Explainable AI framework maps every step of the machine learning workflow

Artificial intelligence is reshaping science and industry, yet many users still regard its models as opaque black boxes. A new open-access paper in Artificial Intelligence Review argues the problem runs deeper than individual hard-to-read predictions. Conventional explainable AI methods clarify single outputs but overlook the upstream decisions and downstream quality checks that determine whether an insight can be trusted at all. The authors, a team led by George Paterakis with Andrea Castellani, George Papoutsoglou, Tobias Rodemann and Ioannis Tsamardinos, respond with a framework they call Holistic Explainable Artificial Intelligence, or HXAI.

HXAI embeds explanation into every stage of the data-analysis workflow rather than treating it as a bolt-on feature applied after a model has produced an answer. The framework unifies six components into a single taxonomy: the data, the analysis set-up, the learning process, the model output, the model quality and the communication channel. Each component is aligned with the needs of three audiences: domain experts, data analysts and data scientists. A clinician, a statistician and a machine learning engineer do not require the same explanation, and a framework that ignores those differences leaves most users underserved.

To make the taxonomy operational, the team assembled a question bank of 112 items covering the needs of these user groups across the workflow. The question bank functions as a measuring instrument: any existing explainable AI toolchain can be checked against it to see which questions it answers and which it leaves open. Using this instrument, the authors surveyed contemporary tools and report that the analysis highlights critical coverage gaps. Even widely used explanation techniques leave substantial portions of the workflow unexplained when judged against what users need to know.

The framework is not built only from engineering considerations. The authors ground HXAI in theories of human explanation, principles from human–computer interaction and findings from empirical user studies. From this base, HXAI identifies the characteristics that make explanations clear, actionable and cognitively manageable. The taxonomy is intended to reduce terminological ambiguity in a field where terms such as interpretability, explainability and transparency are often used loosely, and to enable rigorous coverage analysis of existing toolchains rather than impressionistic comparisons.

One forward-looking contribution concerns AI agents that embed large language models. The authors demonstrate how such agents can orchestrate diverse explanation techniques, translating technical artifacts into stakeholder-specific narratives. The aim is to bridge the gap between AI developers and domain experts: instead of handing a domain expert a saliency map and expecting them to interpret it, an agent could assemble explanations from across the workflow and present them in language suited to that expert’s background and task.

The authors are explicit that this is not a conventional survey. Departing from traditional surveys or perspective articles, the work melds concepts from multiple disciplines, lessons from real-world projects and a critical synthesis of the literature to advance what they describe as a novel, end-to-end viewpoint on transparency, trustworthiness and responsible AI deployment. The end-to-end framing is the core claim: trust in an AI result depends on choices made before any model is trained, on how the learning process is conducted, on how output quality is verified, and on how the result is communicated, not merely on whether a single prediction can be rationalised after the fact.

The practical implication is that organisations deploying AI should evaluate their explanation practices across the full workflow rather than at the model-output stage alone. A tool that explains individual predictions well may still leave users unable to judge whether the training data were appropriate, whether the analysis set-up was sound, or whether model quality was adequately checked. The 112-item question bank offers a concrete way to run such an audit, mapping each question to the workflow component and user group it serves. The paper is published open access, and the version was shared early to provide faster access to peer-reviewed, accepted research, subject to further edits before the final Version of Record.

Subject of Research: A user-centric framework for explainable AI that spans the full machine learning workflow

Article Title: A comprehensive perspective on explainable AI across the machine learning workflow

Article References: Paterakis, G., Castellani, A., Papoutsoglou, G., Rodemann, T., & Tsamardinos, I. (2026). A comprehensive perspective on explainable AI across the machine learning workflow. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11718-z

Image Credits: AI Generated

DOI: 10.1007/s10462-026-11718-z

Keywords: explainable AI, HXAI, machine learning, taxonomy, large language models, AI agents, transparency, trustworthiness, human-computer interaction, data analysis workflow, responsible AI, question bank