Skill pool management and scheduling system for AI digital employees

Abstract

The invention discloses a skill pool management and scheduling system for AI digital employees, and relates to the field of AI digital employees. Comprising a task historical data reverse serialization module, a reverse skill gap propagation module, a skill gap active identification module and a skill self-generation module, and the system performs reverse modeling on task historical records, establishes a skill gap scoring mechanism, dynamically judges gap skills and combines and generates new skills from existing skills. The problem that an AI digital employee system cannot actively identify and cope with a novel task skill gap is solved.

Description

Technical Field

[0001] The present invention relates to AI digital employees, and more particularly, to a skill pool management and scheduling system for AI digital employees. Background Art

[0002] With the rapid development of artificial intelligence (AI) technology, more and more companies are deploying AI-powered digital workforce systems to automate business processes, including routine office tasks, customer service, data analysis, and approval processes. The core capabilities of these systems typically rely on predefined skill sets, leveraging a skill pool to respond to and execute external task demands.

[0003] However, most current AI skill pool building and scheduling solutions have significant limitations. First, systems often rely on manual definitions or task templates to develop skills. This model lacks the ability to perceive future task trends, causing skill updates to lag behind business development. Second, existing skill scheduling methods are mostly based on keyword matching, fixed rules, or static tags. They lack the ability to proactively identify skill gaps and are difficult to adapt to the rapid changes and heterogeneous growth of task scenarios.

[0004] In practice, enterprise tasks are often highly frequent and fragmented, and subtle evolutionary relationships may exist between different tasks. For example, a task originally focused on "loan approval" may evolve into a new type of task, "green credit," driven by green policies. If the system fails to identify insufficient skill coverage early in its evolution, the enterprise will face reduced processing efficiency or reliance on manual intervention.

[0005] Therefore, how to enable AI digital employee systems to identify skill gaps and automatically generate new skills has become an important technical challenge in the current design of intelligent office systems. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a skill pool management and scheduling system for AI digital employees to solve the problems mentioned in the background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A skill pool management and scheduling system for AI digital employees, including:

[0009] The task history data reverse serialization module is used to arrange the enterprise's historical task execution data in reverse order of task completion time to form a task reverse sequence;

[0010] A reverse skill gap propagation module is used to build a reverse skill propagation network, determine the skill gap required by the task according to the reverse sequence of the task, and calculate the potential gap value of the skill node;

[0011] A skill gap active identification module is used to actively identify potential gap values of the skill nodes through threshold comparison and determine a skill gap set;

[0012] The skill self-generation module is used to generate an initial version of new skills through transfer learning based on the skill gap set and incorporate them into the existing skill pool.

[0013] In an optional embodiment, the task history data reverse serialization module adopts a Transformer self-attention model to capture the potential relevance of tasks from the task reverse sequence to form a task reverse sequence feature matrix.

[0014] In an optional embodiment, the skill reverse propagation network uses existing skills as network nodes, skill semantic feature vectors as node features, and skill gap contribution degrees as edge weights between nodes.

[0015] In an optional embodiment, the potential gap value calculation formula is as follows:

[0016]

[0017] In the formula:

[0018] G(S i ) represents the skill node S i The potential gap value, 0<γ<1 is the attenuation factor, P(S i ,j) represents the skill node S i The accumulated gap contribution value of the jth back propagation in the task sequence, N is the number of propagation rounds.

[0019] In an optional embodiment, the skill gap active identification module is further configured to calculate the mean and standard deviation of the potential gap values of the skill nodes and determine the threshold value according to the following formula:

[0020] θ=μ+α·σ;

[0021] Where θ represents the threshold for active gap identification, μ represents the mean of the potential gap value of the skill node, σ represents the standard deviation of the potential gap value of the skill node, and α is the adaptive coefficient dynamically adjusted by the deep reinforcement learning model.

[0022] In an optional embodiment, the deep reinforcement learning model dynamically adjusts the adaptive coefficient based on feedback results of skill gap prediction accuracy and enterprise task processing efficiency.

[0023] In an optional embodiment, the skill self-generation module generates an initial version of the new skill by searching the existing skill in the skill pool that is most similar to the gap skill and extracting the task execution sub-module combination of the skill using the knowledge transfer network.

[0024] In an optional embodiment, the knowledge transfer network adopts a reinforcement feedback mechanism to continuously optimize the sub-module combination according to the performance of the generated skills in the actual application of the task.

[0025] In an optional embodiment, the skill semantic feature vector is composed of skill definition, skill usage and skill attribution label.

[0026] In an optional embodiment, the task history data includes task instruction information, skill call records and task execution result information.

[0027] This invention's advantage over existing technologies lies in its ability to proactively identify and assess potential skills gaps through the AI digital workforce system, building a reverse-sequence analysis mechanism based on historical task data. This addresses the inability of traditional skill pools to foresee business evolution trends. After a task is completed, the system no longer merely archives historical records; instead, it identifies overlooked structural requirements in reverse chronological order, enhancing the system's adaptability to future task scenarios.

[0028] Based on the constructed skill reverse propagation network, the system can model the functional structure of existing skills and conduct propagation analysis, forming a precise skill gap scoring mechanism, thereby achieving data-driven skill priority determination. Combining statistical models with reinforcement learning adjustment mechanisms, the system can dynamically establish skill gap determination thresholds, effectively avoiding the risks of over-generation and omission, and achieving high-accuracy gap identification.

[0029] In terms of skill generation, the system extracts key functional modules from existing skills through a knowledge transfer network, recombining and reconstructing them based on interface compatibility and logical coherence, ultimately forming new skills with rational structure and reusable functions. During actual operation, the newly generated skills are continuously optimized through a reinforced feedback mechanism, resulting in continuous performance improvements, high scalability, and stable task execution.

[0030] This invention not only improves the automation and adaptability of AI systems, but also significantly reduces manual maintenance costs and system response delays, and has broad application prospects for future-oriented multi-scenario intelligent task platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is an overall flow chart of the system of the present invention;

[0032] Figure 2 It is a flow chart of the reverse serialization processing of the system task history data of the present invention;

[0033] Figure 3 It is a flow chart of the reverse skills gap propagation analysis of the system of the present invention. DETAILED DESCRIPTION

[0034] The specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0035] like Figure 1 As shown, the overall process of the system of the present invention includes the following modules:

[0036] The task history data reverse serialization module is used to arrange the enterprise's historical task execution data in reverse order of task completion time to form a task reverse sequence;

[0037] A reverse skill gap propagation module is used to build a reverse skill propagation network, determine the skill gap required by the task according to the reverse sequence of the task, and calculate the potential gap value of the skill node;

[0038] A skill gap active identification module is used to actively identify potential gap values of the skill nodes through threshold comparison and determine a skill gap set;

[0039] The skill self-generation module is used to generate an initial version of new skills through transfer learning based on the skill gap set and incorporate them into the existing skill pool.

[0040] More specifically, Figure 2 As shown, the present invention first performs reverse serialization processing of task history data. Specifically, the execution records of various tasks processed by the AI digital employee system in the enterprise are arranged in reverse order according to the chronological order of each task completion to obtain a reverse time series of the tasks. In this process, the system extracts the task instruction text, the called skill identifier, and the corresponding execution result evaluation information of each task, encapsulates them into a unified data structure, and constructs a task reverse sequence set. The purpose is to enhance the system's sensitivity to "upcoming but not yet available" skill requirements and identify potential skill gaps from the development path of historical tasks.

[0041] like Figure 3 As shown, reverse skill gap propagation analysis is then performed. The system uses existing skills in the current skill pool as nodes to construct a skill graph network. The calling relationships between these skills and the evolutionary trends between tasks are modeled as a graph structure. Edges between skills represent the probability of them being jointly called or derived from each other in historical tasks. Each skill node is represented by its semantic feature vector, which combines multiple dimensions such as the skill's definition text, the task type, and the function label. This vector can be obtained using a pre-trained language model or a feature fusion network.

[0042] In this skill network, the system traces back through each task in the reverse sequence, analyzing whether there is a semantic functional gap between the skill combination actually called upon and the task objective. If so, the task is considered to have a "gap requirement" for uncalled-for but potentially relevant skills. Each time a skill gap is identified, the system propagates a "gap signal" through the graph network, starting with the skills called upon by the current task. During this propagation process, skill nodes accumulate their "potential gap value," reflecting the extent to which they are required by future tasks but currently unmet.

[0043] Specifically, the potential gap value calculation formula is as follows:

[0044]

[0045] In the formula:

[0046] G(S i ) represents the skill node S i The potential gap value, 0<γ<1 is the attenuation factor, P(S i ,j) represents the skill node S i The accumulated gap contribution value of the jth back propagation in the task sequence, N is the number of propagation rounds.

[0047] After the gap value calculation is completed, the active skill gap identification phase begins. The system performs statistical processing on the values of all skill nodes and calculates the mean and standard deviation of the skill gap value. Based on this, a screening criterion based on the distribution threshold is constructed:

[0048] θ=μ+α·σ;

[0049] In the formula, θ represents the threshold for proactive gap identification, μ represents the mean of the potential gap values for skill nodes, σ represents the standard deviation of the potential gap values for skill nodes, and α is an adaptive coefficient dynamically adjusted by the deep reinforcement learning model. This coefficient is continuously optimized by the reinforcement learning agent based on feedback metrics such as historical prediction accuracy and task completion efficiency. This ensures that the threshold remains adaptable to varying business pressures and skill distributions. All skill nodes are included in the skill gap set and marked as priority target skills for generation.

[0050] Subsequently, the system starts the skill self-generation process. For each gap skill, the system retrieves the existing skill with the closest semantic vector from the skill pool as the source skill for transfer learning. For example, if the target gap skill is "green credit approval", and there are "general loan approval" and "environmental policy analysis" in the skill pool, then these two skills will be identified as highly similar source skills. The system uses a knowledge transfer network to extract and combine the task execution modules of the source skills (such as data preprocessing, approval scoring model, document generation module, etc.) at the functional level, and splice them into a new skill structure. The entire skill combination process follows the task logic consistency and interface compatibility rules to ensure that the modules work together.

[0051] The knowledge transfer network also includes a reinforced feedback mechanism. After a new skill is deployed, the system adjusts the weights and combinations of each module based on its performance in actual task execution, such as average processing time, user evaluation scores, and error rates, thereby iteratively optimizing the skill structure. The system also supports synchronously writing these optimization results back to the skill pool for future use in similar tasks or secondary generation.

[0052] As the fundamental information source for the entire process, task history data should at least include task instruction text (i.e., goals described in natural language), skill call records (i.e., a list of skills actually scheduled and their execution sequences), and task execution results (including quantitative and qualitative indicators such as success, time consumption, and degree of human intervention). The system will standardize and model this data for subsequent processing.

[0053] The present invention can be more clearly illustrated by the following Example 1:

[0054] A financial institution has deployed an AI-powered digital employee system to automate workflows for corporate loans, government subsidy applications, and risk assessments. Over the past six months, the system has automatically processed tens of thousands of tasks. However, with the recent launch of the national "XXXXXX Plan," customer needs have begun to shift, but the institution has yet to establish a corresponding "green credit approval" skills module.

[0055] In this context, the system first performs reverse serialization of task history data. Specifically, the system retrieves all task execution records from the past 180 days, arranges them in reverse order of completion time, and processes the most recently completed tasks in order (because the most recent tasks are closest to future developments). For example, the three most recent tasks are:

[0056] 1. Task number T_18000, task instruction: "Approve a green technology upgrade project loan for a high-energy-consuming enterprise." Skills: "Loan Approval (S_01)" and "Project Legality Verification (S_05)." Task execution result: "Taking a long time, the system interrupted the call multiple times, and manual intervention was required."

[0057] 2. Task number T_17999, task instruction is "Approve financing for the purchase of environmentally friendly equipment", and the skills called are: "Loan Approval (S_01)" and "Credit Rating (S_03)". The task execution result is "Processing successful but the approval logic is marked as 'mismatch'".

[0058] 3. Task number T_17998, task instruction: "Confirm the guarantee qualifications of new energy enterprises and evaluate the loan amount." Skills called: "Guarantee Verification (S_04)" and "Risk Assessment (S_02)." Task result: "Manually added environmental scoring module."

[0059] The system arranges all tasks in reverse chronological order, extracts the task text, skill call, and execution results of each task, and forms a unified data structure. For example:

[0060]

[0061] The entire historical task is encapsulated into a reverse time series for subsequent analysis.

[0062] In the second step, the system constructs a skill network with skills as nodes. For example, if skill S_01 "Loan Approval" has co-occurred with S_05 "Project Legality Verification," S_03 "Credit Rating," and S_02 "Risk Assessment," edges are created between them to indicate their connection within the task.

[0063] Each skill node will be labeled with a semantic vector. For example, "loan approval" contains keywords such as "financing", "risk control", and "process classification"; "environmental assessment" includes dimensions such as "emission verification", "policy adaptability", and "energy saving".

[0064] The system traced back from the most recent task in the reverse sequence and found that while several tasks ostensibly invoked skills like S_01, they all showed signs of artificially supplemented "environment-related" content. These traces indicated that the existing skill set did not cover all task logic.

[0065] Therefore, the system determines that there is a potential skill gap in these tasks that has not yet been explicitly defined, and preliminarily classifies them as "green credit approval (S_new)".

[0066] For each such "gap identification", the system will propagate a "gap signal" to the skill graph starting from the called skills of the current task, and the gap value of the corresponding skill will be accumulated.

[0067] For example:

[0068] In the T_18000 task, the skill "Loan Approval (S_01)" is marked with the Gap Propagation tag.

[0069] Because of its association with "Credit Rating (S_03)", S_03 also received a gap signal.

[0070] With each additional round of transmission, the signal attenuates, and the attenuation factor is set to γ = 0.85, and a total of 5 rounds of transmission are performed.

[0071] The system accumulates the values of all nodes and derives the potential gap value for each skill.

[0072] Next, let's assume that "Green Credit Approval" isn't in the current skill pool, but its potential gap is automatically assessed by the system based on upstream and downstream relationships. The potential gaps for all skill nodes are statistically calculated to have a mean of μ = 0.37 and a standard deviation of σ = 0.12. According to the formula:

[0073] θ=μ+α·σ;

[0074] Assuming that the current system setting is α=1.5, then:

[0075] θ = 0.37 + 1.5 · 0.12 = 0.55;

[0076] The system found that the potential gap value of "Green Credit Approval" = 0.62, which is higher than the threshold value of 0.55. Therefore, it is judged that this skill has constituted a significant gap and the supplementary process should be initiated.

[0077] In the fourth step, the system searches the skill pool for existing skills with similar semantics to "green credit approval." The search found:

[0078] "Loan Approval (S_01)" has a scoring and quota control module;

[0079] “Environmental Policy Analysis (S_06)” includes policy judgment and green classification modules;

[0080] "Interpretation of Corporate Carbon Emission Reports (S_08)" includes a module for recognizing environmental-related text.

[0081] The system extracts reusable modules from these three skills and combines them into a preliminary structure. The structure is as follows:

[0082] S_new (Green Credit Approval) = [Loan Scoring Module] + [Green Classification Judgment] + [Emission Compliance Analysis];

[0083] The assembly process is verified through interface standards and input and output consistency to ensure consistent operation. The generated skill is automatically registered in the skill pool with the number S_99.

[0084] In the fifth step, S_99 is used by the system for the next related task: "Approving green loans for new energy vehicle battery factories."

[0085] The results showed that the skill automatically completed most of the tasks, but it misinterpreted a section of a government green standards document. The system recorded this performance, adjusted the matching weight of the "Policy Analysis Submodule," and replaced that section with another text module.

[0086] After several rounds of tasks, the skill's execution success rate increased from 68% to 92% and was marked as a "stable skill."

[0087] It can be seen that when enterprises have not yet realized that "green credit approval" is an independent demand, the system built and launched this skill two weeks in advance through reverse time series, graph network propagation, gap value identification, module combination and enhanced feedback process, saving enterprises manual analysis time and improving customer response efficiency.

[0088] The following Example 2 provides a code example for implementing the present invention:

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100] The reverse_serialize_tasks function sorts the task list in reverse order by timestamp. The sample data includes the task number, instruction, called skill, execution result, and timestamp information.

[0101] In the build_skill_network function, skills are extracted from each task, skill nodes are established, and edges are established between skills that are commonly called in the same task. At the same time, a random semantic vector is initialized for each skill node (simulating the vector generated by the pre-trained language model in practice).

[0102] The function propagate_gap_signal takes the skill called in the task as the starting node, propagates the attenuated gap signal (γ=0.85, propagation for 5 rounds) to its neighboring nodes in rounds, and accumulates it in the "gap_value" of each skill node.

[0103] Use compute_threshold to calculate the mean and standard deviation of the gap values of all skills and construct a threshold θ = μ + α·σ; use identify_gap_skills to identify skills with gap values exceeding the threshold as gap skills.

[0104] The simulation assumes that the skill node (S_new) corresponding to "Green Credit Approval" is not in the skill pool, and the gap value is set to 0.62. When the gap value exceeds the threshold, the system calls the find_similar_skills function to retrieve skills from the existing skill pool that are semantically similar to the target skill by calculating the cosine similarity of their semantic vectors. It then uses the generate_new_skill function to extract modules from these similar skills (such as loan scoring, green classification, and emissions compliance analysis), combine them to form a new skill, and register it in the skill pool (with the new number "S_99").

[0105] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A skill pool management and scheduling system for AI digital employees, characterized by: include: The task history data reverse serialization module is used to arrange the enterprise's historical task execution data in reverse order of task completion time to form a task reverse sequence; A reverse skill gap propagation module is used to build a reverse skill propagation network, determine the skill gap required by the task according to the reverse sequence of the task, and calculate the potential gap value of the skill node; A skill gap active identification module is used to actively identify potential gap values of the skill nodes through threshold comparison and determine a skill gap set; The skill self-generation module is used to generate an initial version of new skills through transfer learning based on the skill gap set and incorporate them into the existing skill pool.

2. The skill pool management and scheduling system for AI digital employees according to claim 1 is characterized in that: The task history data reverse serialization module adopts the Transformer self-attention model to capture the potential relevance of tasks from the task reverse sequence and form a task reverse sequence feature matrix.

3. The skill pool management and scheduling system for AI digital employees according to claim 1 is characterized in that: The skill reverse propagation network uses existing skills as network nodes, skill semantic feature vectors as node features, and skill gap contribution degrees as edge weights between nodes.

4. The skill pool management and scheduling system for AI digital employees according to claim 3 is characterized in that: The potential gap value calculation formula is as follows: In the formula: G(S i ) represents the skill node S i The potential gap value, 0<γ<1 is the attenuation factor, P(S i ,j) represents the skill node S i The accumulated gap contribution value of the jth back propagation in the task sequence, N is the number of propagation rounds.

5. The skill pool management and scheduling system for AI digital employees according to claim 1 is characterized in that: The active skill gap identification module is further configured to calculate the mean and standard deviation of the potential skill gap values and determine the threshold value according to the following formula: θ=μ+α·σ; Where θ represents the threshold for active gap identification, μ represents the mean of the potential gap value of the skill node, σ represents the standard deviation of the potential gap value of the skill node, and α is the adaptive coefficient dynamically adjusted by the deep reinforcement learning model.

6. The skill pool management and scheduling system for AI digital employees according to claim 5 is characterized in that: The deep reinforcement learning model dynamically adjusts the adaptive coefficient based on feedback results of skill gap prediction accuracy and enterprise task processing efficiency.

7. The skill pool management and scheduling system for AI digital employees according to claim 1 is characterized in that: The skill self-generation module retrieves the existing skills in the skill pool that are most similar to the gap skills, and uses the knowledge transfer network to extract the task execution sub-modules of the skills to generate an initial version of the new skills.

8. The skill pool management and scheduling system for AI digital employees according to claim 7 is characterized in that: The knowledge transfer network adopts a reinforcement feedback mechanism to continuously optimize the sub-module combination according to the performance of the generated skills in the actual application of the task.

9. The skill pool management and scheduling system for AI digital employees according to claim 3 is characterized in that: The skill semantic feature vector is composed of skill definition, skill usage and skill attribution label.

10. The skill pool management and scheduling system for AI digital employees according to claim 1, characterized in that: The task history data includes task instruction information, skill call records and task execution result information.

Images