A method and system for monitoring and multi-mode takeover decisions of escort-coordinated outbound calls.

By constructing a dialogue state space and an irreversible failure subgraph (IFS), combined with intent declaration frames and risk prediction, multi-mode takeover decision-making was achieved. This solved the problem that existing systems could not predict failures and misjudgments in high-complexity outbound call scenarios in the early stages, thereby improving outbound call conversion rates and customer satisfaction.

CN123123508APending Publication Date: 2026-11-17KEXUN JIALIAN INFORMATION TECH CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611632732.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing human-machine collaborative outbound calling systems struggle to identify call status evolution in highly complex scenarios, are unable to predict failures early and intervene accurately, are prone to misjudgment in takeover decisions, and have inconsistent information quality, leading to decreased call quality and poor customer experience.

Method used

By constructing a dialogue state space and an irreversible failure subgraph (IFS), the potential gradient of the call state migration to the failure region is calculated based on the potential function. Combined with intent declaration frames, risk prediction, and belief strength assessment, multi-mode takeover decision-making is achieved, including autonomous, advisory, negotiation, cooling-off, and emergency takeover modes. The decision-making level is dynamically adjusted and a dialogue state compensation strategy is generated.

Benefits of technology

It significantly improved outbound call conversion rates and customer satisfaction, avoided misjudgments caused by premature or late intervention, provided accurate basis for takeover decisions, ensured a smooth transition from fully automated to human-led systems, and enhanced the system's robustness under complex conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123508A_ABST
    Figure CN123123508A_ABST
Patent Text Reader

Abstract

This invention relates to the fields of artificial intelligence and call center technology, specifically to a method and system for monitoring and multi-mode takeover decisions in collaborative outbound calls. It adopts a core architecture of "AI-driven employee-led outbound calls, real-time agent-led monitoring, and multi-dimensional intelligent takeover decisions." By constructing a dialogue state space and an irreversible failure subgraph (IFS), it calculates the potential gradient of the current call state migrating to the failure region based on a potential function. Combining six sub-dimensions—dialogue path conflict, interaction link conflict, risk acceptance conflict, dialogue resource conflict, time conflict, and irreversible failure potential energy conflict—it weights and obtains the intent-risk prediction conflict degree, and executes a five-level hierarchical decision-making process: autonomous, suggestive, negotiated, cooling-off, and emergency takeover. The system supports multi-channel concurrent split-screen monitoring by a single agent, achieving seamless context transfer and belief strength weighting, effectively improving outbound call efficiency, customer experience, and conversion quality.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and call center technology, specifically to a method and system for monitoring and multi-mode takeover decision-making in collaborative outbound calls. Background Technology

[0002] With the continuous evolution of artificial intelligence technology, intelligent voice interaction has been widely applied in call centers and telephone customer service. The maturity of large language models, automatic speech recognition, and natural language processing technologies has enabled the large-scale deployment of automated outbound calling robots in scenarios such as standardized notifications, satisfaction surveys, and simple marketing, significantly improving outbound calling efficiency and reducing labor costs. However, in highly complex business scenarios such as financial debt collection, insurance follow-ups, and sales follow-ups, customer demands are highly diverse, emotional states are complex and changeable, and objections are difficult to handle. Purely automated robots often struggle to cope, easily falling into repetitive responses or making incorrect decisions, leading to decreased call quality, deteriorated customer experience, and even compliance risks related to complaints.

[0003] Against this backdrop, the human-machine collaborative "flying" customer service model has gradually become a focus of industry attention. This model aims to balance the high efficiency of AI's large-scale concurrent processing with the core value of human intervention in handling complex requests and building emotional connections. It achieves an optimal balance between efficiency and quality through AI-led outbound call execution, real-time human monitoring, and intelligent dynamic takeover. However, existing human-machine collaborative solutions still have significant shortcomings in areas such as dialogue state modeling, failure risk prediction, takeover decision-making mechanisms, and information credibility assessment, making it difficult to meet the needs of highly complex outbound call scenarios for precise intervention and robust decision-making.

[0004] Existing technologies have the following main drawbacks when handling highly complex outbound call scenarios: Existing systems largely rely on pre-defined dialogue flow trees and simple rules for dialogue management, failing to systematically model the evolution of the dialogue state and thus unable to identify "irreversible failure zones" that are difficult to recover from once a call enters a phase. Therefore, the system can only passively trigger takeover after a customer explicitly expresses refusal or has an emotional outburst, unable to predict or intervene in the early stages of the call's progression towards failure, thus missing the optimal intervention window.

[0005] Existing takeover strategies typically rely solely on keyword matching, sentiment recognition, or single risk indicators for triggering, failing to comprehensively consider multi-dimensional factors such as dialogue path deviation, call quality, risk acceptance bias, remaining dialogue resources, time constraints, and irreversible failure potential. This leads to misjudgments in takeover decisions: either premature intervention wastes manpower, or late intervention results in conversion failure.

[0006] Existing systems often employ a binary, sequential switching mechanism for transferring intervention to human intervention, lacking a tiered decision-making system that progresses from autonomous mode, suggestion mode, negotiation mode, cooling-off mode to emergency takeover mode. The system cannot dynamically select the intervention level based on the severity of the conflict, nor can it provide AI with an opportunity to adjust its intentions during the cooling-off period, resulting in abrupt takeovers and poor customer experience.

[0007] During outbound calls, the quality of information such as speech recognition confidence, intent recognition confidence, prediction model accuracy, and AI intent reliability varies. Existing systems typically treat all input information equally, without reducing the weight of low-confidence information, which may lead to distorted conflict assessment and incorrect takeover decisions.

[0008] Therefore, a method and system for monitoring and multi-mode takeover decisions of escort-coordinated outbound calls are designed. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for monitoring and multi-mode takeover decisions in escort-coordinated outbound calls, so as to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention aims to provide a method for monitoring and multi-mode takeover decision-making of flight companion collaborative outbound calls, including: S1, Intent declaration: the AI ​​outbound call terminal sends a structured intent declaration frame to the flight companion monitoring system at a preset period, the intent declaration frame including at least: outbound call target, dialogue planning path, communication strategy, risk acceptance level, confidence level, and a summary of decision reasons.

[0011] The intent declaration frame described in S1 includes: Outbound call targets: including target type, parameters, priority, and alternative targets; the target type includes at least one of notification, survey, collection, marketing, and follow-up calls; Dialogue planning path: including configuration of at least one dialogue node, estimated arrival time, speech progression rate, emotional escalation rate, and objection escalation rate; Communication strategies include: primary voice channel requirements, backup SMS or ticket channels, and data priority allocation; Risk tolerance: The acceptable threshold for risk categories such as customer hanging up, complaints, compliance, and emotional outbursts; Confidence level: A score indicating how well the AI-powered outbound caller understands the intent. Summary of Decision Justification: A text outlining the basis for the decision, not exceeding the preset word limit.

[0012] S2. Risk Prediction: The accompanying monitoring system performs forward risk prediction for the intent declaration frame and maps the current call state to a pre-constructed dialogue state space. Based on the potential function, it calculates the potential gradient of the current state to migrate to the attraction domain of the irreversible failure subgraph IFS. The irreversible failure subgraph is a set of states in the dialogue state space that will cause the call to fail once entered, regardless of how the subsequent dialogue is executed.

[0013] The potential function described in S2 includes: A Markov potential function is constructed based on the dialogue state transition probability. The value of the potential function increases as the distance between the state and the IFS boundary decreases, and the gradient of the potential function points to the IFS attraction domain.

[0014] The potential energy gradient calculation described in S2 includes: The current call state is encoded as a state vector, and the gradient of the potential function of the state vector is calculated in the dialogue state space. The gradient represents the speed and direction of the current state's migration to the IFS attraction domain.

[0015] The accompanying flight monitoring system described in S2 independently accesses outbound calls through signaling bypass and media bypass, does not participate in the generation of dialogue content, and is deployed decoupled from the AI ​​outbound call process.

[0016] S3. Conflict Measure: Calculate the intention-risk prediction conflict degree, which is a scalar in the interval [0,1] and is calculated by weighted combination of six sub-dimensions.

[0017] S4. Layered takeover decision: Automatically decides whether to enter autonomous mode, suggestion mode, negotiation mode, cooling-off mode or emergency takeover mode based on the degree of conflict.

[0018] S5. Belief Strength Weighting: Assign a belief strength label Belief∈[0,1] to each type of input information. The belief strength reflects the credibility of the information in the current context and is applied in conflict degree calculation and decision weight adjustment. When the belief strength of the information is lower than a preset threshold, the contribution weight of the information to the conflict degree is reduced accordingly.

[0019] S6. Feedback and Learning: Record the complete decision-making process for each frame, including the original input data, the calculation process of conflict degree for each sub-dimension, the final conflict degree and the takeover mode entered, the response of the AI ​​outbound call terminal or human agent to the suggestion, and the deviation between the actual execution result and the prediction; dynamically adjust the weight of each sub-dimension and / or the conflict degree threshold of each takeover mode based on the feedback data.

[0020] S7. Traceability Guarantee: Maintain an immutable decision audit log, with each record containing at least a timestamp, intent statement frame content, risk prediction results, conflict degree calculation results, takeover mode decision, execution actions, and response records.

[0021] S8. Compensation Strategy Generation: When it is decided to take over, a dialogue state compensation strategy is generated based on the reverse path analysis of IFS. The compensation strategy is used to repair the dialogue topology damage caused by AI dialogue in the first N rounds of dialogue after the human agent takes over.

[0022] As a further improvement to this technical solution, the intention declaration frame described in S2 performs forward risk prediction, which includes at least: path reachability assessment for each dialogue node on the dialogue planning path; prediction of call link quality; probability distribution estimation of customer hang-up risk, complaint risk, compliance risk, and emotional loss of control risk; and detection of time-event conflicts.

[0023] Path reachability assessment is used to determine whether there is a feasible path from the current dialogue node to subsequent dialogue nodes in the dialogue planning path declared by the AI ​​outbound caller. Specifically, the accompanying monitoring system extracts the node sequence of the dialogue planning path and the preconditions for each node from the intent declaration frame (such as customer identity verification, risk disclosure completion, and response to a specific objection). Combining this with the set of preconditions already met in the current call state, it assesses the reachability of each node. If a node's preconditions cannot be met in the current state, or if reaching the node along the planned path requires traversing a state marked as an irreversible failure subgraph (IFS), then the node is determined to be unreachable. The assessment results are output in three levels: reachable, conditionally reachable, and unreachable, and serve as input for calculating the risk acceptance conflict degree and the irreversible failure potential conflict degree.

[0024] Call link quality prediction is used to estimate whether the call link can continuously support the accuracy requirements of speech recognition and natural language processing as it progresses along the dialogue planning path. Specifically, the accompanying monitoring system acquires real-time link parameters of the current call, including speech recognition confidence, speech activity detection frame drop rate, network round-trip latency, jitter, and packet loss rate. Based on historical call data, a mapping relationship between link quality and speech recognition confidence is established, and forward prediction is performed on the link quality at each subsequent dialogue node along the dialogue planning path. When the predicted speech recognition confidence is lower than a preset threshold, or the packet loss rate or latency exceeds a preset upper limit, the dialogue node is determined to have link quality risk. The prediction results are output as a link quality score or risk level and used as input for calculating the interaction link conflict degree.

[0025] The probability distribution estimates of customer hang-up risk, complaint risk, compliance risk, and emotional outburst risk are used to quantify the likelihood of the current call migrating to different failure types. In practice, the accompanying monitoring system extracts the following features from real-time call data: negative emotion words, expressions of objection, complaint keywords, and sensitive compliance words in the speech-to-text; acoustic features such as speech rate, volume, pauses, and silence duration; dialogue rounds, duration of the current point of contact, and number of times the customer repeats inquiries; and risk tags for the customer or similar customers in historical calls.

[0026] The aforementioned features are input into a pre-trained risk prediction model, which can be implemented using gradient boosting trees, recurrent neural networks, or fine-tuning of large language models, etc., and outputs the probability values ​​and confidence intervals for each of the four types of risks. Hang-up risk focuses on predicting the probability that the customer will actively hang up in subsequent rounds; complaint risk focuses on predicting the probability that the customer will express a complaint intention or request to be transferred to a higher level; compliance risk focuses on predicting the probability that AI-generated dialogue may trigger regulatory prohibitions, overstepping of commitments, or information leakage; and emotional outburst risk focuses on predicting the probability that the customer's emotions will escalate from negative to intense confrontation. The probability distributions of the four types of risks together constitute the risk prediction result, serving as the risk acceptance conflict level. Conflict degree with irreversible failure potential energy The calculation input.

[0027] Time-event conflict detection is used to determine whether the planned execution time window of the current outbound call overlaps with or times out of a scheduled constraint event within the system. Specifically, the accompanying monitoring system extracts the expected completion time of the dialogue planning path from the intent declaration frame, obtains the current call duration from the real-time call status, and queries the pre-registered constraints in the system calendar. These constraints include at least: customer appointment time or customer-agreed follow-up time, agent scheduling and shift handover times, compliant no-call periods, outbound call deadlines, system maintenance windows, and single call duration limits.

[0028] The system represents each constraint as a point in time or a time interval and calculates the estimated end time window for the current call. Then, the system checks whether the estimated end time window overlaps with the time interval of each constraint, or whether the estimated completion time is later than the deadline of the constraint. If overlap or timeout occurs, the system records the conflict name, conflict type, overlap duration, remaining available time, and estimated timeout duration, and determines the conflict severity based on the difference between the remaining available time and the estimated completion time. The detection results are output as a list of conflict events and a severity score; if there are no constraints or none overlap, a no-conflict flag is output. This detection result serves as the time conflict level in S3. The calculation input is as follows. Those skilled in the art can implement the above-described time-event conflict detection using other equivalent methods, all of which fall within the scope of protection of this invention.

[0029] As a further improvement to this technical solution, the irreversible failure subgraph IFS described in S2 is constructed in the following way: filtering calls that ultimately failed and where the customer clearly expressed negative emotions from historical call records; performing reverse analysis on the dialogue state sequence of the filtered calls to identify the key state transitions that led to the failure; and determining the boundary of the irreversible failure subgraph through clustering and boundary detection, wherein the boundary is the set of states that cannot be recovered through subsequent dialogues after entering the subgraph.

[0030] As a further improvement to this technical solution, the intent-risk prediction conflict degree is calculated in the following way: the dialogue planning path declared by the AI ​​outbound call terminal is sequentially aligned with the safe dialogue path recommended by the flight monitoring system, and the dialogue path conflict degree is obtained by nonlinear compression and weighted fusion based on the node semantic distance deviation, speech progression rate deviation and emotion escalation rate deviation of the aligned node pair.

[0031] For each dialogue node on the dialogue planning path declared by the AI ​​outbound call terminal, the interaction link conflict degree is obtained through nonlinear compression, degradation compensation and weighted fusion based on the delay gap, packet loss rate and recognition confidence gap between the predicted delay, predicted packet loss rate and predicted speech recognition confidence and the declared link requirements.

[0032] The risk acceptance conflict degree is obtained by weighting the risk category according to the deviation between the predicted risk level of each risk category and the acceptable threshold declared by the AI ​​outbound call terminal.

[0033] Based on the number of dialogue rounds required to complete the outbound call target, the number of human agents, the knowledge base retrieval, and the remaining resources between concurrent resources and the currently available resources of the system, the dialogue resource conflict degree is obtained by weighting by resource type.

[0034] The time conflict degree is obtained by taking the maximum overlap ratio between the expected completion time of the dialogue planning path and the time of the scheduled constraints.

[0035] The current call state is mapped to the dialogue state space. The potential gradient of the current state to the attraction domain of the irreversible failure subgraph (IFS) is calculated based on the potential function. The gradient magnitude is normalized to obtain the irreversible failure potential conflict degree.

[0036] Weighted fusion yields the degree of conflict between intention and risk prediction.

[0037] As a further improvement to this technical solution, the autonomous mode is as follows: the flight monitoring system enters a pure observation state, records the intention and execution process of the AI ​​outbound call terminal, and does not actively intervene.

[0038] Recommended mode: The escort monitoring system pushes the conflict analysis report to the AI ​​outbound call terminal or human agent, who can choose to accept the suggestion or stick to the original intention.

[0039] Negotiation Mode: The escort monitoring system triggers an objection negotiation agreement, presents detailed evidence of risk prediction, sets a timeout, and waits for explicit confirmation from the AI ​​outbound call terminal or human agent whether to accept the risk prediction.

[0040] Cooling mode: The escort monitoring system stops executing the current intent but does not force takeover, allowing a cooling-off period to adjust the intent; if the conflict level decreases during the cooling-off period, it will be recalculated; if no adjustment is made by the end of the cooling-off period, it will enter emergency takeover mode.

[0041] The calculation involves re-executing risk prediction in S2 and conflict measurement in S3 based on the modified intent declaration frame from the AI ​​outbound caller during the cooling-off period, resulting in a new total conflict degree Γ and its six sub-dimensions. The object of the calculation is the total conflict degree Γ, the basis for the calculation is the modified intent declaration frame, and the purpose of the calculation is to determine whether the conflict has been reduced to an acceptable level.

[0042] Emergency Takeover Mode: The accompanying monitoring system immediately takes over the outbound call decision, routing the call to a human agent or executing a preset emergency script, overriding the current instructions of the AI ​​outbound call terminal.

[0043] As a further improvement to this technical solution, the belief strength mentioned in S5 specifically includes: belief strength assessment of the call data source: dynamic assessment based on speech recognition confidence, natural language processing intent recognition confidence, multi-source recognition consistency, background noise or dialect interference, and data timeliness.

[0044] The belief strength of the prediction results of the escort monitoring system is evaluated based on the prediction method type, the historical accuracy of the method, the prediction confidence interval width, and the consistency with cross-validation with other independent predictions.

[0045] Assessment of the strength of belief in the intent of the AI ​​outbound caller: Based on a comprehensive assessment of the reliability of the AI ​​outbound caller's historical intent, the complexity of the intent, the current confidence level of the AI ​​outbound caller, and whether it is abnormal behavior.

[0046] When the belief strength of a certain type of information is lower than a preset threshold, the effective weight of that type of information on the degree of conflict is reduced, and the adjusted weight is renormalized.

[0047] As a further improvement to this technical solution, the feedback and learning described in S6 also include: recording the actual effect and counterfactual prediction results of each intervention action, calculating the calibration signal and updating the parameters of the potential function and / or conflict degree assessment model online; and dynamically adjusting the weights of each sub-dimension and the conflict degree threshold of each takeover mode using reinforcement learning or Bayesian update methods.

[0048] As a further improvement to this technical solution, the dialogue state compensation strategy based on IFS-based reverse path analysis described in S8 includes: searching backwards from the current state towards the IFS boundary to the nearest recoverable state; determining the shortest dialogue path from the current state to the recoverable state; and generating a compensation dialogue sequence based on the path. The compensation dialogue sequence is used to repair the dialogue topology damage caused by the AI ​​dialogue in the first N rounds of dialogue after the human agent takes over.

[0049] As a further improvement to this technical solution, an online graph isomorphism detection step is also included: constructing a local dialogue topology graph of the current call online, and searching for isomorphic or nearly isomorphic subgraphs in the historical call graph library; if the retrieved isomorphic subgraph is strongly correlated with IFS, an early warning is triggered as a supplement to the potential energy gradient calculation.

[0050] In a second aspect, the present invention provides a flight escort collaborative outbound call monitoring and multi-mode takeover decision system, comprising: an AI outbound call terminal, used to send structured intent declaration frames to the flight escort monitoring system at preset intervals, wherein the intent declaration frame includes at least: outbound call target, dialogue planning path, communication strategy, risk acceptance level, confidence level, and a summary of decision reasons.

[0051] The accompanying flight monitoring system includes: a risk prediction module, which performs forward risk prediction for intent declaration frames and maps the current call state to a pre-constructed dialogue state space, and calculates the potential gradient of the current state to migrate to the attraction domain of the irreversible failure subgraph IFS based on the potential function.

[0052] The conflict degree calculation module is used to calculate the intent-risk prediction conflict degree, which is a weighted combination of six sub-dimensions, including dialogue path conflict degree, interaction link conflict degree, risk acceptance conflict degree, dialogue resource conflict degree, time conflict degree, and irreversible failure potential energy conflict degree.

[0053] The tiered takeover decision-making module is used to automatically determine whether to enter autonomous mode, suggestion mode, negotiation mode, cooling-off mode, or emergency takeover mode based on the degree of conflict.

[0054] The belief strength assessment module is used to assign belief strength labels to input information and apply them in conflict degree calculation and decision weight adjustment.

[0055] The compensation strategy generation module is used to generate a dialogue state compensation strategy based on the reverse path analysis of IFS when a takeover is decided to be executed.

[0056] The feedback learning module is used to record the decision-making process and dynamically adjust weights and thresholds.

[0057] The audit log module is used to maintain an immutable decision audit log.

[0058] The takeover execution module is used to route calls to human agents or execute preset emergency scripts after the emergency takeover mode or cooldown mode expires.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a dialogue state space and an irreversible failure subgraph (IFS). Based on a potential function, it calculates the potential gradient of the current call state migrating to the attraction domain of the IFS, quantifying the speed and direction of the call's migration to the irreversible failure region. The system can provide early warnings and trigger intervention before the call completely fails, avoiding missing critical intervention windows and significantly improving outbound call conversion rates and customer satisfaction.

[0060] This invention integrates six sub-dimensions—dialogue path conflict degree, interaction link conflict degree, risk acceptance conflict degree, dialogue resource conflict degree, time conflict degree, and irreversible failure potential energy conflict degree—and calculates the intention-risk prediction conflict degree using a weighted average. This conflict degree comprehensively reflects various risks such as dialogue deviation, link quality, risk bias, resource constraints, time conflict, and failure potential energy, providing a precise and quantifiable basis for takeover decisions.

[0061] This invention provides a five-level decision-making system: autonomous mode, suggestion mode, negotiation mode, cooling-off mode, and emergency takeover mode. The system can automatically determine the intervention level based on the degree of conflict, allowing AI to adjust its intentions during the cooling-off period and take over immediately in emergency situations. This achieves a smooth transition from fully automated to human-led operation, avoiding premature intervention that wastes manpower while ensuring that critical intervention opportunities are not missed.

[0062] This invention assesses the belief strength of call data sources, prediction results from the accompanying flight monitoring system, and the intent of the AI ​​outbound caller. When the credibility of a certain type of information falls below a preset threshold, its weight in the conflict calculation is automatically reduced and the system is re-normalized. This mechanism effectively prevents low-quality information from misleading decisions and significantly improves the system's robustness under complex conditions such as noise, dialects, and prediction errors. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention.

[0065] Figure 2 This is a schematic diagram of the system structure connection of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Example: Please refer to Figure 1 As shown, a method for monitoring and multi-mode takeover decision-making of escort-assisted outbound calls is provided, including: S1, Intent declaration: The AI ​​outbound call terminal sends a structured intent declaration frame to the escort monitoring system at a preset period. The intent declaration frame includes at least: outbound call target, dialogue planning path, communication strategy, risk acceptance level, confidence level, and summary of decision reasons.

[0068] The intent declaration frame described in S1 includes: Outbound call targets: including target type, parameters, priority, and alternative targets; the target type includes at least one of notification, survey, collection, marketing, and follow-up calls; Dialogue planning path: including configuration of at least one dialogue node, estimated arrival time, speech progression rate, emotional escalation rate, and objection escalation rate; Communication strategies include: primary voice channel requirements, backup SMS or ticket channels, and data priority allocation; Risk tolerance: The acceptable threshold for risk categories such as customer hanging up, complaints, compliance, and emotional outbursts; Confidence level: A score indicating how well the AI-powered outbound caller understands the intent. Summary of Decision Justification: A text outlining the basis for the decision, not exceeding the preset word limit.

[0069] S2. Risk Prediction: The accompanying monitoring system performs forward risk prediction for the intent declaration frame and maps the current call state to a pre-constructed dialogue state space. Based on the potential function, it calculates the potential gradient of the current state to migrate to the attraction domain of the irreversible failure subgraph IFS. The irreversible failure subgraph is a set of states in the dialogue state space that will cause the call to fail once entered, regardless of how the subsequent dialogue is executed.

[0070] This means: put the current real-time status of the call onto a pre-drawn "dialogue map" to see how fast and in which direction this point is sliding towards the "unsalvageable" failure zone.

[0071] In one specific embodiment, the intention declaration frame described in S2 performs forward risk prediction, which includes at least: path reachability assessment for each dialogue node on the dialogue planning path; prediction of call link quality; estimation of probability distributions for customer hang-up risk, complaint risk, compliance risk, and emotional outburst risk; and detection of time-event conflicts.

[0072] Path reachability assessment is used to determine whether there is a feasible path from the current dialogue node to subsequent dialogue nodes in the dialogue planning path declared by the AI ​​outbound caller. Specifically, the accompanying monitoring system extracts the node sequence of the dialogue planning path and the preconditions for each node from the intent declaration frame (such as customer identity verification, risk disclosure completion, and response to a specific objection). Combining this with the set of preconditions already met in the current call state, it assesses the reachability of each node. If a node's preconditions cannot be met in the current state, or if reaching the node along the planned path requires traversing a state marked as an irreversible failure subgraph (IFS), then the node is determined to be unreachable. The assessment results are output in three levels: reachable, conditionally reachable, and unreachable, and serve as input for calculating the risk acceptance conflict degree and the irreversible failure potential conflict degree.

[0073] Call link quality prediction is used to estimate whether the call link can continuously support the accuracy requirements of speech recognition and natural language processing as it progresses along the dialogue planning path. Specifically, the accompanying monitoring system acquires real-time link parameters of the current call, including speech recognition confidence, speech activity detection frame drop rate, network round-trip latency, jitter, and packet loss rate. Based on historical call data, a mapping relationship between link quality and speech recognition confidence is established, and forward prediction is performed on the link quality at each subsequent dialogue node along the dialogue planning path. When the predicted speech recognition confidence is lower than a preset threshold, or the packet loss rate or latency exceeds a preset upper limit, the dialogue node is determined to have link quality risk. The prediction results are output as a link quality score or risk level and used as input for calculating the interaction link conflict degree.

[0074] The probability distribution estimates of customer hang-up risk, complaint risk, compliance risk, and emotional outburst risk are used to quantify the likelihood of the current call migrating to different failure types. In practice, the accompanying monitoring system extracts the following features from real-time call data: negative emotion words, expressions of objection, complaint keywords, and sensitive compliance words in the speech-to-text; acoustic features such as speech rate, volume, pauses, and silence duration; dialogue rounds, duration of the current point of contact, and number of times the customer repeats inquiries; and risk tags for the customer or similar customers in historical calls.

[0075] The aforementioned features are input into a pre-trained risk prediction model, which can be implemented using gradient boosting trees, recurrent neural networks, or fine-tuning of large language models, etc., and outputs the probability values ​​and confidence intervals for each of the four types of risks. Hang-up risk focuses on predicting the probability that the customer will actively hang up in subsequent rounds; complaint risk focuses on predicting the probability that the customer will express a complaint intention or request to be transferred to a higher level; compliance risk focuses on predicting the probability that AI-generated dialogue may trigger regulatory prohibitions, overstepping of commitments, or information leakage; and emotional outburst risk focuses on predicting the probability that the customer's emotions will escalate from negative to intense confrontation. The probability distributions of the four types of risks together constitute the risk prediction result, serving as the risk acceptance conflict level. Conflict degree with irreversible failure potential energy The calculation input.

[0076] Time-event conflict detection is used to determine whether the planned execution time window of the current outbound call overlaps with or times out of a scheduled constraint event within the system. Specifically, the accompanying monitoring system extracts the expected completion time of the dialogue planning path from the intent declaration frame, obtains the current call duration from the real-time call status, and queries the pre-registered constraints in the system calendar. These constraints include at least: customer appointment time or customer-agreed follow-up time, agent scheduling and shift handover times, compliant no-call periods, outbound call deadlines, system maintenance windows, and single call duration limits.

[0077] The system represents each constraint as a point in time or a time interval and calculates the estimated end time window for the current call. Then, the system checks whether the estimated end time window overlaps with the time interval of each constraint, or whether the estimated completion time is later than the deadline of the constraint. If overlap or timeout occurs, the system records the conflict name, conflict type, overlap duration, remaining available time, and estimated timeout duration, and determines the conflict severity based on the difference between the remaining available time and the estimated completion time. The detection results are output as a list of conflict events and a severity score; if there are no constraints or none overlap, a no-conflict flag is output. This detection result serves as the time conflict level in S3. The calculation input is as follows. Those skilled in the art can implement the above-described time-event conflict detection using other equivalent methods, all of which fall within the scope of protection of this invention.

[0078] In one specific embodiment, the irreversible failure subgraph IFS described in S2 is constructed in the following way: filtering calls that ultimately failed and where the customer explicitly expressed negative emotions from historical call records; performing reverse analysis on the dialogue state sequence of the filtered calls to identify the key state transitions that led to the failure; and determining the boundary of the irreversible failure subgraph through clustering and boundary detection, wherein the boundary is the set of states that cannot be recovered through subsequent dialogues after entering the subgraph.

[0079] The potential function itself does not have a unique, fixed mathematical formula. The potential function is a mathematical tool, and its specific form depends on how "how close to the IFS boundary" is defined and "which dimensions affect this distance." In the context of this invention, one feasible definition is as follows: Let the dialogue state vector be... Each dimension may include customer sentiment value, customer intent state, current conversation node, number of dialogue turns, number of triggered risk tags, etc. Let the IFS boundary be B, then the potential function can be defined as:

[0080] in Representing state To the IFS boundary The weighted distance can be expressed as:

[0081] in The weights for each state dimension, For state With boundary points In the Differences in several dimensions. The range of values ​​is (0,1], when The closer to the boundary, The smaller, The closer to 1; when The farther away from the boundary, The closer it is to 0.

[0082] It can also be expressed in exponential form: ,in The attenuation coefficient controls the sensitivity of the potential function to changes with distance.

[0083] The potential gradient is the gradient of the potential function. Mathematically, for a state vector... The potential energy gradient is:

[0084] The gradient points in the direction of the fastest increase in the potential function value, i.e., in the direction of the IFS attraction region; the magnitude of the gradient... This indicates the speed of migration to the IFS.

[0085] In practice, since the dialogue state space is usually discrete or high-dimensional continuous, the gradient can be approximated as follows: Numerical difference method: for each state dimension ,calculate , divided by This yields an approximate value of the partial derivative in that dimension; Analytical differentiation: When the potential function takes a differentiable form (such as...) When ), the gradient vector can be obtained directly by taking the derivative; Lookup table method: When the state space is discretized into a finite set of states, the potential function value of each state can be pre-calculated, and the gradient is approximated by the difference in potential functions between adjacent states.

[0086] The calculation of the potential gradient based on the potential function includes: encoding the current call state into a state vector in the dialogue state space; defining a potential function based on the weighted distance between the state vector and the IFS boundary, wherein the value of the potential function increases as the distance between the state and the IFS boundary decreases; calculating the gradient vector of the potential function at the state vector in the dialogue state space, wherein the direction of the gradient vector points to the IFS attraction domain, and the magnitude of the gradient vector represents the speed at which the current state migrates to the IFS attraction domain; and normalizing the magnitude of the gradient vector to the [0,1] interval as the input for calculating the irreversible failure potential conflict degree γ_IFS.

[0087] The potential function described in S2 includes: A Markov potential function is constructed based on the dialogue state transition probability. The value of the potential function increases as the distance between the state and the IFS boundary decreases, and the gradient of the potential function points to the IFS attraction domain.

[0088] The potential energy gradient calculation described in S2 includes: The current call state is encoded as a state vector, and the gradient of the potential function of the state vector is calculated in the dialogue state space. The gradient represents the speed and direction of the current state's migration to the IFS attraction domain.

[0089] The accompanying flight monitoring system described in S2 independently accesses outbound calls through signaling bypass and media bypass, does not participate in the generation of dialogue content, and is deployed decoupled from the AI ​​outbound call process.

[0090] S3. Conflict Measure: Calculate the intention-risk prediction conflict degree, which is a scalar in the interval [0,1] and is calculated by weighted combination of six sub-dimensions.

[0091] In one specific embodiment, the intent-risk prediction conflict degree is calculated as follows: the dialogue planning path declared by the AI ​​outbound call terminal is sequentially aligned with the safe dialogue path recommended by the flight monitoring system, and the dialogue path conflict degree is obtained by nonlinear compression and weighted fusion based on the node semantic distance deviation, speech progression rate deviation and emotion escalation rate deviation of the aligned node pair.

[0092] For each dialogue node on the dialogue planning path declared by the AI ​​outbound call terminal, the interaction link conflict degree is obtained through nonlinear compression, degradation compensation and weighted fusion based on the delay gap, packet loss rate and recognition confidence gap between the predicted delay, predicted packet loss rate and predicted speech recognition confidence and the declared link requirements.

[0093] The risk acceptance conflict degree is obtained by weighting the risk category according to the deviation between the predicted risk level of each risk category and the acceptable threshold declared by the AI ​​outbound call terminal.

[0094] Based on the number of dialogue rounds required to complete the outbound call target, the number of human agents, the knowledge base retrieval, and the remaining resources between concurrent resources and the currently available resources of the system, the dialogue resource conflict degree is obtained by weighting by resource type.

[0095] The time conflict degree is obtained by taking the maximum overlap ratio between the expected completion time of the dialogue planning path and the time of the scheduled constraints.

[0096] The current call state is mapped to the dialogue state space. The potential gradient of the current state to the attraction domain of the irreversible failure subgraph (IFS) is calculated based on the potential function. The gradient magnitude is normalized to obtain the irreversible failure potential conflict degree.

[0097] Weighted fusion yields the degree of conflict between intention and risk prediction.

[0098] Dialogue path conflict Compare the node deviations, speech progression rate deviations, and emotional escalation rate deviations between the dialogue planning path declared by the AI ​​outbound call terminal and the safe dialogue path recommended by the escort monitoring system. Let the dialogue planning path declared by the AI ​​outbound caller be:

[0099] Where M is the total number of nodes in the AI-declared path. Indicates the first 1 node .

[0100] The recommended secure communication path for the flight companion monitoring system is:

[0101] Where N is the total number of nodes in the safe path. Indicates the first 1 node .

[0102] Dynamic Time Warping (DTW) is used to align the two paths in time and space, resulting in Q aligned node pairs:

[0103] in: : The sequence number of the aligned node pair ; In the q-th alignment pair, the AI ​​declares the path node in The original index in ; In the q-th alignment pair, the safe path node is... The original index in ; The total number of aligned node pairs is determined by the DTW cumulative distance minimization path, typically... .

[0104] For the Alignment node pairs Calculate the following three sub-deviations respectively: 1. Node distance deviation

[0105] in: :node The semantic embedding vector, with dimension d, can be 128 or 256, and is generated by a pre-trained sentence vector model; :node The semantic embedding vectors are generated from the same sentence vector model; Euclidean norm; : Reference distance constant, taken as the 90th percentile of the semantic distance between all aligned node pairs in historical calls, used to eliminate the influence of dimensions.

[0106] 2. Deviation in the pace of conversation progression

[0107] in: :node The dialogue progression rate at a given point is defined as the reciprocal of the number of dialogue turns required to advance from that point to the next, measured in turns. -1 ; :node The recommended safety script implementation rate is measured in rounds. -1 ; Minimum reference rate constant, taken as 0.1 rounds. -1 To prevent the denominator from being zero.

[0108] 3. Bias in the rate of emotional escalation

[0109] in: :node The emotion escalation rate at a given point is defined as the average change in the customer's emotion value in one round of dialogue before and after that point, expressed in emotion values ​​per round. The emotion value can be a scalar value in the range of [-1, +1], and is output by the emotion recognition model. :node Recommended safe emotional escalation rate, in units of emotional value / round; Minimum reference rate constant, set to 0.05 emotion value / round, to prevent the denominator from being zero.

[0110] Applying the Sigmoid compression function to each of the three sub-biases above, they are mapped to the interval [0, 1]:

[0111]

[0112]

[0113] in: , , Steepness parameter controls the sharpness of the transition of the three sub-deviations near the center point. An empirical value of 1.5 can be used. , : Center parameter, which controls the compression center point of the three sub-deviations respectively, and can be taken as an empirical value of 1.0.

[0114]

[0115] in: , , The fusion weights of node distance, conversational progression rate, and emotional escalation rate satisfy... , It can be preset to (0.4, 0.3, 0.3) or dynamically adjusted according to the business scenario; : Total number of aligned node pairs; The larger the value, the more serious the deviation between the AI's declared dialogue planning path and the recommended safe path.

[0116] Interaction link conflict degree : Measures the discrepancy between the call link requirements declared by the AI ​​outbound call client and the actual quality of available call links;

[0117] , , These are three conflict components: latency, packet loss rate, and identification confidence. As a downgrade compensation factor, , , To integrate weights, For nodes The comprehensive link conflict components; The downgrade compensation factor Determined based on the availability of backup SMS or work order channels and their expected delivery or response rates, to meet [the requirements]. . Specifically, It can be calculated using the following formula:

[0118] in: The compensation intensity coefficient represents the ability of the backup channel to compensate for conflicts with the main voice channel. The larger the value, the stronger the compensation capacity. This is an indicator function; it takes the value 1 when the backup channel is available and meets the preset quality requirements, and 0 otherwise. When the system has no backup channel configured or the backup channel is unavailable, This indicates no compensation; when the backup channel is fully available and has the strongest compensation capability. It can take values ​​close to 0. As an example, If 0.5 is chosen, then when the backup channel is available... This indicates that the conflict has been partially compensated; if If we take 0.8, then This indicates that the conflict has been largely compensated. Those skilled in the art can adjust this according to the actual business scenario. The value of .

[0119] Delay gap: for nodes Obtain the predicted latency of this node from the accompanying flight monitoring system. The maximum acceptable delay declared by the AI ​​outbound calling client. Comparison: When the predicted delay exceeds the demand, calculate the ratio of the excess to the demand delay and treat it as a zero constant; when the predicted delay does not exceed the demand, the gap is 0.

[0120] Packet loss gap: for nodes Obtain the predicted packet loss rate The maximum acceptable packet loss rate stated. Comparison: When the predicted packet loss rate exceeds the demand, calculate the ratio of the excess to the demand packet loss rate and treat it as a zero constant; when the predicted packet loss rate does not exceed the demand, the gap is 0.

[0121] Identify confidence gaps: for nodes Obtain the confidence score for predicted speech recognition. The minimum acceptable confidence level for identification of the statement. Comparison: When the prediction confidence level is lower than the demand confidence level, calculate the ratio of the shortfall to the demand confidence level and treat it as a zero constant; when the prediction confidence level is not lower than the demand confidence level, the gap is 0.

[0122] The three gaps mentioned above are mapped to the [0,1] interval using a nonlinear compression function (such as the Sigmoid function), resulting in... , , The steepness and center parameters of the compression function can be set empirically, for example, a steepness of 1.5 and a center parameter of 1.0.

[0123] Risk acceptance conflict Compare the acceptable risk range declared by the AI ​​outbound calling terminal with the actual risk level assessed by the flight monitoring system; in: Total number of risk categories, for example This corresponds to the risks of customers hanging up, complaints, compliance, and emotional outbursts. Risk Category Index ; The priority weight of risk category r, satisfying , It can be declared by the AI ​​outbound call terminal in the intent declaration frame or preset by the system; Risk Category The conflicting components are ∈ [0,1]; The larger the value, the more severe the deviation between the risk acceptance range declared by AI and the actual predicted risk level.

[0124] Risk Category Conflict components For risk category r, extract the predicted risk level of that risk category from the risk prediction results of the escort monitoring system. Extract the acceptable threshold for that risk category from the risk acceptance level declared by the AI ​​outbound call terminal. When the predicted risk level exceeds the acceptable threshold, the conflict value is calculated as the ratio of the excess to the remaining acceptable space. ,in To prevent a zero constant (e.g., 0.01); when the predicted risk level does not exceed the acceptable threshold, .

[0125] Priority weight The priority of risk categories is determined by the AI-powered outbound caller in the intent statement frame, with priority given to compliance risks, followed by complaint risks, then hang-up risks, and finally, emotional outburst risks. Alternatively, the priority can be preset by the system based on the business scenario.

[0126] Dialogue resource conflict Assess whether the number of conversation rounds, human agent resources, knowledge base retrieval resources, and concurrent resources required to complete the outbound call target are sufficient;

[0127] in: Total number of resource types, for example This corresponds to resources for dialogue rounds, human agents, knowledge base retrieval, and concurrency. Resource type index ; The fusion weight of resource type s, satisfying , The weights can be preset to equal weights or dynamically adjusted according to the business scenario. : Conflicting components of resource type s, ∈ [0,1]; The larger the value, the more strained the resources are required to complete the outbound call target.

[0128] Resource types Conflict components For resource type s, based on the dialogue planning path, outbound target type, and communication strategy declared by the AI ​​outbound call client, estimate the amount of resources required to complete the intent. Get the current available resources of the system. ;Calculate resource reserves ,in To prevent zero constant (e.g., 0.01), when the resource balance is lower than the preset safety threshold. When this happens, the conflict value is calculated based on the ratio of the insufficient portion to the safety threshold, i.e. When the remaining resources are not lower than the safety threshold, .

[0129] The value range is (0, 1), and can be set according to the resource type and business scenario. As an example, the security threshold for dialogue round resources... A value of 0.2 is acceptable, indicating that resource conflicts begin to occur when the remaining dialogue rounds are below 20%; this represents the safe threshold for human agent resources. A value of 0.3 is acceptable, indicating that a resource conflict occurs when the agent idle rate is below 30%; this is the safety threshold for retrieving resources from the knowledge base. A value of 0.15 is acceptable; this is the safe threshold for concurrent resources. A value of 0.25 is acceptable. The security thresholds for different resource types can be the same or different, and those skilled in the art can set them flexibly according to the actual system capacity and business needs.

[0130] Estimation methods for each resource type: Dialogue turn resources are estimated based on the number of nodes in the dialogue planning path and the expected number of dialogue turns; human agent resources are estimated based on the historical takeover rate or the current agent idle rate; knowledge base retrieval resources are estimated based on the number of knowledge base queries bound to the dialogue nodes; and concurrency resources are estimated based on the current system concurrency occupancy rate and task concurrency requirements.

[0131] Time conflict : Detect the degree of time-event conflict within the planning scope;

[0132] in: Total number of constraints; Constraints Index ; : The time conflict component of constraint e, ∈[0,1]; Take the maximum value of the conflict component among all constraints; the larger the value, the more severe the time conflict.

[0133] Constraints Time conflict components Extract the expected completion time of the dialogue planning path from the intent declaration frame declared by the AI ​​outbound call client. ; Query all scheduled constraint items in the system, for each item This is represented as (item name, start time, duration), including customer appointment time, agent scheduling and shift handover time, compliant no-call periods, outbound call deadline, system maintenance window, etc. For each item... Check its time interval and the planned interval. Does it overlap? If it overlaps, calculate the overlap ratio. ,in For the duration of overlap, For matters The duration; if there is no overlap, .

[0134] Selection of the maximum value: Take the maximum value of the conflicting components among all constraints as the time conflict degree, which represents the most urgent time constraint.

[0135] Irreversible failure potential conflict degree Based on the potential energy gradient, the risk of irreversible failure in migrating the current call state to the IFS attraction domain is determined.

[0136] in: : Current call state vector; : A pre-constructed potential function whose value increases as the distance between the state and the IFS boundary decreases; The gradient vector of the potential function at state s, pointing towards the IFS attraction region; The magnitude of the gradient vector represents the speed at which the current state migrates into the IFS attraction domain. The reference gradient magnitude constant is used for normalization and can be taken as the 90th percentile of the gradient magnitude in historical calls. The larger the value, the faster the current call state migrates to the irreversible failure zone.

[0137] Potential function Construction: Encode the current call state into a state vector in the dialogue state space. ,in The number of state dimensions can include customer sentiment value, customer intent state, current dialogue node, dialogue turn, number of triggered risk tags, etc. A Markov potential function is constructed based on the dialogue state transition probability; the value of the potential function increases as the distance between the state and the IFS boundary decreases. One implementable potential function is defined as... ,in The weighted distance from state s to the IFS boundary B is... The attenuation coefficient; , The weights for each state dimension.

[0138] The attenuation coefficient Used to control the sensitivity of the potential function to changes in distance between the state and the IFS boundary. . The larger the value, the faster the potential function decays with increasing distance; that is, the potential function value decreases more rapidly as the state moves away from the IFS boundary. The smaller the value, the smoother the change in the potential function. As an example, when the weighted distance in the state space... When normalized to the [0,1] interval, We can take a value of 1.0, at which point the potential function value at a distance of 0.5 is... The potential function value at a distance of 1.0 is If we want the potential function to be more sensitive to distance, Version 2.0 is acceptable; if a smoother change is desired, A value of 0.5 is acceptable. Those skilled in the art can adjust this value based on the dimensions and distance distribution of the dialogue state space. The value of .

[0139] Calculation of gradient magnitude: Calculating the state in the dialogue state space Potential function gradient vector The gradient direction points towards the IFS attraction region. Gradient magnitude It can be approximated using the numerical difference method, that is, for each state dimension calculate and divide by This yields an approximate value of the partial derivative in that dimension, and then the square root of the sum of the squares of the partial derivatives in all dimensions is taken. The gradient magnitude is then divided by the reference gradient magnitude constant. And truncate it to the [0,1] interval to obtain the irreversible failure potential energy conflict degree. .

[0140]

[0141] in, Intent-Risk Prediction Conflict Degree , , , , , To integrate weights, .

[0142] S4. Layered takeover decision: Automatically decides whether to enter autonomous mode, suggestion mode, negotiation mode, cooling-off mode or emergency takeover mode based on the degree of conflict.

[0143] In one specific embodiment, the autonomous mode is as follows: the flight monitoring system enters a pure observation state, records the intention and execution process of the AI ​​outbound call terminal, and does not actively intervene.

[0144] Recommended mode: The escort monitoring system pushes the conflict analysis report to the AI ​​outbound call terminal or human agent, who can choose to accept the suggestion or stick to the original intention.

[0145] Negotiation Mode: The escort monitoring system triggers an objection negotiation agreement, presents detailed evidence of risk prediction, sets a timeout, and waits for explicit confirmation from the AI ​​outbound call terminal or human agent whether to accept the risk prediction.

[0146] Cooling mode: The escort monitoring system stops executing the current intent but does not force takeover, allowing a cooling-off period to adjust the intent; if the conflict level decreases during the cooling-off period, it will be recalculated; if no adjustment is made by the end of the cooling-off period, it will enter emergency takeover mode.

[0147] The calculation involves re-executing risk prediction in S2 and conflict measurement in S3 based on the modified intent declaration frame from the AI ​​outbound caller during the cooling-off period, resulting in a new total conflict degree Γ and its six sub-dimensions. The object of the calculation is the total conflict degree Γ, the basis for the calculation is the modified intent declaration frame, and the purpose of the calculation is to determine whether the conflict has been reduced to an acceptable level.

[0148] Emergency Takeover Mode: The accompanying monitoring system immediately takes over the outbound call decision, routing the call to a human agent or executing a preset emergency script, overriding the current instructions of the AI ​​outbound call terminal.

[0149] The conflict thresholds for the five takeover modes in S4 are configured as follows: Autonomous mode; : Suggested mode; : Negotiation mode, set negotiation timeout, wait for explicit confirmation from AI outbound caller or human agent; Cooling mode: Sets the cooldown period and allows adjustment of intent; Emergency takeover mode.

[0150] The emergency takeover mode described in S4 further includes: Script takeover: The escort monitoring system pushes a standard emergency script to the human agent, who then continues the call according to the script; Silent takeover: The escort monitoring system keeps the call uninterrupted before the human agent takes over and pushes the complete call context to the human agent; Direct takeover: The escort monitoring system directly routes the call to a human agent, who then seamlessly takes over.

[0151] S5. Belief Strength Weighting: Assign a belief strength label Belief∈[0,1] to each type of input information. The belief strength reflects the credibility of the information in the current context and is applied in conflict degree calculation and decision weight adjustment. When the belief strength of the information is lower than a preset threshold, the contribution weight of the information to the conflict degree is reduced accordingly.

[0152] In one specific embodiment, the belief strength mentioned in S5 specifically includes: evaluating the belief strength of the call data source: dynamically evaluating based on speech recognition confidence, natural language processing intent recognition confidence, multi-source recognition consistency, background noise or dialect interference, and data timeliness.

[0153]

[0154] in, Speech recognition confidence level; Confidence level for natural language processing intent recognition; Multi-source recognition consistency score: This score represents the degree of consistency between speech recognition, intent recognition, and contextual information. : The degree of background noise or dialect interference; the higher the value, the stronger the interference. Data timeliness factor: the higher the value, the newer the data; : Integrate weights to satisfy , It can be preset to (0.25, 0.25, 0.20, 0.15, 0.15); A larger value indicates a more reliable data source for the call.

[0155] The belief strength of the prediction results of the escort monitoring system is evaluated based on the prediction method type, the historical accuracy of the method, the prediction confidence interval width, and the consistency with cross-validation with other independent predictions.

[0156]

[0157] in: The base belief value for the prediction method type, for example, 0.85 for physical model simulation, 0.70 for statistical models, 0.65 for machine learning models, and 0.50 for heuristic rules; The accuracy of this method on historical data; : The normalized value of the predicted confidence interval width; the smaller the value, the more accurate the prediction. Consistency with cross-validation of other independent prediction methods; : Integrate weights to satisfy , It can be preset to (0.20, 0.30, 0.25, 0.25); A larger value indicates a more reliable prediction result.

[0158] Assessment of the strength of belief in the intent of the AI ​​outbound caller: Based on a comprehensive assessment of the reliability of the AI ​​outbound caller's historical intent, the complexity of the intent, the current confidence level of the AI ​​outbound caller, and whether it is abnormal behavior.

[0159]

[0160] in: The reliability of historical intents in AI-powered outbound calling can be defined as the proportion of historical intents that did not lead to dangerous events. : The normalized value of the current intent complexity; the larger the value, the more complex the intent. The level of confidence that the AI-powered outbound calling client declares in the intent declaration frame; : Abnormal behavior flag, set to 1 if the current intent deviates from the historical behavior pattern by more than a threshold, otherwise set to 0; : Integrate weights to satisfy , It can be preset to (0.40, 0.20, 0.25, 0.15); The larger the value, the more credible the intent of the AI ​​outbound caller.

[0161] When the belief strength of a certain type of information is lower than a preset threshold, the effective weight of that type of information on the degree of conflict is reduced, and the adjusted weight is renormalized.

[0162] For each sub-dimension conflict degree Determine the corresponding belief strength. . It can be obtained by synthesizing the belief strength of the information sources on which this sub-dimension depends, for example: Relying on AI intent and prediction results is advisable. Or weighted average; Based on call data sources and prediction results, it is advisable to... ; Relying on prediction results and AI intent is preferable. ; Relying on AI intent and prediction results is advisable. ; Relying on prediction results and AI intent is preferable. ; Based on call data sources and prediction results, it is advisable to... .

[0163] Let the preset belief strength threshold be... (For example, 0.4). When At that time, reduce the effective weight of the conflict degree of this sub-dimension: ,when ; ,when .

[0164] Then renormalize all valid weights:

[0165] The final adjusted total conflict level is:

[0166] This is used for S4 hierarchical takeover decision-making. When the belief strength of a certain type of information is too low, the influence of its sub-dimension conflict degree in the total conflict degree is automatically weakened, preventing low-quality information from misleading decisions.

[0167] S6. Feedback and Learning: Record the complete decision-making process for each frame, including the original input data, the calculation process of conflict degree for each sub-dimension, the final conflict degree and the takeover mode entered, the response of the AI ​​outbound call terminal or human agent to the suggestion, and the deviation between the actual execution result and the prediction; dynamically adjust the weight of each sub-dimension and / or the conflict degree threshold of each takeover mode based on the feedback data.

[0168] In one specific embodiment, the feedback and learning described in S6 further includes: recording the actual effect and counterfactual prediction results of each intervention action, calculating the calibration signal and updating the parameters of the potential function and / or conflict degree assessment model online; and dynamically adjusting the weights of each sub-dimension and the conflict degree threshold of each takeover mode using reinforcement learning or Bayesian update methods.

[0169] S7. Traceability Guarantee: Maintain an immutable decision audit log, with each record containing at least a timestamp, intent statement frame content, risk prediction results, conflict degree calculation results, takeover mode decision, execution actions, and response records.

[0170] S8. Compensation Strategy Generation: When it is decided to take over, a dialogue state compensation strategy is generated based on the reverse path analysis of IFS. The compensation strategy is used to repair the dialogue topology damage caused by AI dialogue in the first N rounds of dialogue after the human agent takes over.

[0171] In one specific embodiment, the dialogue state compensation strategy generated by the reverse path analysis based on IFS in S8 includes: searching backward from the current state to the nearest recoverable state in the direction of the IFS boundary; determining the shortest dialogue path from the current state to the recoverable state; and generating a compensation dialogue sequence based on the path. The compensation dialogue sequence is used to repair the dialogue topology damage caused by the AI ​​dialogue in the first N rounds of dialogue after the human agent takes over.

[0172] In one specific embodiment, an online graph isomorphism detection step is also included: constructing a local dialogue topology graph of the current call online, and retrieving isomorphic or nearly isomorphic subgraphs from the historical call graph library; if the retrieved isomorphic subgraph is strongly correlated with the IFS, an early warning is triggered as a supplement to the potential gradient calculation.

[0173] Please see Figure 2 As shown, a system for monitoring and multi-mode takeover decision-making in flight escort coordination is provided, including: an AI outbound call terminal, used to send structured intent declaration frames to the flight escort monitoring system at preset intervals, wherein the intent declaration frame includes at least: outbound call target, dialogue planning path, communication strategy, risk acceptance level, confidence level, and a summary of decision reasons.

[0174] The accompanying flight monitoring system includes: a risk prediction module, which performs forward risk prediction for intent declaration frames and maps the current call state to a pre-constructed dialogue state space, and calculates the potential gradient of the current state to migrate to the attraction domain of the irreversible failure subgraph IFS based on the potential function.

[0175] The conflict degree calculation module is used to calculate the intent-risk prediction conflict degree, which is a weighted combination of six sub-dimensions, including dialogue path conflict degree, interaction link conflict degree, risk acceptance conflict degree, dialogue resource conflict degree, time conflict degree, and irreversible failure potential energy conflict degree.

[0176] The tiered takeover decision-making module is used to automatically determine whether to enter autonomous mode, suggestion mode, negotiation mode, cooling-off mode, or emergency takeover mode based on the degree of conflict.

[0177] The belief strength assessment module is used to assign belief strength labels to input information and apply them in conflict degree calculation and decision weight adjustment.

[0178] The compensation strategy generation module is used to generate a dialogue state compensation strategy based on the reverse path analysis of IFS when a takeover is decided to be executed.

[0179] The feedback learning module is used to record the decision-making process and dynamically adjust weights and thresholds.

[0180] The audit log module is used to maintain an immutable decision audit log.

[0181] The takeover execution module is used to route calls to human agents or execute preset emergency scripts after the emergency takeover mode or cooldown mode expires.

[0182] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for monitoring and multi-mode takeover decision-making in escort-coordinated outbound calls, characterized in that, include: S1. Intent Declaration: The AI ​​outbound call terminal sends a structured intent declaration frame to the flight monitoring system at a preset period. The intent declaration frame includes at least: outbound call target, dialogue planning path, communication strategy, risk acceptance level, confidence level, and summary of decision reasons. S2. Risk Prediction: The escort monitoring system performs forward risk prediction for the intent declaration frame and maps the current call state to the pre-constructed dialogue state space. Based on the potential function, it calculates the potential gradient of the current state to migrate to the attraction domain of the irreversible failure subgraph IFS. S3. Conflict Measurement: Calculate the intention-risk prediction conflict degree, which is a scalar in the interval [0,1] and is calculated by weighted combination of six sub-dimensions; S4. Layered Takeover Decision: Automatically determines whether to enter autonomous mode, suggestion mode, negotiation mode, cooling-off mode, or emergency takeover mode based on the degree of conflict. S5. Belief Strength Weighting: Assign a belief strength label to each type of input information. When the belief strength of information is lower than a preset threshold, the contribution weight of that information to the degree of conflict is reduced accordingly. S6. Feedback and Learning: Record the complete decision-making process for each frame, including the original input data, the calculation process of the conflict degree of each sub-dimension, the final conflict degree and the takeover mode entered, the response of the AI ​​outbound call terminal or human agent to the suggestion, and the deviation between the actual execution result and the prediction. The weights of each sub-dimension and / or the conflict thresholds of each takeover mode are dynamically adjusted based on feedback data. S7. Traceability Guarantee: Maintain an immutable decision audit log, with each record containing at least a timestamp, intent statement frame content, risk prediction results, conflict degree calculation results, takeover mode decision, execution actions, and response records; S8. Compensation Strategy Generation: When it is decided to take over, a dialogue state compensation strategy is generated based on the reverse path analysis of IFS. The compensation strategy is used to repair the dialogue topology damage caused by AI dialogue in the first N rounds of dialogue after the human agent takes over.

2. The accompanying coordinated outbound call monitoring and multi-mode takeover decision-making method according to claim 1, characterized in that: The intention declaration frame described in S2 performs forward risk prediction, including at least: path reachability assessment for each dialogue node on the dialogue planning path; prediction of call link quality; probability distribution estimation of customer hang-up risk, complaint risk, compliance risk, and emotional outburst risk; and detection of time-event conflicts.

3. The accompanying coordinated outbound call monitoring and multi-mode takeover decision-making method according to claim 1, characterized in that: The irreversible failure subgraph IFS described in S2 is constructed in the following way: The process involves filtering calls from historical call records that ultimately failed and where the customer explicitly expressed negative emotions; performing reverse analysis on the dialogue state sequences of the filtered calls to identify key state transitions that led to the failure; and determining the boundaries of irreversible failure subgraphs through clustering and boundary detection, where each boundary represents a set of states that cannot be recovered through subsequent dialogues after entering the subgraph.

4. The accompanying coordinated outbound call monitoring and multi-mode takeover decision-making method according to claim 1, characterized in that: The degree of conflict between intent and risk prediction is calculated as follows: The dialogue planning path declared by the AI ​​outbound call terminal is sequence aligned with the safe dialogue path recommended by the flight monitoring system. Based on the node semantic distance deviation, speech progression rate deviation, and emotion escalation rate deviation of the aligned node pair, the dialogue path conflict degree is obtained through nonlinear compression and weighted fusion. For each dialogue node on the dialogue planning path declared by the AI ​​outbound call terminal, the interaction link conflict degree is obtained by nonlinear compression, degradation compensation and weighted fusion based on the delay gap, packet loss rate and recognition confidence gap between the predicted delay, predicted packet loss rate and predicted speech recognition confidence and the declared link requirements. The risk acceptance conflict degree is obtained by weighting the risk category according to the deviation between the predicted risk level of each risk category and the acceptable threshold declared by the AI ​​outbound call terminal. Based on the number of dialogue rounds, human agents, knowledge base retrieval, and the remaining resources between concurrent resources and the currently available resources of the system required to complete the outbound call target, the dialogue resource conflict degree is obtained by weighting the resources by resource type. The time conflict degree is obtained by taking the maximum overlap ratio between the expected completion time of the dialogue planning path and the time of the scheduled constraints. Map the current call state to the dialogue state space, calculate the potential gradient of the current state to the attraction domain of the irreversible failure subgraph IFS based on the potential function, and normalize the gradient magnitude to obtain the irreversible failure potential conflict degree. Weighted fusion yields the degree of conflict between intention and risk prediction.

5. The accompanying coordinated outbound call monitoring and multi-mode takeover decision-making method according to claim 1, characterized in that: The autonomous mode: The flight monitoring system enters a pure observation state, records the intention and execution process of the AI ​​outbound call terminal, and does not actively intervene; Recommended mode: The escort monitoring system pushes the conflict analysis report to the AI ​​outbound call terminal or human agent, who can choose to accept the suggestion or stick to the original intention. Negotiation Mode: The escort monitoring system triggers an objection negotiation agreement, presents detailed evidence of risk prediction, sets a timeout, and waits for explicit confirmation from the AI ​​outbound call terminal or human agent whether to accept the risk prediction. Cooling mode: The flight monitoring system stops executing the current intent but does not force takeover, allowing a cooling-off period for the intent to be adjusted; If the conflict level decreases during the cooling-off period, it will be recalculated; if no adjustment is made by the end of the cooling-off period, it will enter emergency takeover mode. Emergency Takeover Mode: The accompanying monitoring system immediately takes over the outbound call decision, routing the call to a human agent or executing a preset emergency script, overriding the current instructions of the AI ​​outbound call terminal.

6. The accompanying coordinated outbound call monitoring and multi-mode takeover decision-making method according to claim 1, characterized in that: The strength of belief described in S5 specifically includes: Belief strength assessment of call data sources: dynamically assessed based on speech recognition confidence, natural language processing intent recognition confidence, multi-source recognition consistency, background noise or dialect interference, and data timeliness; The belief strength assessment of the prediction results of the escort monitoring system is based on a comprehensive evaluation of the prediction method type, the historical accuracy of the method, the prediction confidence interval width, and the consistency with cross-validation of other independent predictions. Assessment of the strength of belief in the intent of the AI ​​outbound caller: based on a comprehensive assessment of the reliability of the AI ​​outbound caller's historical intent, the complexity of the intent, the current confidence level of the AI ​​outbound caller, and whether it is abnormal behavior; When the belief strength of a certain type of information is lower than a preset threshold, the effective weight of that type of information on the degree of conflict is reduced, and the adjusted weight is renormalized.

7. The accompanying coordinated outbound call monitoring and multi-mode takeover decision-making method according to claim 1, characterized in that: The feedback and learning described in S6 also include: Record the actual effects and counterfactual prediction results of each intervention action, calculate the calibration signal and update the parameters of the potential function and / or conflict assessment model online; dynamically adjust the weights of each sub-dimension and the conflict threshold of each takeover mode using reinforcement learning or Bayesian update methods.

8. The accompanying coordinated outbound call monitoring and multi-mode takeover decision-making method according to claim 1, characterized in that: The dialogue state compensation strategy based on IFS-based reverse path analysis described in S8 includes: Search backwards from the current state towards the IFS boundary to the nearest recoverable state; determine the shortest dialogue path from the current state to the recoverable state; generate a compensation dialogue sequence based on the path, which is used to repair the dialogue topology damage caused by the AI ​​dialogue in the first N rounds of dialogue after human agent takeover.

9. The method for monitoring and multi-mode takeover decision-making in escort-coordinated outbound calls according to claim 1, characterized in that: It also includes an online graph isomorphism detection step: The system constructs a local dialogue topology graph of the current call online and retrieves isomorphic or nearly isomorphic subgraphs from the historical call graph library. If the retrieved isomorphic subgraph is strongly correlated with the IFS, an alert is triggered as a supplement to the potential gradient calculation.

10. A system for monitoring and multi-mode takeover of escorted outbound calls, used to execute the method for monitoring and multi-mode takeover of escorted outbound calls as described in any one of claims 1-9, characterized in that: include: The AI ​​outbound calling terminal is used to send structured intent declaration frames to the flight monitoring system at preset intervals. The intent declaration frame includes at least: outbound calling target, dialogue planning path, communication strategy, risk acceptance level, confidence level, and summary of decision reasons. The escort monitoring system includes: The risk prediction module is used to perform forward risk prediction for intent declaration frames and map the current call state to a pre-built dialogue state space. It calculates the potential gradient of the current state to migrate to the attraction domain of the irreversible failure subgraph IFS based on the potential function. The conflict degree calculation module is used to calculate the intention-risk prediction conflict degree, which is a scalar in the range [0,1] and is calculated by weighted combination of six sub-dimensions; The tiered takeover decision-making module is used to automatically determine whether to enter autonomous mode, suggestion mode, negotiation mode, cooling-off mode, or emergency takeover mode based on the degree of conflict. The belief strength assessment module is used to assign belief strength labels to input information and apply them in conflict degree calculation and decision weight adjustment. The compensation strategy generation module is used to generate a dialogue state compensation strategy based on the reverse path analysis of IFS when it is decided to take over. The feedback learning module is used to record the decision-making process and dynamically adjust weights and thresholds; The audit log module is used to maintain tamper-proof decision audit logs; The takeover execution module is used to route calls to human agents or execute preset emergency scripts after the emergency takeover mode or cooldown mode expires.