Source-linked AI summary
A Unified Conversational Assistant Framework for Business Process Automation
Yara Rizk, Abhishek Bhandwalder, Scott Boag, Tathagata Chakraborti, Vatche Isahagian, Yasaman Khazaeni, Falk Pollock, Merve Unuvar
TL;DR
Business-process automation needs accessible interfaces for users to monitor and customize autonomous agents, while existing systems face integration and deployment challenges. The paper presents a multi-agent orchestration framework that composes heterogeneous skills and agents through a conversational interface, and demonstrates it in travel preapproval and loan application processes. The framework supports multiple capabilities with minimal integration overhead, but scalability and agent overlap remain open challenges.
Problem
Business-process automation adoption is limited by the difficulty business users face in monitoring and customizing autonomous agents, along with integration, coding-overhead, scalability, and compatibility challenges.
Source: Introduction · p. 1, Business Process Automation · p. 2, Conversational Agents · p. 3Method
The paper presents a multi-agent conversational framework that composes heterogeneous agents through an orchestrator using defined agent contracts, skills, and execution controls.
Source: Introduction · p. 2, Agents · p. 3, Alerting Agents · p. 4Results
The framework is demonstrated through travel preapproval and loan application assistants supporting querying, task execution, alerts, visualization, and document analysis.
Source: Introduction · p. 2, Use Case – 2: Loan Application · p. 7Takeaways & Limitations
An orchestrator can combine independently developed skills into a more capable conversational assistant with minimal integration overhead.
Source: Coding Overhead · p. 7, Conclusion · p. 8The framework still faces scalability and agent-overlap challenges, and effective skill composition requires compatible interfaces and a unified vocabulary for variables.
Source: Agents · p. 4, Scalability · p. 7, Conclusion · p. 8
Abstract
from arXiv · showhide
Business process automation is a booming multi-billion-dollar industry that promises to remove menial tasks from workers' plates -- through the introduction of autonomous agents -- and free up their time and brain power for more creative and engaging tasks. However, an essential component to the successful deployment of such autonomous agents is the ability of business users to monitor their performance and customize their execution. A simple and user-friendly interface with a low learning curve is necessary to increase the adoption of such agents in banking, insurance, retail and other domains. As a result, proactive chatbots will play a crucial role in the business automation space. Not only can they respond to users' queries and perform actions on their behalf but also initiate communication with the users to inform them of the system's behavior. This will provide business users a natural language interface to interact with, monitor and control autonomous agents. In this work, we present a multi-agent orchestration framework to develop such proactive chatbots by discussing the types of skills that can be composed into agents and how to orchestrate these agents. Two use cases on a travel preapproval business process and a loan application business process are adopted to qualitatively analyze the proposed framework based on four criteria: performance, coding overhead, scalability, and agent overlap.
Introduction
Service automation expanded from customer-care chatbots and RPA toward broader business-process automation, but deployment remains constrained by trust, expertise, scalability, and user-experience challenges. The paper proposes a proactive multi-agent conversational framework with heterogeneous agents and an orchestration layer, demonstrated in travel preapproval and loan application processes.
- Business process automation extends RPA beyond repetitive tasks to decision making, data management, content digitization, and workflow improvement.RPA covers simple activities such as form filling and data mining, whereas BPA incorporates technologies including OCR and decision-making models.Source: Introduction · p. 1
- Autonomous-system deployment remains limited by trustworthiness, coding overhead, scalability, lifecycle management, technical expertise, and user experience.The paper frames a conversational interface as a way to reduce the technical expertise required to interact with, monitor, and customize autonomous agents.Source: Introduction · p. 1
- The paper proposes a multi-agent orchestration framework combining dialog, retrieval, task-execution, and alerting agents with scorer, selector, and sequencer components.The framework is intended to integrate independently developed algorithms through a single conversational interface with minimal coding overhead.Source: Introduction · pp. 1–2
- Two sandbox proof-of-concept assistants apply the framework to travel preapproval and loan application business processes.These use cases are used to discuss how the framework addresses integration, interaction, monitoring, and customization challenges.Source: Introduction · p. 2
Related Work
Related work describes automation progressing from RPA for repetitive tasks toward broader BPA and conversational enterprise interfaces. The paper positions multi-agent composition as an alternative to single agents whose expansion requires substantial coding and may face compatibility constraints.
Business Process Automation
- RPA automates repetitive interface tasks and has been applied across accounting, auditing, human resources, banking, public administration, and energy.Related approaches include document-flow automation, learned form-to-rule agents, and machine-learning methods for identifying RPA-eligible tasks.Source: Business Process Automation · p. 2
- BPA broadens automation toward decision making and process adaptation, while deep-learning components improve performance but remain costly to train and difficult to explain.Automated planning is discussed as a way to generate process models, adapt behavior to unforeseen circumstances, and perform conformance checking.Source: Business Process Automation · p. 2
- End-to-end BPA has struggled in real-world deployment because trust, reliability, and limited stakeholder technical skills hinder monitoring and intervention.The related-work discussion identifies natural-language interaction with autonomous agents as a possible route toward greater adoption in service enterprises.Source: Business Process Automation · p. 2
Conversational Agents
- Enterprise chatbots commonly prioritize predictable, controllable behavior through domain-specific natural-language systems, although data-driven deep-learning approaches are also used.Applications include customer support, shopping, delivery, software-development analysis, and business-process dialog trees; some approaches require substantial overhead and domain knowledge.Source: Conversational Agents · p. 2
- Prior chatbot designs generally focus on single agents, whose expansion can require significant coding and whose components may be difficult to reuse because of compatibility constraints.The proposed alternative uses multiple independently designed specialized agents with narrower scopes that present as one conversational agent to users.Source: Conversational Agents · pp. 2–3
Overall Framework
The framework organizes a proactive conversational assistant around skills, agents, and an orchestrator. The orchestrator previews candidate responses and confidence scores before selecting agents to execute and return responses.
- The framework consists of skills, agents, and an orchestrator, with agents composed of understand, act, and respond skills.Skills perform defined input-output tasks, and each agent supplies an execution pattern ranging from a sequential script to an execution graph.Source: Overall Framework · p. 3
- For each user utterance, the orchestrator forwards input to agents, requests response previews and confidence scores, then selects one or more agents for execution.The selected agents execute according to the adopted orchestration pattern and return responses to the user.Source: Overall Framework · p. 3
Skills
Skills are atomic, role-specific functions that agents compose to understand inputs, perform actions, and formulate responses. They are distinguished by whether they change the world, enabling both conversational and event-driven business-process automation.
Skills
- Skills are atomic building blocks categorized as understand, act, or respond, and as world-changing or non-world-changing.This classification organizes both their role within an agent and their effect on system state.Source: Skills · p. 3
Understand skills
- Understand skills identify intents, entities, keyphrases, or system events that determine when an agent should respond.In conversational settings they classify utterances and annotate entities; outside dialogue they can watch for logins, database changes, and other events.Source: Understand skills · p. 3
Act skills
- Act skills process inputs through world-changing or non-world-changing operations such as checking credit scores, sending email, reading email, or querying information sources.Within business processes, they can gather information, advance decision points, submit applications, or forward collected information.Source: Act skills · p. 3
Respond skills
- Respond skills transform act-skill outputs into user-facing responses ranging from templates to full natural-language generation, visualizations, or images.The response modality can therefore extend beyond text to plots and other visual outputs.Source: Respond skills · p. 3
Agents
The framework composes modular skills into agents with a shared contract, covering dialogue, information retrieval, task execution, and alerting functions.
Agents
- A shared agent contract exchanges an utterance and system context for preview and execute responses, updated context, and confidence.Source: Agents · p. 3
- Static composition is straightforward, while planning-based composition requires compatible skills and a unified vocabulary for their inputs and outputs.Source: Agents · pp. 3–4
- Agents combine understand, act, and respond skills, with zero or more act skills selected according to functionality.Source: Agents · p. 4
Dialog Agents
- Dialog agents answer queries without changing the world, whereas information retrieval agents query external services and may treat cost or side effects as world changing.Source: Dialog Agents · p. 4, Information Retrieval Agents · p. 4
Task Execution Agents
- Task execution agents advance business processes by submitting applications or making decisions, so they change the world and may require preview and execute modes.Source: Task Execution Agents · p. 4
Alerting Agents
- Alerting agents watch system events, generate notifications, and use natural-language queries to support broader, more portable, lower-overhead trigger configuration.Source: Alerting Agents · p. 4
Orchestration
The framework uses a posterior 3S orchestrator that scores agent previews, selects responders, and sequences their outputs while managing context and state trade-offs.
Orchestration
- Stateless orchestration passes context between agents and the orchestrator, while stateful orchestration centralizes state at potential deployment and scalability cost.Source: Orchestration · p. 4
- The 3S orchestrator is a posterior design whose pipeline scores, selects, and sequences agents using confidence and other metrics.Source: Orchestration · p. 5, Scorer · p. 5, Selector · p. 5
Scorer
- The scorer receives preview responses and confidence values, then may forward, normalize, or augment scores with conversational and feedback-based measures.Source: Scorer · p. 5
Selector
- The selector chooses one or more agents using top-score, heuristic, supervised, or reinforcement-learning criteria before requesting execution responses.Source: Selector · p. 5
- A sequencer orders multiple execution responses using scores, rules, or learned models, while Top 1 selection requires no sequencing.Source: Sequencer · p. 5
Use Case – 1: Travel Preapproval Application
TravelBot applies the framework to a Slack-based travel-preapproval process, combining conversational, retrieval, execution, estimation, visualization, and alerting agents.
Travel Preapproval Business Process
- The process collects employee travel requests, routes them to managers for approval, rejection, or revision, and sends approved applications to directors for final decisions.Source: Travel Preapproval Business Process · pp. 5–6
Agents in Travelbot
- TravelBot coordinates multiple agents through the 3S orchestrator to answer manager requests or execute actions in a simplified travel-preapproval process.Source: Use Case – 1: Travel Preapproval… · p. 5, Agents in Travelbot · p. 6
Chit-Chat Agent
- A Chit-Chat dialog agent handles greetings and introductory questions without side effects, giving preview and execute modes identical behavior.Source: Chit-Chat Agent · p. 6
Publication Query Agent
- The publication query agent retrieves employee and conference information to autofill applications and support managerial decisions through conversation.Source: Publication Query Agent · p. 6
Business Process Task Execution Agent
- The task execution agent submits, approves, rejects, or returns applications using intent recognition and selectively invoked directory, paper, pricing, and process-execution skills.Source: Business Process Task Execution Agent · p. 6
Travel Estimation Agent
- The travel estimation agent uses the Amadeus API to estimate flights and hotels so managers can verify requested amounts.Source: Travel Estimation Agent · p. 6
Visualization Agent
- The visualization agent converts retrieved data into plots when available, otherwise returning a default no-data response; TravelBot also creates datastore alerts through NLQ and background monitoring.Source: Visualization Agent · p. 6, Alerting using Natural Language Query · p. 7
Use Case – 2: Loan Application
The loan officer chatbot reuses the multi-agent framework to support application queries, business-rule insights, visualization, alerts, and document analysis.
Use Case – 2: Loan Application
- The loan chatbot includes data query, business rules, visualization, alerts, and content-analysis agents for loan applications.Source: Use Case – 2: Loan Application · p. 7
- Its sample interactions are presented through Table 2 and Figure 5, including multimodal responses.Source: Use Case – 2: Loan Application · p. 7, Figure 5 · p. 7
Challenges
The framework addresses deployment challenges by minimizing integration effort while identifying scalability, agent overlap, multi-user, and access-control constraints that require further design.
Coding Overhead
- Minimal coding overhead is supported by integrating independently developed agents through shared input-output contracts, while the 3S orchestrator tolerates noisy confidence estimates.Source: Coding Overhead · p. 7
Scalability
- Naively adding agents can degrade performance when the orchestrator selects agents incorrectly, motivating learning from prior interactions and system changes.Source: Scalability · pp. 7–8
Agent Overlap
- Overlapping agent functionality and knowledge can create response conflicts, so the orchestrator should model overlap and use learning-based selection.Source: Agent Overlap · p. 8
Multi-user Support
- Supporting personalized assistants requires multiple-user sessions and learning from shared interaction data while preserving privacy and security.Source: Multi-user Support · p. 8
Access Control
- Access restrictions are enforced inside agents, which must decline requests from users lacking the required permissions.Source: Access Control · p. 8
Conclusion
The paper presents a multi-agent conversational framework that combines natural-language querying, task execution, alerts, and visualization with minimal integration overhead. It reports preliminary results while identifying scalability and agent overlap as remaining challenges.
- The framework supports querying data, executing business-process tasks, alerting users to changes, and visualizing data through natural-language interaction.Source: Conclusion · p. 8
- An orchestrator combines independently developed skills with minimal overhead to deploy more capable conversational assistants.Source: Conclusion · p. 8
- Preliminary results motivate follow-on research, while scalability and agent overlap remain unresolved challenges.Source: Conclusion · p. 8