Source-linked AI summary

Beyond IVR: Benchmarking Customer Support LLM Agents for Business-Adherence

Sumanth Balaji, Piyush Mishra, Aashraya Sachdeva, Suraj Agrawal

arXiv:2601.00596v1cs.CL

TL;DR

JourneyBench addresses the challenge of evaluating whether customer-support agents follow business policies and procedural requirements. It represents SOPs as graphs, generates diverse user journeys, and measures adherence with UJCS; across 703 conversations in three domains, structured workflow orchestration outperformed prompt-based approaches, with smaller models exceeding larger ones in policy compliance.

  • Problem

    Evaluating whether customer-support agents follow business policies and procedural requirements remains a core challenge, including adherence to required steps, validations, and exception handling.

  • Method

    JourneyBench represents SOPs as directed acyclic graphs, generates diverse user journeys through those workflows, and uses the User Journey Coverage Score to measure adherence to required action sequences.

  • Results

    Across 703 conversations in three domains, Dynamic-Prompt-Agent orchestration significantly outperformed Static-Prompt-Agent approaches, enabling smaller models to exceed larger ones in policy compliance.

  • Takeaways & Limitations

    The findings demonstrate the value of structured workflow orchestration for policy-aware customer-support agents beyond rigid IVR systems.

  • Takeaways & Limitations

    The Dynamic-Prompt-Agent depends on precise business-logic modeling, while simulation-based evaluation may not capture all nuances of real-world user behavior.

Abstract

from arXiv · show

Traditional customer support systems, such as Interactive Voice Response (IVR), rely on rigid scripts and lack the flexibility required for handling complex, policy-driven tasks. While large language model (LLM) agents offer a promising alternative, evaluating their ability to act in accordance with business rules and real-world support workflows remains an open challenge. Existing benchmarks primarily focus on tool usage or task completion, overlooking an agent's capacity to adhere to multi-step policies, navigate task dependencies, and remain robust to unpredictable user or environment behavior. In this work, we introduce JourneyBench, a benchmark designed to assess policy-aware agents in customer support. JourneyBench leverages graph representations to generate diverse, realistic support scenarios and proposes the User Journey Coverage Score, a novel metric to measure policy adherence. We evaluate multiple state-of-the-art LLMs using two agent designs: a Static-Prompt Agent (SPA) and a Dynamic-Prompt Agent (DPA) that explicitly models policy control. Across 703 conversations in three domains, we show that DPA significantly boosts policy adherence, even allowing smaller models like GPT-4o-mini to outperform more capable ones like GPT-4o. Our findings demonstrate the importance of structured orchestration and establish JourneyBench as a critical resource to advance AI-driven customer support beyond IVR-era limitations.

1 Introduction

JourneyBench addresses the gap between customer-support agents’ conversational flexibility and their need to follow multi-step business policies. It introduces graph-based policy evaluation and shows the value of structured workflow control.

  • Traditional IVR systems enforce compliance through rigid decision trees but often provide inflexible, frustrating user experiences.
  • Existing benchmarks assess tool selection and state transitions but inadequately evaluate complete task sequences with complex inter-task dependencies.
  • JourneyBench represents standard operating procedures as graphs to generate diverse customer-support scenarios involving branching logic, missing inputs, and tool failures.
  • The User Journey Coverage Score measures whether agents follow the required sequence of SOP-defined actions.
  • A Dynamic-Prompt Agent guided by workflow structure performs more reliably than a Static-Prompt Agent, highlighting the value of structured control in business settings.

2 JourneyBench Framework

JourneyBench models business workflows as graph-structured SOPs and derives reproducible conversational scenarios from their paths. Its framework combines explicit workflow representations, simulated interactions, and varied conditions for testing policy adherence.

  • JourneyBench comprises SOP graphs, task nodes, user journeys, and scenarios that test agent robustness under varied conditions.
  • SOPs are modeled as directed acyclic graphs whose nodes represent tasks and whose edges encode valid transitions, decision points, and policy constraints.
  • Each node specifies a natural-language task description, available tools, parameters, and conditional pathways that determine the next node at runtime.
  • Validated SOP graphs generate user journeys through Breadth-First Search, after which multiple scenarios vary inputs and tool behavior while preserving expected outcomes.
  • 703 datapoints average 10.91 turns and 3.34 tool calls per conversation across a 41-tool benchmark.

3 Evaluation Metrics

The evaluation measures strict adherence to expected tool-call journeys and parameter execution through the User Journey Coverage Score (UJCS).

  • Tool Trace Alignment compares predicted and expected tool-call sequences, treating missing, extra, or misordered calls as SOP violations.
  • For each tool call, the score accounts for the number of correctly supplied parameters relative to the expected parameters.
  • The expected trace length is L = |Texp|, and conversation-level performance is represented by Sconv.
  • UJCS aggregates adherence across N simulated conversations for an SOP graph.

4 Experimentation

The experiments compare static prompting with dynamic workflow-state control across controlled scenarios, robustness disturbances, and production deployment. Results favor DPA for policy adherence and robustness, including for a smaller model.

  • Agent designs: DPA processes one SOP node at a time and updates orchestration after each tool execution, whereas SPA places the entire SOP in one static prompt.
  • Agent designs: The study excludes explicit planning approaches because their latency is considered unsuitable for real-time customer support, using a custom runtime-state framework instead.
  • Experimental setup: Experiments use a 40-turn limit and a GPT-4o simulated user following predefined journeys without information leakage.
  • Experimental setup: JourneyBench generates failed-function and missing-parameter scenarios from correct-context journeys to test robustness.
  • Deployment: Production DPA-based systems handle 6,000+ daily calls while applying the evaluated text-based workflow logic to voice interactions.
  • Results: GPT-4o with DPA achieves a UJCS of 0.717 versus 0.564 for SPA, while DPA maintains stable coverage under failed functions and missing parameters.
  • Results: GPT-4o-mini with DPA scores 0.649, exceeding GPT-4o with SPA at 0.564 and showing that structured orchestration can benefit smaller models.
  • Error analysis: Error analysis identifies dependency violations, parameter hallucination, and user-simulator failures as distinct error classes.

5 Related Work

Prior benchmarks advance tool-use and stateful execution evaluation but generally do not measure fidelity to complex, prescribed journeys or robustness to disturbances. JourneyBench targets these gaps while decoupling workflow evaluation from tool implementations.

  • Tau Bench, Tau2-Bench, ToolSandBox, and AppWorld evaluate multi-turn tool use, dual-control environments, stateful execution, or world-state tracking.
  • These frameworks primarily measure tool selection or state changes rather than fidelity to prescribed multi-step journeys with complex dependencies.
  • JourneyBench additionally evaluates policy adherence, task dependencies, and robustness to conversational disturbances and dynamic tool failures.
  • JourneyBench separates workflow evaluation from tool internals by treating tools as modular components with defined interfaces and evolving implementations.

6 Conclusion

JourneyBench evaluates policy-aware customer-support agents with graph-based workflows and UJCS, finding that structured DPA orchestration outperforms SPA across broad testing and deployment settings.

  • JourneyBench addresses the need to combine conversational flexibility with strict policy adherence beyond rigid IVR systems.
  • Across 703 conversations in three domains, DPA significantly outperforms SPA and enables smaller models to exceed larger ones in policy compliance.
  • DPA-based production systems reliably handle 6,000+ daily customer interactions while meeting real-time and policy-adherence requirements.

7 Limitations and Future Work

The framework’s effectiveness is bounded by the quality of business-logic modeling, simulation realism, and evaluation cost. Future work targets automated graph construction, richer dependencies, and cheaper evaluation.

  • DPA success depends on precise business-logic modeling, which is difficult in dynamic or poorly documented fields.
  • Simulation-based evaluation may not capture all nuances of real-world user behavior.
  • $388.88 in evaluation cost constrained the range of tested models.
  • Future research could explore semi-automated graph generation, more cost-effective evaluation, and complex dependency structures.

8 Ethical Considerations

JourneyBench’s synthetic workflows and conversations introduce risks related to bias and evaluation validity, while automation may affect workforce composition. The paper recommends expert validation, human assessment, real-world bias checks, and responsible transition planning.

  • Synthetic workflows and conversations may inherit LLM biases, motivating domain-expert validation and real-world bias checks.
  • LLM-generated evaluation data can introduce circularity, so JourneyBench should be paired with human assessment and production-behavior alignment.
  • The paper suggests customer-support automation may shift work toward higher-complexity tasks rather than directly displacing staff.
  • Organizations should plan responsible workforce transitions and training as customer-support automation is deployed.

G.1 Missing Parameters

The missing-parameter examples contrast an SPA that proceeds without required customer information with a DPA that preserves verification requirements and halts when necessary inputs or tools are unavailable.

  • The SPA proceeds to other tasks after the user lacks a Customer ID, rather than preserving the verification dependency.
  • The DPA states that verification cannot proceed without the Customer ID because it is essential for account verification.
  • When an internal server issue prevents verification, the DPA says it cannot complete the verification at that moment.
  • After the verification issue, the user redirects the conversation toward checking a Downtown service outage.
  • The SPA responds that an outage affects Downtown and offers further assistance.

H Hallucinations in Parameter Values

The example shows parameter hallucination: the agent uses a tool-description example value instead of the user-provided credit score. Both agent types exhibit this error.

  • Both Dynamic-Prompt-Agent and Static-Prompt-Agent examples are presented in the parameter-hallucination discussion.
  • The agent should use the user-provided credit score of 720 but instead uses the example value 700 from the parameter description.The resulting analysis reports a score of 700 as good.
  • The parameter definition identifies creditScore as an integer extracted from the report when the credit report is available.
  • In the Dynamic-Prompt-Agent conversation, the user supplies a credit score of 720 before the tool analyzes the value as 700.

I User Input Hallucination Example

The examples illustrate two user-agent failures: a simulated user supplies information absent from its seed, and another conversation ends before completing the prescribed journey.

  • User Input Hallucination Example: The simulated user provides a payment method even though no payment method was included in the user seed.The conversation therefore should not have proceeded.
  • User Input Hallucination Example: The agent asks the user to choose among Credit Card, PayPal, and Bank Transfer, after which the user answers Credit Card.
  • User Input Hallucination Example: The full journey requires Service Calls Analysis after Billing Info Retrieval, but the agent ends the conversation after retrieving billing information.
  • K Human Validation Protocol: Human validation reviews graphs for logical structure, coherence, and complexity, accepting only graphs that all five annotators unanimously pass.

L Realism Validation Summary

Synthetic conversations are evaluated against production-oriented quality criteria and achieve averages comparable to production QA bands. The examples also show that Dynamic-Prompt-Agent can handle a user correction within a workflow node.

  • L Realism Validation Summary: The realism rubric evaluates conversational proficiency through questions about tone, empathy, friendliness, concision, and clarity.
  • L Realism Validation Summary: Goal Attainment evaluates whether the agent identifies the caller’s intent and resolves or properly escalates the primary request.
  • L Realism Validation Summary: Synthetic conversations achieve 84.37% overall, with 82.33% Conversational Proficiency and 87.78% Goal Attainment.Scores average binary rubric responses across conversations.
  • L Realism Validation Summary: Grouping dependent tools within a node lets the Dynamic-Prompt-Agent incorporate a user correction without restarting the process.The example updates the income category from Salaried to Self-Employed and recollects the income details.
Loading 2601.00596v1…