full-cycle sales

This work addresses the discrepancy between high offline metrics and poor online performance of lead-ranking models in CRM systems by proposing SalesLoop, a closed-loop reinforcement learning framework. SalesLoop introduces a performance-aware reward mechanism and a novel Discriminative Group Relative Policy Optimization (Discriminative GRPO) method, which for the first time adapts group relative policy optimization to discriminative ranking models. This enables listwise objective optimization and dynamic policy adaptation under temporal distribution shifts. Experimental results demonstrate that the approach improves NDCG@K and P@K by 7.9% and 15.8%, respectively. A 160-day A/B test shows a significant 4.7%–8.7% increase in cumulative conversion rate, achieves a 44.1% recall rate within the top-10% ranked leads, and enhances conversion rates for high-intent leads by 2.3×.

This work addresses the limitations of traditional enterprise messaging systems, which rely heavily on external connectors and consequently suffer from fragmented data between messages and CRM records, achieving only eventual consistency through brittle synchronization tasks that hinder seamless automation and unified state management. To overcome these challenges, the authors propose and implement a platform-native messaging architecture that treats messages as first-class CRM entities. By leveraging platform events, asynchronous delivery, multi-tenant decoupling, and standard database object models, this approach deeply integrates the entire message lifecycle into CRM transactions, workflows, and reporting systems. The architecture has been successfully deployed at scale across multiple industries—including healthcare, sales, and field service—demonstrating significant improvements in system consistency, scalability, and automated integration capabilities.

To address challenges in commercial management system development—including poor alignment between process models and execution platforms, low model reusability, and suboptimal development efficiency—this paper proposes a metamodel-based Model-Driven Development (MDD) approach. We design an evolvable and extensible business process metamodel framework and introduce a staged model transformation mechanism supporting QVT/ATL, enabling automated adaptation of extended BPMN models to diverse execution platforms. Crucially, we deeply integrate MDD into BPM system construction, establishing business models as the authoritative source governing system behavior. Experimental evaluation demonstrates significant improvements in development productivity and model consistency, robust cross-platform model reuse, and validates the metamodel’s effectiveness and flexibility in extended application scenarios such as resource management and customer relationship management.

Existing CRM evaluation benchmarks fail to capture real-world business complexity, hindering the integration and validation of AI agents in professional settings. Method: We introduce CRMArena—the first industrial-grade CRM workflow benchmark—featuring nine realistic tasks across three roles (service agent, analyst, manager), grounded in 16 highly interdependent object types and latent-variable modeling to encode intricate business logic and regulatory constraints. It uniquely integrates domain expert knowledge with latent-variable formalization and evaluates agents via dual paradigms: ReAct prompting and structured function calling, under high-fidelity object-relational modeling and dynamic data distribution simulation. Contribution/Results: CRMArena systematically exposes critical LLM agent limitations in rule adherence and structured function invocation. Experiments show state-of-the-art LLM agents achieve only 40% task completion under ReAct and 55% under function calling—highlighting the stringent demands of real-world CRM on robustness, compliance, and structured operational fidelity.

Business process optimization remains challenging due to fragmented methodologies across process mining, predictive process monitoring, and process-aware recommendation—each operating in isolation without a unified theoretical foundation or integration framework. Method: This paper proposes a closed-loop optimization framework that systematically integrates Alpha algorithm/Inductive Miner for process discovery, LSTM/Transformer for runtime prediction, collaborative filtering/graph neural networks for action recommendation, and explainable AI (XAI) for interpretability—enabling automated bottleneck identification, anomaly forecasting, and prescriptive optimization from event logs. Contribution/Results: We establish the first unified conceptual boundary, evolutionary taxonomy, and synergy paradigm across the three domains; construct a comprehensive classification schema covering 120+ studies; clarify application scopes and standardized evaluation benchmarks; and deliver an industrially actionable methodology selection guide with validated deployment pathways.

Traditional revenue forecasting approaches struggle to uncover the underlying customer behavioral drivers—such as customer acquisition, repeat purchase rates, and average transaction value—that influence revenue dynamics. To address this limitation, this work proposes the Customer-Based Multi-Task Transformer (CBMT), which uniquely integrates multi-task learning with a Transformer architecture to jointly model customer behavioral metrics and total revenue through shared representations. Furthermore, CBMT incorporates a downstream alignment mechanism to enhance both interpretability and predictive accuracy. Empirical evaluation on real-world customer transaction panel data demonstrates that CBMT outperforms existing methods across 23 out of 24 evaluation metrics, achieving a 30% reduction in total sales prediction error compared to the strongest baseline and significantly surpassing single-task models employed by 74.3% of firms.

This study addresses the challenge of skill acquisition for LLM-based agents in industrial planning, which is hindered by heterogeneous and incomplete multi-source knowledge. To this end, we propose a cross-source skill induction and execution verification framework. The method extracts coordination patterns from discrepancies between documentation and practice, compiles them into standardized skill packages anchored by COM specifications, and achieves iterative refinement through unlabeled compliance screening and signal-conditioned trajectory attribution. Evaluated on the PIMS-Bench benchmark across four LLM backbones, the proposed approach yields absolute improvements of 14%–30% in component matching F1 scores, with particularly significant gains observed on complex tasks.

This work addresses the lack of evaluation benchmarks for end-to-end operational capabilities of large language model (LLM) agents in real-world business settings. We introduce the first simulated marketplace grounded in authentic cross-border trade data from Alibaba.com, enabling AI agents to execute long-horizon decisions—including procurement, pricing, sales, and compliance—to maximize profitability. To assess performance realistically, we propose an end-to-end evaluation framework incorporating opportunity estimation, skill-level metrics, and action-level reward attribution, thereby avoiding overreliance on single profit indicators or environmental shortcuts. Systematic evaluation of 15 state-of-the-art LLMs reveals up to a nine-fold difference in final net worth, with even the best-performing agent significantly underperforming human strategies. Our analysis further uncovers key value-creating and value-destroying behaviors and distinct operational styles across agents.

This study addresses the vulnerability of language models to misleadingly optimistic assertions from CRM stakeholders during sales qualification, wherein incentive-misaligned statements are erroneously treated as objective evidence, thereby compromising decision-making. To investigate this, we propose a diagnostic framework that distinguishes persuasion effects from information deficits, employing bucket analysis, same-information controlled experiments, and computational step-length control for systematic evaluation. Our findings reveal that neither scaling nor enhanced reasoning mitigates such biases. Across seven mainstream models, false approval rates reach 87–97% under contradictory assertions, confirming that this failure mode is fundamentally rooted in persuasion rather than information insufficiency. This work offers a novel perspective on the reliability of large language models in high-stakes business applications.

This work addresses the heavy reliance on manual effort in chemical process modeling, which is prone to catastrophic failure due to single-point errors. To overcome this limitation, the authors propose a role-adaptive collaborative framework that decomposes the modeling task into seven specialized sub-roles. By integrating natural language, process flow diagrams, and domain knowledge, the framework generates structured models through typed intermediate representations and a deterministic engineering gating mechanism, enabling automated optimization. Leveraging a fine-tuned Qwen large language model for three critical roles—visual, topological, and specification—the system is integrated with a LangGraph workflow and the IDAES/Pyomo solvers. Evaluated on 82 held-out cases from the OpenIDAES-450 dataset, the approach achieves a 91.5% model construction success rate, with F1 scores of 0.815, 0.791, and 0.782 for unit operations, material streams, and connections, respectively.