Large Language Models (LLMs) have demonstrated impressive conversational capabilities, yet their susceptibility to hallucinations and inconsistent recommendations poses significant risks in high-stakes domains such as finance. This paper presents an interactive chatbot for loan application guidance that leverages a case-based reasoning (CBR) approach to generate actionable counterfactual explanations within an agentic framework. Our system employs a supervisor agent, built using the LangGraph framework, to orchestrate four specialised agents: a classifier agent that provides an initial loan prediction, a causally-aware counterfactual explanation agent that proposes minimal yet feasible modifications to reverse an unfavourable decision, a Feature Actionability Taxonomy (FAT) agent that updates user-specific immutability constraints based on feedback, and a template-based natural language generation (NLG) agent that transforms counterfactual suggestions into clear, user-friendly explanations. A key strength of our design is the automated feedback loop: when users indicate that certain suggestions are unworkable, the FAT agent revises the constraints and instructs the counterfactual generation agent to produce a refined explanation. We detail the system architecture and workflow and outline an experimental plan that compares our full agentic chatbot to ablated variants and a LLM-Only Baseline. And finally we outline a planned user study to evaluate how controlled reasoning afects trust in high-stakes lending.
Large language models (LLMs) such as GPT-4 have demonstrated remarkable conversational abilities,
making them attractive for use as automated assistants in decision-making domains. However, in
highstakes applications like financial loan assessments, the unreliability of LLM outputs, especially their
tendency to generate hallucinations, i.e., plausible-sounding but incorrect or unfounded information,
poses a significant risk [
Explainable AI (XAI) is crucial in lending and other domains to justify algorithmic decisions and
provide users with recourse options [
Despite their usefulness, counterfactual explanations in practice face challenges regarding feasibility
and personalisation [
In this paper, we address these challenges by combining case-based reasoning (CBR) with LLM capabilities to create a conversational agent for loan application guidance. CBR, an approach where reasoning is grounded in specific instances or “cases” naturally complements LLMs by providing concrete examples or analogies that the generative model can use to formulate explanations. By retrieving and adapting similar past cases (or generating counterfactual cases), a CBR component can supply factual anchors that reduce the risk of hallucination from the LLM. Meanwhile, the LLM enables a flexible dialogue with the user, clarifying their context and preferences and presenting explanations in fluent natural language.
We present a novel interactive conversational AI system that implements this CBR-LLM synergy in the context of loan application decisions. In this synergy, CBR anchors explanations in concrete cases while the LLM contextualises those cases into conversational, user-specific advice which using the strengths of both symbolic retrieval and generative language. Our system is built with an agentic workflow using a framework called LangGraph, 1 which allows the LLM to act as a controller orchestrating multiple specialised modules. The supervisor loan agent integrates four key agents: (1) an agent with trained loan approval classifier that provides an initial prediction, (2) a causally-aware instance-based counterfactual explanation discovery agent that finds how the unsuccessful input case could change to yield a positive outcome, (3) a Feature Actionability Taxonomy (FAT) agent that encodes user-specific immutability and ethical constraints on feature changes, and (4) a template-based natural language generation (NLG) agent that guide the LLM to convert explanations in a proper format considering fairness and actionability then putting it into user-friendly dialogue responses. By decomposing the task among these modules, we use the strengths of each (e.g., reliable numeric computation and causal logic in the structured modules, and language expressiveness in the LLM) while mitigating the weaknesses of an unconstrained LLM.
In summary, our contributions are as follows: • We introduce a novel interactive CBR-enhanced LLM agentic framework that ensures modularity: the supervisor agent interacts with dedicated agents for prediction, explanation, updating user constraints and NLG, improving the handling of numerical computations, counterfactual generation, and ethical constraints compared to a monolithic LLM approach. • We integrate a Feature Actionability Taxonomy into the explanation process, enabling the system to respect user-specific immutable features as well as generating personalised feasible, respectful suggestions. • We demonstrate how an LLM can serve as a dialogue orchestrator to parsing user intent, invoking specialised tools via LangGraph, and tailoring the final wording, thereby contributing reasoning, tool-use, and adaptive NLG beyond simple template filling • We outline an evaluation strategy comparing our full system against a LLM-Only Baseline and an ablated version, to quantify the benefits in terms of explanation quality, user trust, and the reduction of hallucinated or infeasible recommendations.
To our knowledge, this is one of the first works to systematically combine case-based reasoning with large language models in an agentic setting, aiming to enhance explainability and reliability in high-stakes decision support. Next, we discuss related literature before detailing the methodology of our system, experimental setup, followed by our planned user study, and conclusions.
Conversational XAI and Interactive CBR. Researchers have increasingly recognised the need for
making AI explanations more conversational and interactive. Traditional explanation interfaces (e.g.,
static texts or visualisations) do not allow users to seek clarification or explore “what-if” scenarios in
depth. Conversational XAI attempts to fill this gap by using dialogue to deliver and refine explanations.
For instance, Wijekoon et al. [
Counterfactual explanations and actionable recourse. Counterfactual explanations have become
a cornerstone of interpretable machine learning [
LLM augmentation and hallucination mitigation. The advent of powerful LLMs has led to
explorations of how they can be integrated with existing AI models or knowledge sources to improve
reliability. A key concern is the phenomenon of hallucination in generative models, where the model
outputs incorrect information with high confidence [
For our interactive chatbot, we developed a workflow using the LangGraph framework, which allows us to explicitly define the sequence and branching of actions the agent (driven by an LLM) should perform. The architecture is illustrated in Figure 1. At a high level, the system involves a supervisor agent (which is powered by an LLM) interacting with external modules through a structured graph of operations. This design ensures that the supervisor agent consults the right agent or tool at the right time, rather than relying on the LLM to do everything in one prompt.
At the heart of this system lies a large language model (LLM) supervisor, responsible for more than just template polishing. It acts as (i) a conversational router that understands free-form user utterances and maps them onto the appropriate branch of the LangGraph, (ii) an orchestrator that conditionally calls external tools (classifier, counterfactual module, FAT checker, template NLG) and stitches their outputs together, and (iii) an adaptive explainer that revises template text to match the user’s tone, education level, and follow-up questions while strictly preserving the factual content returned by the downstream agents. These three capabilities are essential: without them the workflow could neither decide which agent to invoke next nor maintain a coherent, context-aware dialogue.
The interaction begins with the user either providing their loan application details or asking a
question about their application status. The supervisor agent orchestrates the following steps:
1. Classification: The agent first calls a classifier Agent, passing in the user’s application features
(e.g., income, loan amount, credit score, etc.). This agent has a classifier as a tool (a neural network
with 4 linear layers) that has been trained on German Credit dataset contains 32,581 instances
with 11 features (9 mutable) to predict approval or rejection of the loan status, 2. It returns a
prediction (approved/rejected).
2. Counterfactual Generation: If the loan is predicted to be rejected (or if the user asks “How
can I get approved?” or “How can I improve my loan application?”), then the supervisor agent
triggers the counterfactual explanation agent. This agent uses an instance-based approach to
identify one or more plausible modifications to the user’s input that would result in an approval
(see algorithm 1). In our case-based approach, the module searches a database of past successful
applications for a case similar to the user’s to find the nearest decision boundary crossing [
This agent takes the refined set of counterfactual recommendations and constructs a conversational explanation. It uses predefined sentence templates that incorporate the user’s data, the model’s decision, and the suggested changes. The template choice and phrasing are informed by the FAT categories for each feature to ensure the explanation is polite, understandable, and appropriately hedged (especially for sensitive factors). The LLM may also be used at this stage in a lfil-in-the-blank manner to smooth out the text or adjust it to the user’s tone (e.g., formal vs. informal), without altering the factual content. 5. User Interaction: The compiled explanation is presented to the user by supervisor agent. The conversation can continue: the user might ask follow-up questions (e.g., “Why is my credit score considered low?” or “I can’t change X, what else can I do?”). The supervisor agent handles these by invoking diferent branches of the agentic workflow as needed. For instance, if the user asks “I cannot decrease my loan amount by $5,000”, then FAT agent will update Feature Actionability Taxonomy for that feature in the user’s profile and loop back to the counterfactual generation agent to provide another solution tailored to user’s specific situation.
Algorithm 1 Iterative Adaptation for Counterfactual Generation (adapted from [
foreach feature in priority order do if ∈/ and ′ ̸= then ′ ← if Classifier (′) = approved then
return ′ return FAIL
This agentic loop continues until the user’s queries are resolved. By explicitly encoding the workflow graph, we ensure that at each turn the supervisor agent knows which agent outputs to incorporate, maintaining consistency and factuality across the dialogue. One benefit of this modularisation is that the LLM never has to internally compute or assume the outcome of changes so that it always queries the actual model, thereby eliminating a potential source of error or hallucination. Similarly, knowledge about what is changeable is codified in FAT rather than left to the LLM’s general knowledge (which might be incomplete or biased regarding what a user can or should change).
Our use of LangGraph diferentiates from simpler pipeline approaches by allowing conditional branching and iterative flows. For example, based on whether the model outcome is positive or negative, diferent branches in the graph are followed; based on whether a user-proposed change is feasible or not, the agent can decide how to respond (perhaps by invoking the FAT logic again or by politely explaining limitations). The graph structure thus ofers flexibility in the conversation flow beyond a fixed sequence, which is crucial for interactive dialogue. This modular architecture means improvements or updates to one component (say, a better classifier or a more nuanced FAT categorisation) can be integrated without retraining the entire system.
In our agentic framework, every arrow in the figure 1 represents a directed communication act which is an intentional message exchange that drives the system through its reasoning, acting, and interacting phases. Each specialised agent, including the supervisor, is configured with its own system prompt. These prompts guide the agents on how to interact with one another in a cost-eficient and purposedriven manner. For instance, when the supervisor agent receives a user’s request, it issues a directive to the classifier agent. The classifier, informed by its system prompt, responds with a brief result (e.g., “accepted” or “rejected”) to eficiently update the supervisor’s internal belief state.
This structured communication mirrors the principles of speech act theory, where the expressive act
(such as a request or assertion) not only transmits information but also performs an action. Here, the
supervisor’s directive (a “request” act) and the classifier’s corresponding response (an “inform” act) are
essential steps in the internal virtuous cycle [
Together, these communication acts are guided by predefined system prompts to ensure that the agentic LLM operates as a cohesive, eficient, and adaptive system.
The counterfactual explanation agent addresses the question “What minimal changes would flip the
decision to a positive outcome?” by integrating instance-based search with causal reasoning. In a
highlevel process, the agent first performs a nearest unlike neighbor (NUN) search in the training dataset.
This search uses a weighted feature distance where the weights are computed from two components: (1)
Individual Treatment Efect (ITE) values that obtained via DECI which is a causal discovery framework
[
After selecting the nearest unlike neighbour, the agent generates a counterfactual explanation by iteratively adapting the query instance. For each prioritised feature, the agent updates its value to match that of the selected instance; importantly, any modification to a feature also triggers a recursive adaptation of its causally dependent (child) features as defined by the DAG. This parent/child adaptation ensures that the proposed changes maintain the inherent causal structure, thereby generating counterfactuals that are both causally efective and practically actionable.
Once a valid counterfactual is obtained (i.e., one that reverses the model’s decision to the desired outcome) the counterfactual explanation agent communicates this result back to the supervisor agent, completing its role within the overall agentic workflow.
The FAT agent is responsible for ensuring that counterfactual explanations are personalised and practically actionable by updating the user’s feature constraints based on their feedback. Conceptually, the FAT agent analyses counterfactual suggestions and identifies if any proposed changes involve features that the user deems immutable (e.g. the loan Amount, normally adjustable by applicants, becomes immutable if the user specifically needs exactly $5,000 to cover a planned expense). When such features are detected, the FAT agent revises its internal FAT profile to reflect these constraints.
Once the FAT agent updates this information, it first communicates the revised FAT details back to the supervisor agent. The supervisor agent, acting as the central coordinator, then relays the updated FAT guidelines to the counterfactual generation agent, instructing it to generate a new counterfactual explanation that respects the user’s specified immutability. This dynamic exchange among the FAT agent, supervisor agent, and counterfactual generation agent ensures that the final counterfactual recommendations are both causally informed and tailored to the user’s real-world capabilities. If, under these constraints, no valid counterfactual can be found, the system will explicitly inform the user: “Given the constraints, no feasible changes can overturn the decision.”
In our agentic framework the LLM co-authors the final explanation; it fills the slots of a safe template
and then dynamically re-voices the text (e.g., adjusting formality, adding clarifications requested in
the previous turn), something a rigid template engine alone cannot do [
For example, a final output might look like:
Your loan application was declined due to several factors. To improve your chances: Take steps to increase your monthly income to at least $5,000 (currently $4,200).
Reduce your requested loan amount closer to $10,000 (currently $15,000).
Unfortunately, you cannot change your age (25), and younger applicants often have shorter credit histories.
These changes would address the risk factors identified by the model. I hope this helps, good luck with your next application! After generating its explanation, the NLG agent returns its output to the supervisor agent. The supervisor agent then refines the explanation as needed, enhancing its naturalness and realism, before presenting the final, user-tailored message. This chain of communication ensures that every counterfactual recommendation is accurately translated into clear, actionable, and contextually appropriate advice.
The proposed agentic workflow forms a cohesive, feedback-driven system in which each specialised agent contributes to delivering personalised and actionable explanations. Initially, the classifier agent predicts the loan decision. If the loan is rejected or if a user requests guidance, the counterfactual explanation agent identifies the nearest unlike neighbour and adapts its features to generate a counterfactual explanation for the query.
If a participant indicates that a suggested change is unworkable (e.g., stating “I cannot change X”), the FAT agent is activated. In response, the FAT agent updates its internal profile to mark the specified feature as immutable and promptly communicates these updated constraints to the supervisor agent. The supervisor agent then relays the revised FAT information to the counterfactual explanation agent, which regenerates a personalised counterfactual that adheres to the updated constraints. This revised explanation is forwarded to the template-based NLG agent, which transforms it into a clear and structured natural language message. Finally, after further refinement by the supervisor agent to ensure naturalness and realism, the final explanation is presented to the user.
This iterative process, whereby the system continuously refines its counterfactual based on user feedback via the FAT agent, is central to ensuring that the provided explanations are both causally consistent and tailored to the user’s real-world constraints. In the next section, we describe the experimental setup and planned user study that will evaluate the impact of our integrated agentic framework on explanation quality and user trust.
In order to assess the efectiveness of the proposed system, we outline experiments focusing on two aspects: (1) the quality and feasibility of the explanations, and (2) the impact on user understanding and trust.
We compare three versions of the chatbot: • Full Agentic System: the complete system as described, using the classifier, counterfactual reasoning, FAT, and template NLG (the CBR-LLM agent). • No-FAT Variant: an ablation where the FAT module is disabled. The agent still generates counterfactuals, but it cannot update user’s specific constraints to provide a better counterfactual tailored to the user. • LLM-Only Baseline: a baseline using a state-of-the-art LLM (e.g., GPT-4) with only classifier tool. This baseline is prompted to perform the entire task on its own. For example, we provide it with the user’s input and ask it to give an explanation or advice. The prompt can be engineered with few-shot examples to encourage counterfactual-style outputs. However, the LLM will not have access to the FAT agent or counterfactual agent to generate responses and it relies only on its general knowledge and reasoning.
In order to assess the efectiveness of our proposed system, we will conduct a user study focusing on two main aspects: the quality and feasibility of the counterfactual explanations, and the overall impact on user understanding and trust. In this future study, participants will engage in an interactive chat environment implemented using Streamlit, 3, where the supervisor agent coordinates among internal agents, namely, the classifier agent, the counterfactual generation agent, the FAT agent, and the natural language generation (NLG) agent.
The central focus of our evaluation will be the performance of the automated feedback loop. Specifically, the FAT agent will update the user’s constraints based on their feedback regarding unworkable suggestions, and these updated constraints will prompt the counterfactual generation agent to generate refined and more actionable counterfactual explanations. This iterative process, which continually updates the explanation based on user input, is expected to significantly enhance the personalisation and practical utility of the system’s advice.
Following this interactive phase, participants will complete a detailed survey measuring dimensions
such as overall satisfaction, clarity, helpfulness, fairness, and trustworthiness of the provided
explanations. This comprehensive evaluation approach, informed by recent best practices in user-centred XAI
research [
Participants will be recruited from local university networks and professional online communities. We will use a screening process to select individuals who can realistically assume the role of a loan applicant and have an elementary understanding of financial decision-making. The required sample size will be determined using G*Power analysis, targeting a power of 0.80, an alpha level of 0.05, and an anticipated medium efect size (Cohen’s d = 0.5), which suggests a minimum of 34 participants. To account for potential dropouts and to enhance the robustness of our findings, approximately 40 participants will be recruited. All participants will be provided with an information sheet and will give informed consent prior to their involvement.
• Scenario: You are John Doe, a 36-year-old applicant whose loan request for purchasing a new car has been rejected. Your task is to interact with the intelligent assistant and explore ways to improve your loan application.
In this study, participants interact with each of the three system variants described earlier for up to ifve turns (one user utterance plus one system reply constitutes a turn). For each of the three chatbot variants, we collect the complete five-turn dialogue, five per-variant Likert-scale ratings, and optional free-form feedback, and analyse these data jointly to assess how specific dialogue features (e.g., invoked modules) influence user evaluations. We adopt a within-subject Latin-square counter-balancing: the order of the three variants (3! = 6 possible sequences) is randomly assigned, and the same decision scenario is reused across variants to keep task dificulty constant. The examples of these chat interactions and UI screenshots will be provided in Appendix I.
During these interactions, the supervisor agent orchestrates responses to both general queries and specific requests for guidance on improving a loan application. For general questions, the assistant provides prompt, context-aware answers; however, when a participant seeks concrete solutions or strategies to improve their rejected loan application, the counterfactual generation agent is activated. This agent first evaluates the current scenario (i.e., a loan rejection) and generates a counterfactual suggestion that details minimal modifications to the input features.
This generated counterfactual is then communicated to the supervisor agent, which passes it on to the template-based NLG agent. The NLG agent transforms the counterfactual suggestion into a clear and structured explanation using predefined templates. If the resulting explanation includes any suggestions that the participant finds immutable, for example, due to personal constraints, the FAT agent is triggered. The FAT agent updates the user’s feature constraints (e.g., marking features as non-negotiable) and communicates these updated FAT details back to the supervisor agent. The supervisor agent subsequently instructs the counterfactual generation agent to generate a revised counterfactual explanation that adheres to the updated constraints.
After further refinement by the supervisor to enhance naturalness and realism, the final, personalised explanation is returned to the participant. This coordinated multi-agent interaction ensures that the counterfactual explanations are not only feasible and actionable but also tailored to the user’s realworld circumstances. To safeguard data quality, we embed an instructed-response attention check (“Select Disagree”), a recall question on the chatbot’s last suggestion, and a 5 minutes interaction-time requirement per variant; any participant failing two of these criteria is removed a priori.
After the interaction phase, they will complete a detailed survey designed to capture their perceptions of the system’s explanations.
The survey will include Likert-scale items measuring:
• Overall Satisfaction: How satisfied the participant is with the overall assistance provided by
the chatbot.
• Clarity: The degree to which the counterfactual explanations are clear and understandable.
• Helpfulness: The extent to which the suggestions aid the participant in understanding what
changes could improve their loan application.
• Fairness: The perceived fairness and ethical appropriateness of the recommendations.
• Trustworthiness: The degree to which the participant trusts the assistant’s advice [
For example, an item measuring clarity might ask, “How clear and understandable were the counterfactual explanations provided by the assistant?” with response options ranging from 1 (not clear at all) to 5 (extremely clear). Similar items will be framed for the other dimensions.
In addition to these Likert-scale measures, the survey will include open-ended questions where participants outline the specific steps they would take based on the suggestions they received. This qualitative feedback will ofer further insight into the actionability and practical utility of the explanations.
Because each participant uses all three variants, we will compare their mean Likert ratings with paired -tests for the three variant pairs (full agentic vs. no-FAT, full agentic vs. LLM-only, and no-FAT vs. LLM-only). We will first check the normality of the rating diferences with the Shapiro-Wilk test; if this assumption is violated, we will use the Wilcoxon signed-rank test instead. The GPower sample-size estimate (medium efect = 0.5, = 0.05, power = 0.80) was calculated for this paired-test design.
We introduced an agentic framework for loan-application guidance that combines case-based reasoning with large language models. By orchestrating specialised agents including a classifier, a causally aware counterfactual generator, a Feature Actionability Taxonomy (FAT) checker, and a template-based NLG module. The system is designed to generate explanations that respect the user’s real-world constraints and causal dependencies.
Although empirical validation is still to come, the modular architecture is intended to reduce common LLM issues such as hallucination and to support a feedback loop that should improve clarity and trust. These hypotheses will be tested in a forthcoming user study.
We posit that this agentic-CBR approach can provide personalised, actionable recourse in high-stakes ifnance scenarios and, once validated, could generalise to other domains where trustworthy decision support is required.
During preparation of this work, the authors used ChatGPT for the purpose of: grammar and spelling check, paraphrase and reword. After using this tool, the authors reviewed and edited the content and take full responsibility for the publication’s content.
Figures 2 and 3 present two screenshots of the instructions that will be provided to participants at the beginning of the user study. These images depict the comprehensive guidelines for the study scenario, including task descriptions, interaction guidelines, and post-interaction procedures.