Research journeys · Record 002
An Evidence-First Reference Model for Trustworthy AI in Business Information Systems
How can trustworthy AI be described so that it remains traceable inside a decision-support system? This published BIS 2026 chapter develops an evidence-first reference model for that question.
- Publication
- BIS 2026 · LNBIP 584 · pp. 89–101
- Published
- 21 May 2026
- Context
- Evidence-first reference model
The question behind it
Trustworthiness should remain traceable in the system.
The chapter starts from a persistent gap: in many organisations, trustworthiness is handled through principles, documentation or post-hoc explanations rather than through the design of AI-supported decision systems.
It develops a domain-agnostic evidence-first reference model. A focused synthesis of recent literature is consolidated with a conceptual implementation architecture, without prescribing a particular technology stack.
Understand the modelFive requirements for the reference model
The published chapter’s abstract names five requirements. They are summarised here as orientation, not as an independent validation.
- 01
Evidence binding
Evidence stays connected to the claim it supports.
- 02
Artifact-mandatory evaluation
Evaluation is tied to the artefacts it requires.
- 03
Audit-safe traceability
Traceability remains available for review and audit.
- 04
Effective human oversight
Human oversight remains an effective part of the process.
- 05
Separated execution
Governance and analytical execution remain distinct.
What the paper contributes
A compact reference frame.
The published chapter consolidates the five requirements into a domain-agnostic reference model and a conceptual implementation architecture. The architecture demonstrates that the requirements are constructible in principle, without prescribing a particular tool stack.
This is a reference-model contribution. It does not establish field effectiveness, independent implementation validation or the effectiveness of today’s AIJIM SHARK.
Source: published chapter at SpringerWhere to go from here
A step between diagnosis and specification.
The reference model is one published step in Torsten’s research journey: it translates requirements into a structured model class. The later specification work is a separate, non-peer-reviewed preprint with its own evidence and status.
- Explore the specification paper
The current arXiv preprint, with its own status and source trail.
- Research on TechJournal
The other works and their respective boundaries.
Read the original
The source behind this record.
This record is based on the published Springer conference chapter. Its bibliographic identity and interpretation remain bound to that original source.
- 01
Published conference chapter
An Evidence-First Reference Model for Trustworthy AI in Business Information Systems
Torsten Olivi Tiltack · Yifei Dong · Kun Yu · Fang Chen · BIS 2026 · LNBIP 584 · pp. 89–101
Online since 21 May 2026 · DOI 10.1007/978-3-032-26363-6_7
Publication at Springer
Cite this work
The bibliographic record is available as BibTeX and RIS for reference managers.
Take another look
A model remains bound to its evidence.
Later additions and a personal reflection will have a visible place here. Until then, the boundary between conceptual constructability and empirical validation remains explicit.
- Which requirements can be reproduced in independent environments?
- What evidence would support a credible field study?
- Which elements remain deliberately model- or profile-bound?
Springer metadata and abstract linked as the original source. Reference model, constructability and open validation boundaries kept distinct.
The next question remains open.
More research records connect original sources, context and each work’s own status.
Explore the sources