Serverless automated migration system for routine processing using AI agents

Abstract

To suppress increased costs and operational instability caused by intervening in inference processing each time an artificial intelligence agent repeatedly performs a task. [Solution] Based on the execution history, it detects repetitive processing patterns, identifies definitively executable routine processes without using inference processing, automatically generates and deploys processing logic for a serverless execution environment, executes it without inference processing via triggers, and only notifies the artificial intelligence agent to request action when an exception occurs.

Description

Technical Field

[0006] , , , , , ,

[0001] The present invention relates to the automation of business processes by artificial intelligence agents, and particularly to an information processing system in which part or all of the processes repeatedly executed by an artificial intelligence agent are automatically transferred to a serverless execution environment for execution.

Background Art

[0002] In recent years, artificial intelligence agents that execute business processes including operations of external services in response to natural language input have been used. Such an artificial intelligence agent performs inference processing while referring to dialogue history, business data, information of external services, etc., and executes desired processing by API calls of external services or the like.

[0003] However, in processes that are repeatedly executed such as periodic monitoring, synchronization, notification, etc., in a configuration where inference processing is executed every time, due to the consumption of computational resources for inference processing, delay, and fluctuations in inference results, the operation cost and response time may increase.

[0004] On the other hand, in a serverless execution environment, by executing a function or workflow according to an event or schedule, server management is not required and desired processing can be executed. Such an execution form is useful from the viewpoint of executing routine processing quickly and stably.

[0005] Also, technologies related to the generation or optimization of workflows, technologies related to the orchestration of serverless functions, and technologies related to code generation support have been proposed, but a mechanism in which an artificial intelligence agent extracts routine processing that can be executed without going through inference processing based on its own business process history and automatically transfers the routine processing to a serverless execution environment for operation is not sufficiently established.

Prior Art Documents

Patent Documents

[0006]

Patent Document 1

[0007] The processes that artificial intelligence agents perform iteratively may include parts that can be executed deterministically without using inference, such as routine API calls to external services, fixed data transformations, and notifications based on predetermined conditions.

[0008] However, conventionally, even for processes that include such routine parts, artificial intelligence agents often perform inference processing each time to execute a series of processes, which can lead to increased inference costs, response delays, operational instability, and a lack of auditability.

[0009] Furthermore, even when adopting a configuration that executes routine processes in a serverless execution environment, the determination of which processes can be standardized, the implementation, testing, deployment, and handling of exceptions often depend on humans or separate development processes.

[0010] In view of the above circumstances, the present invention aims to provide an information processing technology that enables an artificial intelligence agent to detect repetitive processing patterns based on its own business process execution history, identify routine processes that can be definitively executed without inference processing, automatically transition to a serverless execution environment, and request the artificial intelligence agent to take action when an exception occurs. [Means for solving the problem]

[0011] The information processing system according to the present invention comprises: an artificial intelligence agent that analyzes a user's natural language input and executes business processes including operations on external services; a serverless execution environment; a pattern detection unit that detects repetitive processing patterns based on the execution history of business processes by the artificial intelligence agent; a standardization determination unit that identifies standard processes that can be definitively executed from the detected processing patterns without using inference processing; a generation unit that automatically generates the standard processes as processing logic that can be executed in the serverless execution environment; a deployment unit that automatically deploys the generated processing logic to the serverless execution environment and causes the standard processes to be executed by a predetermined trigger without going through the inference of the artificial intelligence agent; and an exception escalation unit that, when an exception is detected in the execution of the standard processes, notifies the artificial intelligence agent of information regarding the exception and requests action to be taken.

[0012] Furthermore, the present invention can also be understood as an information processing method and program realized by the above system.

[0013] Furthermore, the present invention can be implemented by appropriately combining the specific contents of the execution history, test case generation and production deployment gate, trigger type, notification destination and correction proposal generation in case of exceptions, monitoring and audit log generation, serverless execution environment type, execution method selection by comparing inference cost and execution cost, and individual deployment on a per-customer basis. [Effects of the Invention]

[0014] According to the present invention, an artificial intelligence agent can automatically detect repetitive processing patterns and automatically transfer routine processing that can be executed without inference processing to a serverless execution environment. This allows for stable execution of routine processing while suppressing inference costs and delays during repetitive execution.

[0015] Furthermore, according to the present invention, since escalation to an artificial intelligence agent is possible only when an exception occurs during the execution of routine processing, decisive execution can be maintained under normal circumstances, while in the event of an exception, the reasoning capabilities of the artificial intelligence agent can be used to deal with the situation flexibly.

[0016] In addition, according to the present invention, since the generation to deployment of processing logic can be automated in a series, the dependency of the implementation and operation of routine processing can be reduced.

[0017] In addition, according to the present invention, by combining monitoring and audit log generation, it becomes easy to record operations on external services and visualize the execution status of routine processing, improving operability and auditability.

[0018] In addition, according to the present invention, by adopting a configuration that selects an execution method by comparing the inference cost and the serverless execution cost, and a configuration that individually deploys processing logic for each customer or user, cost optimization and multi-tenant operation can be facilitated.

Brief Description of the Drawings

[0019] [Figure 1] It is a configuration diagram showing a configuration example of the present invention. [Figure 2] It is a block diagram showing a functional block configuration example of the present invention. [Figure 3] It is a flowchart showing an example of a processing procedure of the present invention. [Figure 4] It is a sequence diagram showing a time series example of the processing of the present invention. [Figure 5] It is a sequence diagram showing a time series example of the processing following FIG. 4.

Modes for Carrying Out the Invention

[0020] Hereinafter, embodiments of the present invention will be described with reference to the drawings. The following embodiments are examples for embodying the present invention, and the technical scope of the present invention is determined by the invention described in the claims.

[0021] As shown in FIG. 1, the information processing system 10 may include a user terminal 11, an artificial intelligence agent 20, a serverless execution environment 40, and an external service 50. The user terminal 11 is a terminal for a user to send natural language input and receive a response or a result.

[0022] The artificial intelligence agent 20 is an execution entity that analyzes natural language input from the user and performs business processing, including tasks that involve operations on external services 50. External services 50 are, for example, SaaS, business systems, storage services, messaging services, etc., and can be operated by API calls.

[0023] The serverless execution environment 40 is an execution platform that executes processing logic in response to triggers such as events or schedules. The serverless execution environment 40 may include at least one of a function execution service 41 and a workflow execution service 42.

[0024] The information processing system 10 is characterized by collecting the history of business processes performed by the artificial intelligence agent 20, detecting recurring processing patterns, identifying routine processes that can be definitively executed without inference processing, and automatically migrating those routine processes to the serverless execution environment 40.

[0025] As shown in Figure 2, the information processing system 10 may include, in addition to the artificial intelligence agent 20, a history storage unit 23, a pattern detection unit 24, a standardization determination unit 25, a generation unit 26, a test unit 27, a deployment unit 28, a trigger management unit 29, an exception escalation unit 30, a monitoring unit 32, and an audit unit 33. The information processing system 10 may also include an execution method selection unit 31.

[0026] The artificial intelligence agent 20 may include a natural language analysis unit 21 and a task execution unit 22. The natural language analysis unit 21 analyzes natural language input received from the user terminal 11 and generates candidate tasks or procedures to be executed. The task execution unit 22 executes tasks by calling APIs of external services 50, etc., and obtains the results.

[0027] The history storage unit 23 records the execution history of business processes performed by the artificial intelligence agent 20. The execution history may include, for example, API calls to external services 50, input parameters, execution results, and execution times.

[0028] The pattern detection unit 24 detects repetitive processing patterns based on the execution history recorded in the history storage unit 23. The processing pattern can be represented, for example, as a combination of API call sequences, data transformation sequences, and notification sequences.

[0029] The standardization determination unit 25 identifies standard processing patterns that can be definitively executed without using inference processing from among the processing patterns detected by the pattern detection unit 24. Standardization determination may be performed based on, for example, the smallness of input fluctuations, the definability of branching conditions, and the degree of uncertainty of external dependencies.

[0030] The generation unit 26 automatically generates the standardized processing identified by the standardization determination unit 25 as processing logic that can be executed in the serverless execution environment 40. The processing logic can be generated as function code, workflow definitions, or a combination thereof.

[0031] The test unit 27 can automatically generate test cases for the processing logic generated by the generation unit 26. The test unit 27 can also execute these test cases in the serverless execution environment 40 or a corresponding verification environment and generate execution results.

[0032] The deployment unit 28 automatically deploys the generated processing logic to the serverless execution environment 40. The deployment unit 28 may adopt a configuration in which it deploys the processing logic to the production environment if the test results from the test unit 27 meet predetermined conditions.

[0033] The trigger management unit 29 manages triggers that initiate routine processing. A trigger may include at least one of the following: a schedule, an event notification, or a webhook.

[0034] The exception escalation unit 30, when it detects an exception while executing routine processing in the serverless execution environment 40, notifies the artificial intelligence agent 20 of the exception information and requests action. The exception may be, for example, an error response from an external service 50, throttling, a timeout, or a data format inconsistency.

[0035] The monitoring unit 32 monitors the execution status of routine processes. The auditing unit 33 generates audit logs regarding the deployment of processing logic or operations on external services 50. Monitoring and auditing may include, for example, the success rate of routine processes, reasons for failure, delays, and access history to external services 50.

[0036] The execution method selection unit 31 may select whether to execute the inference process or the routine process in the serverless execution environment 40, based on the execution cost of the inference process and the execution cost of executing the routine process in the serverless execution environment 40. Here, the execution cost can be defined, for example, as the amount of computing resources or charges required for the inference process and the amount of computing resources or charges required for execution in the serverless execution environment 40.

[0037] Next, an example of the processing procedure of the present invention will be described with reference to Figure 3. First, the history storage unit 23 stores the execution history of business processes performed by the artificial intelligence agent 20 (step S101).

[0038] Next, the pattern detection unit 24 detects a repetitive processing pattern based on the execution history (step S102). Subsequently, the standardization determination unit 25 identifies a standard processing that can be definitively executed from the processing pattern without using inference processing (step S103).

[0039] Next, the generation unit 26 automatically generates the standard processing as processing logic that can be executed in the serverless execution environment 40 (step S104). The test unit 27 then automatically generates test cases for the processing logic and executes these test cases (step S105).

[0040] If the test results meet the predetermined conditions, the deployment unit 28 automatically deploys the processing logic to the serverless execution environment 40 (step S106). Next, the trigger management unit 29 activates the deployed processing logic and executes the routine processing based on a schedule, event notification, or trigger such as a webhook (step S107).

[0041] During or after the execution of routine processing, the monitoring unit 32 monitors the execution status, and the audit unit 33 generates an audit log (step S108). If an exception is detected during the execution of routine processing, the exception escalation unit 30 notifies the artificial intelligence agent 20 of the information regarding the exception and requests action (step S109).

[0042] The artificial intelligence agent 20 may generate a correction procedure or proposed modification based on the notified exception information (step S110). If the artificial intelligence agent 20 decides to update the processing logic, the generation unit 26 may generate the updated processing logic, and the deployment unit 28 may redeploy it.

[0043] Next, a time-series example of the present invention will be described with reference to Figure 4. The artificial intelligence agent 20 operates an external service 50 in response to natural language input from the user terminal 11 and executes business processing. The result of this execution is recorded in the history storage unit 23 (step S101).

[0044] If predetermined conditions are met, the pattern detection unit 24 analyzes the execution history and extracts a repetitive processing pattern (step S102), and the standardization determination unit 25 identifies a process that can be standardized (step S103). The generation unit 26 generates processing logic corresponding to the identified standardized process (step S104), and the test unit 27 executes the test (step S105).

[0045] The deployment unit 28 deploys the processing logic that has passed the test to the serverless execution environment 40 (step S106). The trigger management unit 29 activates the processing logic when a trigger condition is met and causes the serverless execution environment 40 to execute the routine processing (step S107).

[0046] The serverless execution environment 40 performs operations on the external service 50 and obtains the results. The monitoring unit 32 and the auditing unit 33 record the execution status of routine processing and the history of operations on the external service 50 (step S108).

[0047] If an exception occurs, the exception escalation unit 30 notifies the artificial intelligence agent 20 of the exception information (step S109), and the artificial intelligence agent 20 selects and executes at least one of the following: execute a correction procedure, generate a corrective solution, or notify a human user (step S110).

[0048] There are various ways to collect execution history. For example, a configuration can be adopted in which the endpoint identifier, input parameters, response code, summary of the response content, and execution time of API calls issued by the artificial intelligence agent 20 to the external service 50 are recorded sequentially or in batches.

[0049] The pattern detection unit 24 can represent the execution history as a time-series event sequence and determine repetition based on whether identical or similar event sequences appear a predetermined number of times or more. It can also calculate the similarity of the event sequences using a predetermined similarity index and cluster those exceeding a threshold as identical patterns.

[0050] The standardization determination unit 25 can determine whether or not it is executable without inference processing, based on the definability of the branching conditions included in the processing pattern and the determinability of the input data. For example, a configuration can be adopted in which it is determined that standardization is possible if the branching conditions can be expressed as a predetermined rule expression or filter condition.

[0051] The generation unit 26 can adopt a configuration as its processing logic that generates function code for the function execution service 41, or a configuration that generates a workflow definition for the workflow execution service 42, or a combination of these. The workflow definition may include multiple function calls, retries, error handling, and branching.

[0052] The test unit 27 can perform tests using mocks or stubs of the external service 50, tests by making actual calls in a verification environment, or a combination of both. Test cases can be generated to include representative input values, boundary conditions, and exception cases extracted from the execution history.

[0053] The deployment unit 28 can adopt a configuration that adds version information when deploying processing logic, enabling rollback from existing versions. It can also adopt a configuration that gradually promotes the deployment target from the verification environment to the production environment.

[0054] In addition to being configured to start according to a schedule, the trigger management unit 29 can also be configured to start when it receives an event notification from an external service 50 or when it receives a webhook provided by the external service 50.

[0055] The exception escalation unit 30 can generate notification information, including the type of exception, location of occurrence, input value, response from the external service 50, and retry history, as exception information, and send it to the artificial intelligence agent 20. Based on the notification information, the artificial intelligence agent 20 can generate changes to the retry conditions, parameter corrections, or confirmation requests to the human user.

[0056] The monitoring unit 32 can be configured to collect and visualize the success rate, delay, number of retries, and exception rate as the execution status of routine processing. The audit unit 33 can be configured to record audit logs related to operations on the external service 50, associating them with the user or customer identifier, version information of the processing logic, and execution time.

[0057] The execution method selection unit 31 compares the inference resource consumption or inference charge when executing the same task using inference processing with the execution resource consumption or execution charge when executing it in the serverless execution environment 40, and can select an execution method based on predetermined criteria. For example, a configuration can be adopted in which only the parts requiring inference processing are processed by the artificial intelligence agent 20, and the remaining parts are offloaded to the serverless execution environment 40.

[0058] When individual deployments are performed on a customer or user basis, the deployment unit 28 can configure processing logic based on external service connection settings specific to the customer or user, and can adopt a configuration in which the processing logic is deployed individually on a customer or user basis. This makes it possible to optimize and isolate each customer or user even in a multi-tenant environment.

[0059] Next, an embodiment will be described. For example, suppose an artificial intelligence agent 20 repeatedly performs a business process that acquires status information from an external service 50 at predetermined intervals and sends a notification when predetermined conditions are met. In this case, the history storage unit 23 repeatedly records the API call sequence corresponding to the acquisition of the status information and the notification process.

[0060] The pattern detection unit 24 detects the repetitive pattern, and the standardization determination unit 25 identifies it as a standard process based on whether the input can be defined by predetermined parameters and whether the branching conditions can be expressed by rule expressions. The generation unit 26 generates the processing logic corresponding to the standard process, and the test unit 27 executes the test in the verification environment.

[0061] The deployment unit 28 deploys the processing logic that has passed the test to the serverless execution environment 40, and the trigger management unit 29 sets the scheduled triggers. As a result, execution at predetermined intervals is performed in the serverless execution environment 40, without the intervention of the inference processing of the artificial intelligence agent 20.

[0062] If an exception occurs due to a change in the response format of the external service 50, the exception escalation unit 30 notifies the artificial intelligence agent 20, and the artificial intelligence agent 20 generates, for example, a correction of the input parameters or a proposed modification of the processing logic, and the updated processing logic is redeployed, thereby restoring subsequent execution.

[0063] <Example 1> The serverless execution environment 40 is not limited to the cloud, but may include event-driven execution infrastructure in on-premises or edge environments. Furthermore, the processing logic is not limited to function code, but can be expressed as workflow definitions, job definitions, or IPaaS flow definitions.

[0064] This configuration allows the same automated migration framework to be applied to different execution platforms depending on the constraints of the operating environment.

[0065] <Modification 2> The pattern detection unit 24 can be configured to detect processing patterns not only by matching API call sequences, but also by using feature quantities such as the type of external service 50, the purpose label of the call, and the data conversion type. Furthermore, the standardization determination unit 25 can be configured to prioritize processing with an exception rate below a predetermined value as a candidate for standardization.

[0066] This configuration allows for automated migration while increasing the accuracy of repetition detection and the success rate of standardization.

[0067] <Variation 3> The test unit 27 and the deployment unit 28 can adopt a configuration in which, in addition to test results, regression tests based on past failure patterns collected by the monitoring unit 32 are added before deployment to production. Furthermore, during the initial period after deployment, a configuration can be adopted in which the trigger management unit 29 is controlled to activate at a low frequency or for limited targets.

[0068] This configuration reduces the operational risks associated with migrating routine processes.

[0069] <Modification 4> The exception escalation unit 30 can be configured to notify the human user of a confirmation request in addition to notifying the artificial intelligence agent 20, depending on the type of exception. Furthermore, if the exception is minor and automatic retries are effective, the system can be configured to perform a certain number of automatic retries first, and then escalate the exception.

[0070] This configuration allows for an expanded scope of automated exception handling while still enabling human or artificial intelligence agent intervention only when necessary.

[0071] <Modification 5> The execution method selection unit 31 can decompose the same task into a standard part to be executed in the serverless execution environment 40 and an inference part to be executed by the artificial intelligence agent 20, and the processing logic can adopt a configuration in which the inference part is executed as an external call. The inference part can be extracted as, for example, a part that requires summarization, natural language generation, or ambiguous judgment.

[0072] This configuration allows for both high-speed and stable execution of the standardized parts and flexibility in the inference part, while suppressing overall cost and delay. [Explanation of symbols]

[0073] 10. Information Processing Systems 11 User terminals 20 Artificial Intelligence Agents 21 Natural Language Analysis Department 22. Operations Execution Department 23 History Storage Unit 24 Pattern detection unit 25 Stylized judgment section 26 Generation part 27 Test Department 28 Deployment Department 29 Trigger Management Department 30 Exception Escalation Department 31 Execution Method Selection Section 32 Monitoring Department 33 Audit Department 40 Serverless Execution Environments 41 Function Execution Service 42 Workflow Execution Service 50 External Services

Claims

1. An AI agent that analyzes the user's natural language input and performs business processes, including operations on external services, Serverless execution environment and A pattern detection unit detects a repetitive processing pattern expressed as a combination of API call columns, data conversion columns, and notification columns, based on the execution history of business processing performed by the AI ​​agent, which includes API calls to the external service, input parameters, execution results, and execution time. A standardization determination unit evaluates the degree of input fluctuation, the definability of branching conditions, and the degree of uncertainty of external dependencies for the aforementioned processing pattern, and based on the evaluation results, identifies a standardized process that can be definitively executed without using the inference of the AI ​​agent. A generation unit that automatically generates the aforementioned routine processing as processing logic that can be executed in the serverless execution environment, A deployment unit that automatically deploys the generated processing logic to the serverless execution environment and executes the routine processing without the AI ​​agent's inference based on a predetermined trigger, An exception escalation unit that, when an exception is detected during the execution of the aforementioned routine processing, notifies the AI ​​agent of information regarding the exception and requests action to be taken, An information processing system characterized by comprising the following features.

2. In the information processing system described in claim 1, An information processing system characterized in that the pattern detection unit represents the execution history as a time-series sequence of events and determines the repetition based on whether the same or similar sequence of events appears a predetermined number of times or more.

3. In the information processing system described in claim 1, The system further includes a test unit that automatically generates test cases for the processing logic and executes the test cases in the serverless execution environment or a corresponding verification environment. An information processing system characterized in that the deployment unit deploys the processing logic to the production environment when the execution result of the test case satisfies predetermined conditions.

4. In the information processing system described in claim 1, An information processing system characterized in that the trigger includes at least one of a schedule, an event notification, or a webhook.

5. In the information processing system described in claim 1, An information processing system characterized in that the exception escalation unit selects and executes at least one of the following, depending on the type of exception: (i) generation of a correction procedure by the AI ​​agent, (ii) generation of a proposed modification for the standard processing, or (iii) notification to a human user.

6. In the information processing system described in claim 1, An information processing system characterized in that, when the AI ​​agent decides to update the processing logic based on the result of handling the exception, the generation unit generates the updated processing logic, and the deployment unit redeploys the updated processing logic.

7. In the information processing system described in claim 1, A monitoring unit that monitors the execution status of the aforementioned routine process, An audit unit that generates audit logs relating to the deployment of the aforementioned processing logic or operations on external services, An information processing system characterized by further comprising the following features.

8. In the information processing system described in claim 1, An information processing system characterized in that the serverless execution environment includes at least one of a function execution service and a workflow execution service.

9. In the information processing system described in claim 1, Based on the execution cost of the AI ​​agent's inference and the execution cost of performing the routine processing in the serverless execution environment, An information processing system further comprising an execution method selection unit that selects whether to perform the aforementioned routine processing by the inference of the AI ​​agent or by the serverless execution environment.

10. In the information processing system described in claim 1, An information processing system characterized by deploying the processing logic individually on a customer or user basis and executing the routine processing using external service connection settings specific to the customer or user.

11. An information processing system comprising an AI agent that analyzes a user's natural language input and performs business processing including operations on external services, and a serverless execution environment, wherein an information processing method is performed by one or more computers, The process of one or more computers detecting a repeatable processing pattern expressed as a combination of API call columns, data conversion columns, and notification columns based on the execution history of business processing by the AI ​​agent, which includes API calls to the external service, input parameters, execution results, and execution time, The process involves one or more computers evaluating the degree of input fluctuation, the definability of branching conditions, and the degree of uncertainty of external dependencies for the processing pattern, and identifying a routine process that can be definitively executed without using the inference of the AI ​​agent, based on the evaluation results. The process includes one or more computers automatically generating the routine processing as processing logic that can be executed in the serverless execution environment, The process involves one or more computers automatically deploying the generated processing logic to the serverless execution environment and executing the routine processing without the AI ​​agent's inference via a predetermined trigger. The process includes one or more of the above-mentioned computers notifying the AI ​​agent of information regarding the exception and requesting action when an exception is detected during the execution of the routine processing, An information processing method characterized by including

12. A program for causing a computer to execute the information processing method described in claim 11.

Images