Dual AI Engines Predict Timeouts, Diagnose Root Causes in Bank Long Processes

Background: New Governance Challenges After Process Digitization

As digital China and "AI+" strategies advance, banking processes such as corporate credit, inclusive loans, guarantee issuance, and operational review have moved online via an enterprise-level process integration platform. Full-chain traceability now captures initiation, node transitions, returns, corrections, and completion. However, storing process data does not equal using it. Traditional reports show current node and historical duration but cannot timely judge whether a running case will time out, why it slows, or what action to take. The focus must shift from connecting and recording processes to understanding, predicting, and optimizing them.

Core Challenges

Late risk detection: Duration and return anomalies appear only in post-hoc reports or special reviews, often after deadlines have passed, leaving only remedial action.

Slow anomaly diagnosis: Slowdowns may stem from missing materials, abnormal paths, node overload, or collaboration waits. Reports provide statistics but cannot simultaneously link regulations and historical cases, forcing cross-team manual verification at high cost.

Weak experience reuse: Return reasons, handling methods, and expert judgments scatter across approval comments, emails, reports, and personal experience, preventing a maintained, unified organizational knowledge base. Similar problems trigger repeated investigations.

The common root cause: process data, predictive models, regulatory knowledge, and business handling have not formed a stable closed loop. The solution must both anticipate risks and explain them, support handling, and feed actual outcomes back into models and knowledge.

Method and Implementation: A "Prediction + Diagnosis" Dual-Engine System

The overall architecture adopts five layers — process data → specification knowledge → dual engines → intelligent services → business applications — chaining trace processing, risk prediction, knowledge retrieval, human review, and business handling into a complete technical pipeline from data to action.

1. Turning Process Traces into Computable Event Logs

Raw data comes from process instances and node events: initiation, acceptance, transition, return, completion, with timestamps, organizations, and roles. By reconstructing event chains per instance and node order, unifying state and statistical calibers, and extracting features such as node duration, wait time, return count, path variants, and historical performance — then labeling against business SLAs — scattered operation records become predictable, explainable, and traceable event logs.

2. Letting the Prediction Engine and Diagnostic Agent Each Do Their Job

FlowPredict builds separate models for duration, timeout, return, bottleneck, and load tasks, outputting standardized signals: estimated remaining time, risk probability, and abnormal nodes. It identifies which instances need priority attention from the mass of running processes (see Table 1). FlowCopilot combines current process state with retrieval of relevant regulations, process specifications, and historical cases to produce structured diagnostic drafts: risk causes, cited bases, correction prompts, and items for review. The prediction engine answers "which processes need attention and how high is the risk"; the diagnostic agent answers "why attention is needed and what is the evidence." Both engines output analysis references only — they do not replace business personnel's approval decisions.

3. Constraining LLM Diagnosis with Specification Knowledge

The knowledge base is divided into four asset types:

Rules: explicit access conditions, compliance boundaries, and mandatory manual review items.

Specs: metric definitions and output templates.

Skills: solidified operational steps for timeout diagnosis, return attribution, and node anomaly handling.

Knowledge: confirmed return causes, bottleneck cases, and expert opinions.

Initial construction focuses on key processes to form a minimum viable knowledge set, with clear source, version, effective time, and scope. Diagnostic results display cited bases; when bases are insufficient, rules conflict, or scope is exceeded, the case escalates to human verification.

4. Building a "Predict → Diagnose → Review → Handle → Feedback" Business Closed Loop

Process events enter FlowPredict, which outputs estimated remaining time and timeout/return risks. FlowCopilot retrieves regulations, SOPs, and similar cases to generate a structured diagnostic draft. Business staff verify business facts and cited bases, confirm valid prompts, and correct deviations. Reviewed results translate into actions: urging, material supplementation, collaborative communication, or task rescheduling. Actual outcomes, review opinions, and handling effects flow back to calibrate models and update knowledge. The LLM explains signals, finds bases, and lists review items but does not automatically change approval conclusions; all key business judgments remain confirmed by authorized staff.

Construction Effectiveness: From Capability Formation to Business Metric Improvement

Pilots targeted long processes with relatively complete node traces, clear time boundaries, and sufficient historical samples — initially credit, guarantee, and letter-of-guarantee scenarios. A capability system linking event-log processing, multi-task prediction, controlled diagnosis, human review, and result feedback has taken shape.

Phase pilot results (under existing statistical calibers):

Process completion time compressed by 20% .

Business overtime rate reduced by 10% .

Return rate reduced by 20% .

First-pass rate improved by 20% .

Single-case anomaly investigation time cut from average 10 minutes to 2 minutes .

Future work will extend pilots to more businesses, refine pilot cycles, baselines, and metric definitions, and continuously validate model stability and business improvement.

Value Realization: From Point Efficiency to Systematic Process Governance

Governance mode shifted forward: Timeout and return risks move from post-summary to pre-warning, reserving time for material correction, customer communication, and collaborative handling — turning management from remediation to prevention.

Collaboration efficiency improved: Process nodes, regulations, and historical cases are presented in association, assisting cause localization and reducing cross-system verification and repetitive communication between business and tech staff.

Organizational knowledge precipitated: Scattered return reasons, handling steps, and review opinions become maintainable knowledge assets, unifying diagnostic calibers for similar problems and reducing understanding bias from personnel experience gaps.

Method capability reused: Generalized methods — event-log processing, label design, dual-engine collaboration, human review — provide a foundation for expanding to similar processes characterized by many nodes, long cycles, and frequent returns.

Future Outlook: From Key-Scenario Validation to Platform Reuse

The value of process intelligence lies not in slapping a "large model" label on traditional systems, but in giving processes true perception, prediction, explanation, and improvement capabilities. When every risk alert has a data source, every diagnostic suggestion has a regulatory basis, and every key conclusion undergoes human confirmation, bank long-process governance can advance from "seeing results" to "understanding process, intervening early, continuously optimizing" — achieving full-chain governance that is "visible, predictable, explainable."