount, and reduces the corresponding "total dynamically allocated quota" in the budget pool.

[0068] b) If the reason for withdrawal is temporary or foreseeable, such as "scheme adjustment" or "price renegotiation," the model will invoke a predictive sub-model. This sub-model, based on historical data, predicts the probability that an application withdrawn for this reason at the current project stage will be re-initiated within the next N days (with the amount changing within a certain range). Historical data shows that cancellations in some cases are an intermediate stage in business adjustments.

[0069] c) Based on the predicted probability Depending on the amount applied for, the model makes the following decisions: ①If (For example If the applied amount is large (e.g., greater than 100,000 yuan), the decision will be "partial suspension". For example, 50% of the occupied credit limit will be released, the remaining 50% will be marked as "suspended", and a retention period will be set. (For example, 7 calendar days). During the retention period, this "suspended" credit limit is reserved for the applicant or the project and can be quickly reused for new applications without having to go through the full AI authentication process again (but a simple consistency check is required). After the retention period, the remaining credit limit will be automatically released regardless of whether it has been reused.

[0070] ②If (For example )but If the amount is small, the decision is to "release all immediately". ③ If The decision would then be "release all immediately".

[0071] d) In addition, if an application is withdrawn at the “Approved but not executed” node, since it has been approved, indicating a high level of business necessity, the model will tend to have a higher suspension rate and a longer retention period.

[0072] Understandably, intelligent budget release management for withdrawal requests avoids the efficiency losses that might result from a "one-size-fits-all" approach. Immediate release facilitates rapid budget recovery and improves overall utilization. The "suspended" mechanism provides a buffer for reasonable business adjustments, avoiding duplicate approvals and budget occupancy processes caused by temporary withdrawals followed by re-initiation, thus enhancing business continuity and user experience. This refined strategy fully considers the complexity of actual business scenarios and is a significant manifestation of AI-powered dynamic reimbursement capabilities.

[0073] Specifically, upon detecting a payment or reconciliation event related to the aforementioned business, the AI ​​model dynamically adjusts the allocation to the budget pool based on the details of the event, further including: In scenarios where contracts are paid in multiple phases, when an event is detected that a certain phase payment has been completed, the AI ​​model automatically calculates the difference between the amount payable for this phase and the original dynamic occupancy limit based on the payment plan stipulated in the contract and the evaluation information of the current phase completion quality. It then completes the actual write-off of the corresponding amount and generates a new dynamic budget occupancy suggestion based on the updated project progress for the next payment phase, thereby updating the occupancy status in the budget pool.

[0074] It should be further explained that this step solves the problem of dynamic budget management for long-term, phased contracts (such as project-based contracts and multi-year service contracts), and achieves deep synchronization between budget control and project execution progress.

[0075] (1) Scenario Initialization: For a contract with installment payments, when the first payment is requested (such as an advance payment), the system will create a primary dynamic occupancy record. The occupancy amount may be a certain percentage of the total contract amount, and the occupancy weight is predicted based on the overall payment risk of the contract. At the same time, the system will record the contract's payment plan, for example: Phase 1 (after contract signing): 20% payment; Phase 2 (complete solution design): 30% payment; Phase 3 (project goes live): 40% payment; Phase 4 (warranty period expires): 10% payment. Each phase corresponds to an expected completion time.

[0076] (2) Stage Payment Event Handling: When the project progresses to "Completion of Solution Design" and triggers the Stage 2 payment application, the business system submits a payment application (which can be regarded as a special business application). The AI ​​agent will perform real-time authentication and generate a budget allocation decision for Stage 2. Assume that the approved allocation amount is 30% of the total contract price. When the finance system actually completes the payment, a "Stage Payment Completed" event is generated.

[0077] (3) AI Model Cancellation and Re-prediction: After the AI ​​model detects this event, it performs the following operations: a) Actual write-off: The model locates the dynamic occupancy record associated with the contract in Phase 2, updates its status to "write-off," removes its occupancy amount from the "Total Dynamic Occupancy" in the budget pool, and adds it to the "Actually Paid Amount." This completes the final write-off of the budget occupancy for that phase.

[0078] b) Generate new occupancy recommendations for the next phase: The model obtains the latest project progress, such as a "design completion rate" of 100% (from the project management system), and may obtain quality evaluation scores for deliverables in this phase. ( The model reassesses the payment risk in subsequent stages. For example, if the current stage is completed with high quality ( If so, the model may predict the payment probability in the subsequent stage (stage 3). This will increase (e.g., from 0.8 to 0.9). Simultaneously, considering whether the current project is delayed, the expected payment time for the next phase will be re-forecasted. .

[0079] c) Occupancy Status Update: Based on the new predicted payment probability and expected time Based on the aforementioned weighting formula, recalculate the dynamic budget occupancy weight for the next stage (Stage 3). Then, the recommended budget allocation for Phase 3 is calculated: The system then creates a new dynamic occupancy record in the budget pool for Phase 3, with an occupancy limit of [amount missing]. This will update the "Dynamic Total Amount Used" in the budget pool. The amount used in this new record may differ from the amount in Phase 3 of the original contract plan because it has been intelligently adjusted based on the latest project execution status.

[0080] Understandably, this approach transforms static contract payment plans into dynamic budget allocation sequences linked to project execution quality and schedule. Upon completion of each phase, the used budget is automatically and accurately reimbursed, and the budget required for the next phase is intelligently predicted and allocated. This not only achieves comprehensive and refined budget control but also dynamically adjusts budget allocation based on actual project performance. For example, when project execution faces risks, the budget allocation for the next phase will be more conservative, thus providing early warning and controlling financial risks; when project execution is successful, more budget flexibility is released. This represents a profound manifestation of integrated business and financial management in project financial management.

[0081] Specifically, when the AI ​​agent generates budget allocation decisions in real time, it also performs the following steps: For specific business types, key entities in the application information are identified and compared with a risk knowledge base. If a high-risk entity is identified, an enhanced review process is triggered. This enhanced review process includes automatically retrieving the entity's past transaction dispute records and market price fluctuation data, simulating the impact of different approval decisions on the overall project budget expenditure over a future period, and finally generating a report containing risk warnings and multi-scenario decision suggestions for authorized nodes to review.

[0082] It should be further explained that this step adds in-depth risk insight and forward-looking simulation capabilities to the AI ​​agent's real-time decision-making, enabling it to go beyond single-point compliance judgments and perform more complex risk-return analyses.

[0083] (1) High-risk entity identification: The AI ​​agent integrates an entity identification and risk scanning module. For specific business types, such as "large-scale equipment procurement," "external consulting service procurement," and "market promotion cooperation," the system defines key entity types that require focused attention, such as "supplier name," "brand and model," and "service content keywords." When analyzing application information, the AI ​​agent uses Natural Language Processing (NLP) technology to extract these key entities. Subsequently, the extracted entities are compared in real time with internal and external risk knowledge bases.

[0084] a) Internal risk knowledge base: includes the company's internal supplier black / grey list, supplier records of past contract disputes or quality problems, and procurement categories that have been marked as high-risk by audits.

[0085] b) External risk data: By accessing external data services through API interfaces, you can query suppliers' business risk information (such as administrative penalties and legal proceedings), market price trend data of specific materials (such as price volatility in the past six months), or average fee levels of specific service industries.

[0086] (2) Enhanced review process triggering: If the supplier in the application is identified as being on an internal blacklist, or the recent price volatility of the purchased materials exceeds a threshold (e.g., 30%), or the service fee rate applied for is much higher than the industry average, the AI ​​agent will determine that the application involves a "high-risk entity" and automatically trigger the enhanced review process, rather than giving a simple decision of pass / reject.

[0087] (3) Enhance review and enforcement: ① In-depth data retrieval: The system automatically retrieves detailed information about the high-risk entity, such as all transaction records with the supplier over the past three years, payment terms, dispute resolution status, and the purchase price series of the material over the past year.

[0088] ② Decision Impact Simulation: The AI ​​agent incorporates a lightweight budget impact simulation model. This model takes the current application as input and simulates the potential chain reactions that might result from approving or rejecting it over a future period (e.g., this quarter or year). For example, simulating approving this high-priced purchase: it might consume a significant amount of budget, impacting other planned purchases; however, it might also secure crucial technical support. Simulating rejecting the application and suggesting alternatives: it might extend the project timeline, incurring delay costs. The simulation generates 2-3 different scenarios (e.g., "Approve the original application," "Reject and activate alternative supplier A," "Suspend and re-tender") and estimates the impact on total project cost, project schedule, and relevant budget pool balance under each scenario.

[0089] (4) Generating a Decision Support Report: Finally, the AI ​​agent integrates the aforementioned risk information, data comparison results, and multi-scenario simulation analysis into a structured "Enhanced Review Report." The report does not replace human decision-making but provides clear, quantifiable risk warnings (e.g., "This supplier has been involved in 3 contract lawsuits in the past two years; risk level: high") and offers suggested decision options with estimated impact. This report, along with the business application, is pushed to higher-level authorization nodes (such as department directors or CFOs) for final review and approval.

[0090] Understandably, this step combines AI's real-time judgment capabilities with the in-depth decision-making abilities of human experts. For routine, low-risk applications, the AI ​​agent can process them quickly and automatically, improving efficiency. For high-risk, complex decisions, the AI ​​agent acts as a "super assistant," automatically completing tedious data collection, comparative analysis, and simulation exercises. It condenses key information and recommendations into reports, helping the authorizing person make more scientific and evidence-based approval decisions, thereby achieving a better balance between risk control and supporting business development.

[0091] Specifically, the method also includes an online dynamic optimization step for the budget strategy: Continuously collect all business applications and their subsequent payment verification data to form decision-outcome sample pairs; periodically use new samples to train the budget occupancy weight prediction model and compliance judgment model, and optimize their parameters; deploy the optimized model parameters to the AI ​​agent in a hot update manner to realize the adaptive adjustment of the budget control strategy based on actual business feedback. The full-link data is anonymized before being used for training to remove directly personal identification information.

[0092] It should be further explained that this step ensures that the entire intelligent budget control system has the ability to learn and continuously optimize itself, enabling it to adapt to changes in business operations and the external environment, and maintain long-term effectiveness and accuracy.

[0093] (1) Data Collection and Sample Construction: During operation, the system automatically and continuously records the entire data chain of each business application, forming a complete "digital twin" record. For each application, the recorded data includes: a) Input features ( ): All features used by the AI ​​agent in decision-making, including application information, contextual information, real-time budget balance, historical behavior data, etc.

[0094] b) Decision-making actions ( The decision results provided by the AI ​​agent include approval / rejection flags, suggested usage limits, dynamic usage weights, and risk markers.

[0095] c) Business Results The actual business outcomes following this application include: whether payment was ultimately made, the actual payment amount, the actual payment time, whether contract changes occurred, whether disputes arose, and the final cost-benefit evaluation (if any). This data needs to be obtained through integration with downstream procurement systems, financial payment systems, project management systems, etc.

[0096] Will( , By using the difference between the final actual business outcome and the initial decision as input, and taking this difference as the learning signal, a massive number of "decision-outcome" sample pairs can be constructed. For example, a sample could be: the initial AI decision based on weights... The application consumed a portion of the budget, but was ultimately cancelled and not paid. This example suggests to the model that the predicted payment probability for this type of application may have been too high, and the weights could have been lowered.

[0097] (2) Regular Model Training and Optimization: The system is set to a training cycle, such as weekly or monthly. At the end of each cycle: ① Data preprocessing: Clean, anonymize (e.g., map employee IDs and supplier names to anonymous identifiers) and perform feature engineering on the end-to-end data collected in the new cycle to form a new training dataset.

[0098] ② Model retraining: Using accumulated historical data (e.g., data from the past 24 months) plus new data from the current period, retrain the two core models: i. Budgetary Occupation Weighted Prediction Model: This is a regression model whose optimization objective is to optimize the predicted payment probability. and expected time Closer to what actually happened and The mean squared error (MSE) or mean absolute percentage error (MAPE) can be used as the loss function.

[0099] ii. Compliance and Reasonableness Judgment Model: This is a classification (or ranking) model whose optimization goal is to make the model's judgment (approve / reject / transfer to manual review) more consistent with the "optimal decision" verified after the fact. The "optimal decision" can be defined by reverse deduction based on actual business results. For example, for applications that ultimately lead to disputes or serious cost overruns, the "optimal decision" should be "reject".

[0100] ③ Model Evaluation and Validation: Evaluate the performance of the newly trained model on an independent test set, using metrics such as accuracy, recall, and AUC. Deployment only begins when the new model outperforms the current online model.

[0101] (3) Hot Update Deployment: Using non-disruptive deployment technologies such as hot update or blue-green deployment, the optimized model parameter files are updated to the AI ​​agent service in the production environment. The update process is transparent to the front-end business systems and does not affect ongoing approval processes. After the update is completed, new decisions will be made based on the latest model that is more in line with current business trends.

[0102] Understandably, by introducing online dynamic optimization steps, this system evolves from a static, historical rule-based system into a dynamic, self-learning intelligent system. It can continuously learn from actual business results, automatically adjusting its prediction and judgment logic, resulting in more accurate budget allocation and more sensitive risk identification. This effectively solves the problems of rigid strategies and difficulty in adapting to business changes in traditional rule engines, ensuring the long-term vitality and effectiveness of the budget control system, and is key to achieving sustainable intelligent management.

[0103] This invention provides another embodiment, which offers a real-time budget control and AI dynamic reconciliation system embedded in business processes. The real-time budget control and AI dynamic reconciliation system embedded in business processes includes: (1) Deployment module, used to deploy AI agents for real-time budget control in business systems.

[0104] It should be further noted that this system is a device embodiment corresponding to the aforementioned method, describing the system architecture required to achieve real-time budget control and AI dynamic reimbursement embedded in business processes in a modular manner. Each module can be deployed on the same server or multiple distributed servers, communicating via an internal network or service bus. The deployment module provides deployment, configuration, and registration functions for the AI ​​agent. It can be a deployment platform or a set of scripts responsible for deploying the AI ​​agent service (typically in the form of Docker containers or microservices) to enterprise servers or cloud environments and registering its service address (such as an API endpoint) with the enterprise service registry. Simultaneously, this module also provides adapters or SDKs for different business systems (OA, ERP, SRM), facilitating rapid integration and invocation by these systems.

[0105] (2) Triggering module, used to trigger the AI ​​agent when the business system receives the instruction to initiate the business.

[0106] It should be further explained that the triggering module is embedded in or closely related to various business systems. It listens for key events within the business systems (such as "form submission" and "workflow creation"). When a predetermined event is detected, the triggering module is responsible for collecting the complete context data of the business application, encapsulating it into a standard format (such as JSON), and triggering the AI ​​agent service located remotely or locally via a network call (such as an HTTP POST request). It is also responsible for handling communication reliability issues such as call timeouts and retrying on failures.

[0107] (3) Decision module, built into the AI ​​agent, is used to generate budget occupancy decisions in real time based on the application information of the business and the associated budget strategy.

[0108] It's worth noting that the decision-making module is the core of the AI ​​agent and typically runs on a server with high computing power. It receives request data from the triggering module and internally integrates a data acquisition submodule, a feature engineering submodule, a model inference submodule, and a rule engine submodule. The data acquisition submodule is responsible for retrieving the required data in real time from the budget system, historical databases, etc. The feature engineering submodule transforms the raw data into feature vectors usable by the model. The model inference submodule loads the aforementioned budget occupancy weight prediction model and compliance judgment model and performs forward inference calculations. The rule engine submodule executes hard compliance rules. These submodules work together to ultimately generate a structured budget occupancy decision and return it to the caller (business system).

[0109] (4) Execution module, used to execute dynamic budget allocation in the budget pool based on the budget allocation decision.

[0110] It's worth noting that the execution module interacts directly with the enterprise's budget management system or core financial database. It receives instructions from the business system (which has received AI-driven decisions) and performs an "update" operation on the specified budget pool data records. Specifically, it modifies the "Total Dynamic Occupation" field and inserts a new "Dynamic Occupation Record." This module needs to ensure the atomicity and consistency of data updates in concurrent scenarios, typically achieved using database transaction mechanisms or distributed locks.

[0111] (5) Adjustment module, which is used to dynamically adjust the amount of the budget pool occupied by the AI ​​model according to the details of the event when a payment or reimbursement event related to the business is detected.

[0112] It's important to further clarify that the adjustment module is an independent event-driven service. It listens to event message queues from multiple systems within the enterprise (such as the financial system, procurement system, and project management system). When it detects events such as "Invoice verification completed," "Payment order completed," "Contract milestone completed," or "Application order cancelled," the adjustment module is activated. It parses the event content, locates the associated original business application and dynamic occupancy record, calls the internal AI reconciliation model (or logic) to perform difference analysis and adjustment calculations, and then calls the execution module or a similar database operation module to update the status and amount of the original dynamic occupancy record (e.g., marking it as "partially released" or "reconciled"), and updates the total budget pool accordingly. This module is also typically responsible for generating budget adjustment logs and sending related notifications.

[0113] Understandably, this system, through a clear modular design, decomposes the complex intelligent budget control process into discrete, highly cohesive, and loosely coupled functional components. The deployment module solves the system integration problem, the triggering module achieves seamless embedding with business processes, the decision-making module provides the intelligent core, and the execution and adjustment modules ensure precise synchronization between budget status and business status. This architecture gives the system good scalability, maintainability, and reliability, enabling it to support large-scale concurrent business requests at the enterprise level. It is the technical foundation and organizational guarantee for implementing the aforementioned methods.

[0114] Specifically, to more clearly illustrate the present invention, a concrete operational case is provided below. This case uses "an employee of a company's R&D department applying to purchase an AI training server" as an example to demonstrate the complete process from business application, real-time authentication, dynamic budget allocation, to invoice verification and cancellation, application withdrawal and reapplication, and includes specific data calculations.

[0115] It should be further explained that this case simulates a real-world business and financial management scenario. The company has a budget item for "Hardware Equipment Procurement" under "Project A for 2026," with a total budget of RMB 100,000. As of the time of this case, the "Actually Paid Amount" in this budget pool is RMB 20,000, the "Dynamically Occupied Total Amount" is RMB 10,000 (occupied by other previous applications), and there is no "Frozen Amount." Therefore, the current "Available Balance" is calculated as: Total Budget - Actual Paid Amount - Dynamically Occupied Total Amount = RMB 100,000 - RMB 20,000 - RMB 10,000 = RMB 70,000.

[0116] (1) Business application and AI real-time authentication: Zhang San, an employee in the R&D department, submitted a purchase requisition in the company's procurement management system (business system) to purchase an AI training server with the market model "NV-AI-Unit". The requisition information included: project identifier "Project A", budget item "Hardware Equipment Procurement", requisition amount 30,000 yuan, business type "Fixed Asset Procurement", and supplier "Company B". The submission action was "Save and Submit for Review". At this moment, a trigger module deployed in the procurement system was activated, packaging the above requisition information along with contextual information such as Zhang San's job level (Senior Engineer), current date (March 30, 2026), and business urgency level ("High"), and calling the AI ​​agent's service interface.

[0117] The AI ​​agent's decision-making module is triggered. It first performs a hard compliance check: Company B is found to be on the list of qualified suppliers, and Zhang San has the authority to apply for "fixed asset procurement." The single transaction amount of 30,000 yuan does not exceed his job level limit (50,000 yuan), so the rule check passes. Next, a reasonableness analysis and flexible threshold judgment are performed: the decision-making module queries the budget pool in real time and finds an available balance of 70,000 yuan. The company's flexible threshold for "fixed asset procurement" is 15%. The percentage of this application amount is calculated as: 30,000 / 70,000 ≈ 42.9%, exceeding the 15% flexible threshold, thus triggering a deeper analysis.

[0118] The decision-making module retrieves data from the historical database: Zhang San initiated 5 similar purchase requests in the past 12 months, with an actual payment conversion rate of 100% and an average execution deviation rate (|actual payment amount - request amount| / request amount) of 3%, demonstrating good performance. Project A is currently in the "critical development phase." Through access to external data services, it was learned that the market price of the "NV-AI-Unit" server model has remained stable over the past three months. The machine learning model within the decision-making module makes predictions based on these characteristics (business type, project stage, applicant history, market conditions, etc.), outputting a predicted payment probability. ( ), predict the expected payment time Days (calculated from the application date). Here. This represents the probability, as predicted by the model, that the application will ultimately be actually paid out. This is the predicted average time from application submission to payment date.

[0119] Next, calculate the dynamic budget occupancy weight. The formula described in the examples is used: .in, It is the balance coefficient, with a preferred value of 0.7, indicating a greater emphasis on the probability of payment; It's about time. The decay function, in this example, is used. ,in Represents the natural logarithm. Calculation process: Then, calculate the preliminary budget allocation: Yuan. The minimum threshold for "fixed asset procurement" set by the company. It is 50% of the applied amount, which is 15,000 yuan. Because... Therefore, the final recommended budget allocation is... Yuan.

[0120] Based on the above analysis, the decision-making module comprehensively judges the application to be reasonable and generates a decision result: approved, and recommends dynamic budget allocation of 21,507 yuan with a weight of 0.7169. After receiving this result, the procurement system displays "Budget pre-approval passed", and the application is transferred to the subsequent approval stage.

[0121] (2) Execution of dynamic budget allocation: Based on the decision, the execution module increases the "Dynamically Occupied Total" by 21,507 yuan in the "Project A - Hardware Equipment Procurement" budget pool, updating it from 10,000 yuan to 31,507 yuan. Simultaneously, a dynamic occupancy record is created, linked to this purchase order number, recording the occupied amount as 21,507 yuan, a weight of 0.7169, a status of "Occupied," and the predicted write-off time (approximately May 29, 2026). At this point, the "Available Balance" of the budget pool is updated to: 100,000 - 20,000 - 31,507 = 48,493 yuan. It is worth noting that this 48,493 yuan is a strictly available balance. However, due to the existence of the 31,507 yuan "Dynamically Occupied Total," and the fact that a portion of it (based on weight) can be reused by higher-priority requests, the budget resources still retain some flexibility.

[0122] (3) AI-powered dynamic verification (invoice verification scenario): Assuming the purchase request is subsequently approved, Zhang San's department signs a contract with supplier "Company B" for 30,000 yuan. One month later, the server arrives and is accepted. Zhang San registers a VAT invoice from "Company B" in the financial system for verification. The invoice amount excluding tax is 28,301.89 yuan, the tax amount is 3,679.25 yuan, and the total price including tax is 31,981.14 yuan. The financial system completes the invoice verification operation, triggering the "Invoice Verification Completed" event.

[0123] Upon receiving this event, the adjustment module activates the reimbursement AI model. The model retrieves invoice information and compares it with the dynamic occupancy record associated with the original purchase requisition. First, it verifies supplier consistency (which is confirmed). Then, it calculates the price-tax difference. : Yuan. This indicates that the invoice amount exceeds the requested amount. Calculate the discrepancy rate: Assume the company sets a tolerance threshold. The threshold was 5%. Since 6.6% > 5%, exceeding the tolerance range, the reimbursement model did not process it automatically. Instead, it marked the event as a "significant discrepancy," generating a warning notification: "Invoice amount exceeds applied amount by 6.6%, exceeding budget by 1,981.14 yuan," and forwarded it to the finance department for processing. After contacting the business department for verification, the finance department confirmed that it was due to increased equipment transportation insurance costs, which were considered reasonable expenses, and approved the discrepancy within the system.

[0124] After receiving manual approval, the reimbursement model executes adjustments. Because the invoice amount exceeds the original request, additional budget allocation is required. First, the amount requiring additional allocation is calculated. The original dynamic usage limit was 21,507 yuan, corresponding to the original application of 30,000 yuan. Based on the proportion, the current actual budget amount to be used should be: Yuan. Therefore, additional credit is required. The system checks that the available balance in the budget pool is sufficient and executes an additional payment. The dynamic occupancy record is updated, changing the occupied amount from 21,507 yuan to 22,925 yuan. Simultaneously, in the budget pool, the "Total Dynamic Occupancy" is increased by 1,418 yuan from 31,507 yuan to 32,925 yuan. Subsequently, the finance department completes the payment. Upon detecting the "Payment Completed" event, the adjustment module updates the status of the occupancy record to "Write-off," deducts the occupied amount of 22,925 yuan from the "Total Dynamic Occupancy," and adds it to the "Actually Paid Amount." At this point, the "Actually Paid Amount" becomes 20,000 + 22,925 = 42,925 yuan, the "Total Dynamic Occupancy" reverts to 32,925 - 22,925 = 10,000 yuan (i.e., occupancy from other applications), and the "Available Balance" is updated to 100,000 - 42,925 - 10,000 = 47,075 yuan. The entire reconciliation and adjustment process is completed automatically or semi-automatically by the system (after manual confirmation of differences).

[0125] (4) AI-powered dynamic verification (application withdrawal and reapplication scenarios): In a parallel case, suppose Li Si submits a "marketing promotion service" application for 50,000 yuan, linked to the "Project B - Marketing Promotion Fee" budget. After AI agent authentication, it predicts the probability of payment. Expected time Day, calculate weight The application suggested a limit of 32,500 yuan, which was successfully used. Three days later, due to adjustments in the promotion strategy, Li Si withdrew the application during the approval process, citing "plan adjustment" as the reason for withdrawal.

[0126] The adjustment module detected an "application withdrawal" event. The reconciliation model analyzed the context: the application was at the "submitted and awaiting approval" stage, and the reason for withdrawal was "plan adjustment." The model queried historical data and predicted the probability that, in "marketing promotion" type applications withdrawn at the "awaiting approval" stage due to "plan adjustment," they would be resubmitted within 7 days. .because (Threshold here) Given a value of 0.5 and a relatively large amount (50,000 yuan), the model decides to implement a "partial suspension" strategy: immediately release 50% of the occupied quota (i.e., 16,250 yuan), and mark the remaining 50% of the occupied quota (16,250 yuan) as "suspended," with a retention period T set to 7 days. The system updates the budget pool, reducing the "dynamic total occupied amount" by 16,250 yuan. Simultaneously, the original occupied record is split into two entries: one with a status of "released," amounting to 16,250 yuan; and another with a status of "suspended," amounting to 16,250 yuan, valid until 7 days later.

[0127] Five days later, Li Si resubmitted the optimized promotion service application, with the amount increased to 45,000 yuan. This triggered AI proxy authentication. The decision-making module discovered that in the budget pool associated with this new application, there was a "temporary" usage record initiated by the same applicant and associated with the same project, which was still valid. While performing regular authentication, the decision-making module would prioritize reusing this "temporary" quota. Assuming the new suggested usage quota calculated by regular authentication is 40,000 yuan, the system will directly use the 16,250 yuan of the "temporary" quota to offset it and generate a new dynamic usage record for the new application, with a usage quota of 40,000 yuan. At this time, the net change in the budget pool is: the original "temporary" record is consumed and becomes part of the new record, and the "total dynamic usage" of the budget pool actually increases by 40,000 - 16,250 = 23,750 yuan in this operation, instead of a completely new 40,000 yuan. This effectively preserved budgetary resources and expedited the resubmission process.

[0128] It is understandable that this specific operational example fully and clearly demonstrates the application process of the method of the present invention in a real business scenario. This is achieved through specific data calculations (such as weighting). (The calculation, determination of the occupancy amount, and comparison of the difference rate) allows those skilled in the art to understand without a doubt how the AI ​​agent makes real-time decisions, how it executes dynamic occupancy, and how the AI ​​reimbursement model automatically adjusts the budget status based on invoice differences and cancellation events. Case studies demonstrate that this invention successfully moves the budget control point forward to the moment of business application and achieves flexible budget occupancy through intelligent prediction (in the case of an application for 30,000 yuan, only 21,507 yuan was flexibly occupied). When business changes occur (changes in invoice amounts, application cancellations), the system can automatically and accurately adjust budget occupancy, achieving deep synchronization and dynamic balance between budget management and business processes, effectively solving the problems of lag and rigidity caused by traditional post-event control.

[0129] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the real-time budget control and AI dynamic reimbursement method for embedded business processes. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include non-volatile storage media and internal memory. The non-volatile storage media may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0130] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0131] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0132] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0133] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for real-time budget control and AI-driven dynamic reconciliation embedded in business processes, characterized in that, include: S1. Deploy an AI agent for real-time budget control in the business system, and trigger the AI ​​agent when the business system receives an instruction to initiate a business; the AI ​​agent generates a budget allocation decision in real time based on the application information of the business and the associated budget strategy; S2. Based on the budget allocation decision, perform dynamic budget allocation in the budget pool; S3. When a payment or reimbursement event related to the business is detected, the AI ​​model dynamically adjusts the amount of the budget pool occupied based on the details of the event.

2. The method according to claim 1, characterized in that, The AI ​​agent is triggered when the business system receives an instruction to initiate a business, specifically including: When the business system detects that a user has submitted a purchase request, advance payment request, or contract payment request, it calls the interface corresponding to the AI ​​agent and transmits the request information, including project identifier, budget item, amount, and business type, along with user identity and context information of the current business scenario, to the AI ​​agent. The real-time decision-making of the AI ​​agent includes: The system retrieves the real-time budget balance bound to the project identifier and budget item from the business database, and obtains historical business data associated with the user and the project within a preset time period. Based on a preset authentication model that integrates multiple enterprise expense control rules, the system performs a fusion analysis on the application information, historical business data, and current context information to determine whether the application is compliant and whether it exceeds a preset flexible threshold. When the application is approved, the AI ​​agent generates a decision result containing an approval identifier and a suggested occupancy plan; otherwise, it generates a decision result containing a rejection identifier and a reason.

3. The method according to claim 2, characterized in that, The AI ​​agent generates budget allocation decisions in real time based on the application information and associated budget strategy of the business. The specific steps for generating the suggested allocation plan include: Based on the business type, current project stage, and payment conversion rate records of similar past businesses, a machine learning model is invoked to predict the probability and expected time of cancellation, modification, or final payment completion of this application in subsequent business stages. Based on the probability and expected time, a dynamic budget allocation weight between 0 and 1 is calculated. This weight is positively correlated with the predicted final payment probability and negatively correlated with the expected time. Based on the application amount and the dynamic budget allocation weight, a preliminary budget allocation amount is calculated. Combined with preset minimum allocation thresholds for different business types, a final recommended budget allocation amount is determined to form a recommended allocation plan.

4. The method according to claim 3, characterized in that, Based on the aforementioned budget allocation decision, dynamic budget allocation is performed in the budget pool, specifically as follows: Based on the suggested budget allocation in the decision results, the corresponding budget allocation is marked as "dynamically occupied" in the corresponding budget pool. In this state, the budget allocation can still be partially occupied by higher priority business applications, but it is no longer fully included in the available balance for regular approval. The system records the business application identifier, the amount occupied, the dynamic occupancy weight, the occupancy timestamp, and the predicted reimbursement node information for this occupancy, and associates the occupancy record with the business process instance of the business application.

5. The method according to claim 4, characterized in that, When a payment or reconciliation event related to the aforementioned business is detected, the AI ​​model dynamically adjusts the allocation to the budget pool based on the details of the event, including: The payment or cancellation events include invoice registration verification, service completion confirmation, contract payment execution, or business application cancellation. When an invoice registration verification event is detected, the AI ​​model obtains the invoice amount, tax rate, and actual supplier information and matches them with the original business application. If there is a price-tax difference, the model automatically calculates the amount of budget to be released or added from the "dynamically occupied" state according to the preset price-tax difference handling strategy, and updates the actual occupied value in the budget pool.

6. The method according to claim 5, characterized in that, When a payment or reconciliation event related to the aforementioned business is detected, the AI ​​model dynamically adjusts the amount allocated to the budget pool based on the details of the event, and also includes: When an event is detected where a business application is actively withdrawn, the AI ​​model determines, based on the process node of the business application, the reason for withdrawal, and the probability of re-initiating after a previous withdrawal, whether to immediately release all dynamically occupied budget quotas or to retain a portion of the quotas in a "suspended" state for a preset period of time to deal with possible emergency re-applications.

7. The method according to claim 6, characterized in that, Upon detecting a payment or reconciliation event related to the aforementioned business, the AI ​​model dynamically adjusts the allocation to the budget pool based on the details of the event, further including: In scenarios where contracts are paid in multiple phases, when an event is detected that a certain phase payment has been completed, the AI ​​model automatically calculates the difference between the amount payable for this phase and the original dynamic occupancy limit based on the payment plan stipulated in the contract and the evaluation information of the current phase completion quality. It then completes the actual write-off of the corresponding amount and generates a new dynamic budget occupancy suggestion based on the updated project progress for the next payment phase, thereby updating the occupancy status in the budget pool.

8. The method according to claim 1, characterized in that, When the AI ​​agent generates budget allocation decisions in real time, it also performs the following steps: For specific business types, key entities in the application information are identified and compared with a risk knowledge base. If a high-risk entity is identified, an enhanced review process is triggered. This enhanced review process includes automatically retrieving the entity's past transaction dispute records and market price fluctuation data, simulating the impact of different approval decisions on the overall project budget expenditure over a future period, and finally generating a report containing risk warnings and multi-scenario decision suggestions for authorized nodes to review.

9. The method according to claim 1, characterized in that, The method also includes an online dynamic optimization step for the budget strategy: continuously collecting full-link data of all business applications and their subsequent payment verification to form decision-outcome sample pairs; periodically using new sample pairs to train the budget occupancy weight prediction model and the compliance judgment model, and optimizing their parameters; The optimized model parameters are deployed to the AI ​​agent via hot updates, enabling the budget control strategy to be adaptively adjusted based on actual business feedback. The end-to-end data is anonymized before being used for training to remove directly personally identifiable information.

10. A real-time budget control and AI dynamic reconciliation system embedded in business processes, characterized in that, include: The deployment module is used to deploy AI agents for real-time budget control in business systems; The triggering module is used to trigger the AI ​​agent when the business system receives an instruction to initiate a business. The decision-making module, built into the AI ​​agent, is used to generate budget allocation decisions in real time based on the application information of the business and the associated budget strategy. The execution module is used to perform dynamic budget allocation in the budget pool based on the budget allocation decision; The adjustment module is used to dynamically adjust the amount of the budget pool occupied by the AI ​​model based on the details of the event when a payment or reconciliation event related to the business is detected.

Images