Counterfactual explanations have become a significant component in eXplainable AI (XAI), ofering intuitive "what if" scenarios. However, typical numeric or tabular outputs can be vague to non-technical audiences. Additionally, many counterfactual methods ignore causal relationships or suggest inactionable changes such as “be younger by five years,” raising concerns over realism and ethics. To address these issues, we propose a holistic approach that integrates a Feature Actionability Taxonomy (FAT) and causal discovery into counterfactual generation, thereby ensuring realistic, ethically sound, and semantically transparent explanations in natural language. We further introduce an interactive, agentic workflow enabling users to iteratively refine constraints. Through extensive user studies, pilot evaluations, and synergy with Case-Based Reasoning (CBR), this approach yields explanations that are accessible, trust-enhancing, and practically useful in domains such as healthcare, ifnance, and education.
(iii) Natural Language Generation (NLG), presenting counterfactuals in user-friendly textual form. Finally, we propose an agentic framework, in which users can iteratively refine or reject certain suggestions, receiving updated counterfactuals each time. This paper outlines our progress towards a more transparent, practical, and interactive XAI, referred to here as Natural-XAI.
Modern decision-support systems increasingly rely on complex, “black-box” models. Trust and adoption highly rely on efective explanation. Counterfactual methods stand out for their simplicity and actionability: rather than simply stating “your loan was denied because of X,” a counterfactual might say “if your monthly income increased by $200, your loan would have been approved.”
However, user trust can degrade if suggested feature changes are not plausible or ethically-sound.
For instance, suggestions like “be ten years younger” or “change your race” are not only inactionable
but also unethical in many contexts. Prior works such as DiCE [
Despite these advancements, few solutions comprehensively unify actionability, causality, and natural language. Our approach bridges these dimensions, while also employing Case-Based Reasoning (CBR), which complements counterfactuals by providing exemplars of similar cases and how they difer from the current instance.
Building on the limitations of purely numeric or tabular counterfactuals, our proposed framework introduces a holistic pipeline that progressively addresses actionability, presentation, causal coherence, and iterative user engagement. We begin by specifying how features can and cannot be changed in the Feature Actionability Taxonomy (FAT). Next, we incorporate these constraints into a template-based Natural Language Generation (NLG) module to communicate potential changes in a user-friendly manner. We then extend the counterfactual generation process to handle causal dependencies so that recommended interventions remain realistic. Finally, we embed all these components into an agentic workflow, allowing iterative dialogue and refinement of constraints. The sections that follow describe these four core pillars in turn.
A foundational element of our approach is the Feature Actionability Taxonomy (FAT), which classifies features into three categories based on how realistically they can be modified: • Directly Mutable (DM): Easily adjustable features such as requested loan amount. • Indirectly Mutable (IM): Features that are changeable but only through more extended or intricate actions (e.g. educational level, occupation type). • Non-sensitive or Sensitive Immutable: Characteristics like age or gender, which typically cannot or should not be changed for ethical or practical reasons.
FAT was defined using a data-driven methodology that relied on examining features extracted from
six datasets [
By explicitly encoding each feature’s mutability, FAT automatically filters out suggestions that violate
real-world or ethical constraints, thereby increasing user trust in the system [
We employ a three-stage, template-based approach to generate user-friendly counterfactual explanations
in natural language [
Figure 2 (referenced in the text) illustrates the full pipeline, from matching features to templates through to sentence sorting and domain-specific epilogues. For example, a finance application might conclude with “Good luck with your loan!” while a healthcare scenario might close with “Stay healthy!” This layered approach ensures that explanations are both contextually accurate and reassuring in tone.
While FAT addresses the actionability of individual features, many real-world scenarios involve intricate
causal dependencies. Changing one feature (e.g. “Education Level”) may, in fact, afect others (e.g.
“Occupation,” “Income”), which in turn could impact the ultimate prediction (e.g. “Loan Approval”). To
capture these relationships, we integrate causal discovery, for instance, the DECI framework [
Once such a causal graph is in place, our counterfactual discovery engine intervenes on chosen features while propagating changes through the causal network, ensuring each recommended scenario respects real cause-and-efect pathways. This stands in contrast to naive methods that treat all features as mutually independent, potentially yielding contradictory suggestions (e.g. recommending both fewer working hours and a higher income). By linking causal modelling with FAT, we ensure that only permissible features are altered, and do so in a way that maintains internal consistency across all features.
Finally, we embed these components in an agentic workflow, transitioning from static explanations to dynamic, user-driven conversations. Our agentic workflow is simultaneously interactive, which means the user can add or relax constraints in natural language as well as being iterative so that the system regenerates counterfactuals until the user is satisfied. Basically, after the system presents a counterfactual recommendations in natural language, the user can refine constraints or discard unfeasible changes by updating Feature Actionability Taxonomy which is embeded in this workflow. Concretely: • Rejecting a Recommendation: Users may specify, “I cannot reduce my monthly outgoings below $1,000 due to fixed costs.” • Prioritising Feasibility: The system re-executes the counterfactual search, guided by both FAT and the user’s updated constraints, and then re-runs the NLG to produce revised textual explanations.
This interactive loop ensures each subsequent iteration of recommendations is increasingly tailored to the user’s personal limitations and priorities. Consequently, Natural Language Counterfactual Explanations evolve from a one-of advisory statement into an iterative conversation, fostering transparency, trust, and practical usability.
Thus far, our work has centred around three key pillars ensuring actionability, incorporating causal insights, and enabling interactive user refinement and has yielded promising preliminary outcomes.
First, we developed a Feature Actionability Taxonomy (FAT) to categorise features as directly mutable, indirectly mutable, or immutable/sensitive. Using insights from a focused user study, we then crafted a template-based NLG module that transforms raw numeric deltas into readable, context-rich sentences (e.g. emphasising time frames and feasibility cues). Early feedback showed that these text-based explanations were perceived as significantly more transparent and actionable than purely tabular representations.
Next, to address the risk of contradictory or unrealistic suggestions, we integrated causal discovery (e.g. via DECI) into the counterfactual search process. This ensures that interventions on one feature (e.g. increasing “Education Level”) properly cascade to related variables (e.g. “JobType” or “Income”), thus reflecting genuine real-world cause-and-efect. Pilot experiments demonstrated a measurable reduction in contradictory recommendations, particularly in financial scenarios, when compared to correlation-only methods.
Finally, we introduced an agentic workflow that treats users as active participants rather than passive recipients of one-shot explanations. After receiving an initial set of textual counterfactuals, users can impose additional constraints (“I cannot reduce my expenses below $1,000” ) or request alternative actions. The system promptly regenerates revised solutions, again expressed via the NLG module. Preliminary user testing suggests that iterative refinement can boosts clarity and trust.
In moving towards truly Natural XAI, our framework unifies a Feature Actionability Taxonomy (FAT), causal integration, and a template-based NLG approach within an agentic workflow. This design ensures that counterfactual explanations remain feasible, ethically grounded, and responsive to user feedback. As the next major step, we plan a user study that deploys this interactive system in finance, healthcare, and education scenarios each featuring its own domain-specific features and constraints. Participants will provide iterative feedback on the system’s recommendations (e.g. indicating infeasible changes) and then observe how the workflow adapts the generated counterfactuals in real time. We will evaluate how iterative refinement influences trust, clarity, and overall satisfaction, testing whether our method significantly outperforms traditional single-shot explanations. Ultimately, we aim to demonstrate that combining actionability, causality, and user engagement not only enhances the transparency of AI decisions but also ofers a meaningful path towards ethically and practically sound recourse. Acknowledgements. The author would like to thank the supervisory team, domain experts who provided guidance on ethical considerations, and participants in pilot studies whose feedback has substantially shaped the direction of this research.
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.