[Paper Review] Orchestration-Free Customer Service Automation: A Privacy-Preserving and Flowchart-Guided Framework

Mengze Hong, Chen Jason Zhang+4 more|arXiv (Cornell University)|Feb 17, 2026

Business Process Modeling and Analysis0 citations

TL;DR

The paper proposes an orchestration-free, privacy-preserving framework for end-to-end customer service automation using Task-Oriented Flowcharts (TOFs), with a two-step dialogue-to-flowchart construction method and decentralized distillation to train small language models locally.

ABSTRACT

Customer service automation has seen growing demand within digital transformation. Existing approaches either rely on modular system designs with extensive agent orchestration or employ over-simplified instruction schemas, providing limited guidance and poor generalizability. This paper introduces an orchestration-free framework using Task-Oriented Flowcharts (TOFs) to enable end-to-end automation without manual intervention. We first define the components and evaluation metrics for TOFs, then formalize a cost-efficient flowchart construction algorithm to abstract procedural knowledge from service dialogues. We emphasize local deployment of small language models and propose decentralized distillation with flowcharts to mitigate data scarcity and privacy issues in model training. Extensive experiments validate the effectiveness in various service tasks, with superior quantitative and application performance compared to strong baselines and market products. By releasing a web-based system demonstration with case studies, we aim to promote streamlined creation of future service automation.

Motivation & Objective

• Define a Task-Oriented Flowchart (TOF) structure to provide procedural guidance for service automation.
• Develop a cost-efficient method to construct TOFs from service dialogues via representative sampling and iterative flowchart construction.
• Enable orchestration-free automation by flowchart-guided prompting and decentralized distillation to train small language models locally.
• Validate TOFs through quantitative benchmarks and qualitative assessments across multiple service tasks and datasets.
• Release a web-based system demonstration to illustrate practical deployment and guide future research.

Proposed method

• Introduce a five-node TOF schema (Start, Action/Decision, Output, Reflection, End) with directed edges and Mermaid-compatible representation.
• Formulate Utterance Matching Ratio (UMR) and Complete Path Coverage (CPC) as flowchart evaluation metrics grounded in semantic matching and path completeness.
• Solve Weighted Dialogue Intent Coverage (WDIC) via a two-step approach: (i) randomized/Rounding-based selection from dialogues to cover intents; (ii) iterative, domain-aware flowchart construction driven by local small language models (SLMs).
• Construct TOFs from dialogue data using three oracles (domain classifier, intent descriptor extractor, node-typifier) and an ILP-based or LP-rounding dialogue selection pipeline for cost-efficient coverage.
• Propose flowchart-guided prompting (flowchart-aware prompts and active node tracking) and a decentralized flowchart distillation framework to train lightweight models on local data while generating synthetic data from global flowcharts.

Experimental results

Research questions

• RQ1 Can TOFs provide comprehensive procedural guidance for end-to-end customer service automation with reduced need for agent orchestration?
• RQ2 How can TOFs be constructed cost-effectively from service dialogues while ensuring coverage of intents and dialogue paths?
• RQ3 Do flowchart-guided prompting and decentralized distillation enable high task completion with small, locally deployed models?
• RQ4 How do flowcharts perform on benchmark TOD tasks compared to existing training- and prompting-based baselines, and in real outbound service settings?
• RQ5 What privacy protections and practicality benefits do TOFs offer in distributed training environments?

Key findings

• TOFs enable orchestration-free task coordination and preserve privacy via decentralized data processing and synthetic data generation.
• Iterative flowchart construction from dialogues achieves strong CPC and UMR on MultiWoZ 2.0 and SGD, often outperforming human-annotated and abstraction-based baselines in coverage metrics.
• Flowchart-guided prompting with LLaMA-3-8B and GPT-3.5-based prompts yields competitive or superior task success and completion rates compared to several baselines on benchmark TOD datasets.
• In outbound banking deployments, a decentralized flowchart distillation pipeline trained on locally generated synthetic data outperforms commercial baselines in Exemplary Rate and Completion Rate, indicating practical benefits for real-world deployments.
• The approach demonstrates that smaller open-source models can match or exceed larger proprietary models when guided by flowchart-based procedural priors.