Academic Theory and Design Logic

Positions RAIDT as an Information Systems design-science and mid-range theory contribution for governing generative AI in organisational work. This star explains the design logic behind treating the run as the unit of governance, and shows how run-level evidence, pillar scoring, and governance interventions turn responsible AI principles into inspectable organisational practice.

This star sits in the academic and conceptual layer of the RAIDT project. Its role is to explain why RAIDT is not only a practical framework, but also a defensible theory-led contribution to Information Systems governance. It connects foundational concepts to the project's artefacts, mechanisms, and expected outcomes. In project terms, it supports:

• foundations and methodological positioning in Paper 08;
• testable propositions and field validation logic in Paper 09;
• policy translation and standards alignment in Paper 10;
• the design of sector playbooks, evidence packs, scoring logic, and governance interventions.

• What does the academic theory contribution of RAIDT actually consist of?
• Why should RAIDT be understood as design science rather than only as a checklist or framework?
• In what sense is RAIDT a mechanism-based mid-range theory for Information Systems governance?
• Why does RAIDT treat the run, rather than the model alone, as the unit of analysis and governance?
• What problem does this solve in responsible AI, organisational uncertainty, and GenAI risk management?
• What evidence would show that RAIDT works in practice?
• How does this star connect to the run-level evidence pack?
• How does it connect to the five RAIDT pillars: Responsibility, Auditability, Interpretability, Dependability, and Traceability?
• How does this star help supervisors understand the novelty, rigour, and scope of the overall project?

• 10–20 min — How should RAIDT's academic contribution be framed: as a design-science artefact, a governance framework, a mid-range theory, or a combination of all three?
• 20–40 min — What governance problem becomes visible only when the run, rather than the model in abstraction, is treated as the key unit of analysis?
• 40–60 min — Which mechanisms plausibly connect run-level evidence capture to better organisational oversight, contestability, and intervention?
• 60–75 min — What would count as convincing empirical validation that RAIDT improves governance practice rather than merely documenting it?

• S7.01 · Design science research
• S7.02 · Mid-range design theory
• S7.03 · Mechanism-based explanation
• S7.04 · Constructs
• S7.05 · Artefacts
• S7.06 · Mechanisms
• S7.07 · Outcomes
• S7.08 · Boundary conditions
• S7.09 · Socio-technical systems
• S7.10 · IS governance theory
• S7.11 · Audit lineage
• S7.12 · Technical foundation

RAIDT addresses a problem that is increasingly central to organisational use of generative AI: organisations are expected to govern systems whose behaviour is probabilistic, context-sensitive, and operationally shaped by many small design choices. A model card, a policy statement, or a high-level risk category may say something important about an AI system, but they do not adequately explain what happened in a particular use of that system at a particular moment. In practice, organisational risk often emerges not from the model in the abstract, but from the configured run: who used it, for what task, with which prompt, retrieval context, tools, settings, safeguards, and checks. RAIDT makes that operational insight the centre of its theory and design logic.

In academic terms, this positions RAIDT within Information Systems rather than within machine learning theory alone. Information Systems research is concerned with how technologies are configured, embedded, governed, and used within organisational settings. That is important because generative AI in organisations is never just a model. It is a socio-technical arrangement involving users, workflows, access controls, prompts, retrieved documents, tool connections, escalation routes, and accountability structures. RAIDT therefore treats governance as a problem of organisationally situated system use, not only of model performance.

The design-science contribution of RAIDT lies in the purposeful creation of governance artefacts intended to solve a recognised class of problems. The problem class is clear: organisations need a practical way to inspect, compare, justify, and improve individual uses of generative AI under uncertainty. RAIDT responds by producing two core artefacts. The first is a run-level evidence pack, which records the relevant elements of a specific run, including instructions, model and tool configuration, retrieved context where applicable, output, and human or automated checks. The second is a five-pillar RAIDT score profile, which converts the collected evidence into an interpretable governance view across Responsibility, Auditability, Interpretability, Dependability, and Traceability. Together, these artefacts support intervention rather than passive documentation.

The theory contribution is stronger when RAIDT is framed as a mid-range design theory. It is not a universal theory of all AI governance, nor is it merely a local case description. Instead, it proposes a bounded explanation of how certain governance outcomes can be improved under identifiable conditions. Its core constructs include the run, the evidence pack, the pillar scores, governance interventions, actors, tasks, controls, and boundary conditions. Its mechanisms explain how evidence capture and structured review change organisational capability. For example, making prompts, retrieval sources, model settings, LoRA or PEFT adapters, tool use, and checks visible creates audit lineage. Audit lineage then supports review, contestability, and targeted correction. Similarly, converting evidence into a standardised score profile allows cross-run comparison and highlights where uncertainty remains unacceptable.

This mechanism-based account matters because many responsible AI discussions remain at the level of principles. Principles such as accountability, transparency, or robustness are valuable, but they often become weak in operational settings because organisations cannot easily connect them to a specific decision episode. RAIDT translates those principles into governable objects. Responsibility becomes clearer when a run records user role, approval route, and intended task. Auditability improves when there is a stable evidence pack. Interpretability becomes more practical when the basis of the output can be examined through prompt structure, retrieved context, and stated checks, even if the underlying model remains only partially interpretable. Dependability is addressed by documenting checks, repeatability conditions, and failure patterns. Traceability is strengthened through audit lineage across input, configuration, output, review, and intervention.

The run-level focus also helps explain why RAIDT is relevant to AI and uncertainty, including managerial uncertainty. Managers are often not deciding whether a model is perfect. They are deciding whether a particular use is acceptable, contestable, and governable in context. A procurement manager may tolerate lower interpretability for low-risk drafting support but require stronger traceability and human sign-off for contract risk analysis. A public-sector team using retrieval-augmented generation may accept a narrow task if the retrieved documents are logged and the review process is explicit. In both cases, RAIDT provides a way to govern uncertainty by evidencing the concrete run and aligning responses to the risk of the task.

This also explains the relevance of technical foundations such as prompt engineering, RAG, PEFT or LoRA, and RLHF or alignment controls. These are not peripheral technical details. They are part of what shapes the behaviour of a run. A retrieval pipeline changes what contextual evidence is brought into the system. A domain adapter may improve task fit but introduce new versioning and assurance questions. Alignment controls may reduce certain harmful outputs while obscuring how edge cases are handled. RAIDT's design logic does not require a full causal model of internal weights. Instead, it requires enough run-level evidence to make governance judgements inspectable and actionable.

For the project as a whole, this star helps justify why RAIDT belongs in Information Systems governance scholarship. It contributes a theory of governable AI use in organisations, not only a technical method. It shows how the framework links concepts, artefacts, and outcomes: from socio-technical conditions, to run configuration, to evidence capture, to score formation, to governance action. That chain is especially important for Paper 08, where the project must explain its conceptual foundations and methodological pathway. It is equally important for Paper 09, because the theory implies empirical questions that can be tested: do evidence packs improve review quality, do pillar scores improve intervention prioritisation, and do sectoral contexts change which mechanisms matter most? It also supports Paper 10 by showing how policy frameworks such as the EU AI Act, ISO/IEC 42001, and the NIST AI RMF can be operationalised at run level rather than treated as external compliance labels.

The limitations of the theory should be stated clearly. RAIDT does not claim to solve all questions of AI safety, fairness, or legal compliance. It does not eliminate uncertainty, and it does not make opaque models fully transparent. It is best understood as a governance theory and artefact set for organisational use episodes. Its value depends on disciplined evidence capture, reasonable scoring criteria, and institutional willingness to act on the findings. Even so, that is a meaningful contribution: RAIDT explains how responsible AI governance can move from abstract aspiration to auditable practice, one run at a time.

Q1. What is the academic theory contribution of RAIDT?

Answer:
RAIDT contributes a design-science, mechanism-based mid-range theory for governing generative AI use in organisations. Its theoretical novelty lies in shifting attention from the model as a static object to the run as a situated governance unit. This matters because organisational risk, accountability, and assurance arise from specific uses of GenAI, not from model capability statements alone.

Practical example:
A university uses the same model for marketing copy, student support triage drafts, and research administration summaries. The model is constant, but the governance requirements differ by run.

Link to RAIDT:
RAIDT explains why each run should produce an evidence pack and a five-pillar score profile so that governance is tied to actual use, not abstract model claims.

Q2. Why is design science the right framing?

Answer:
Design science is appropriate because RAIDT creates purposeful artefacts to solve a real organisational problem. It does not only describe governance; it designs tools for governance, namely the run-level evidence pack, score profile, and linked intervention logic.

Practical example:
A public body needs a standard way to review GenAI use across departments. RAIDT provides a repeatable structure for documenting runs and comparing them.

Link to RAIDT:
The evidence pack and scoring model are the designed artefacts through which RAIDT produces practical value and researchable outcomes.

Q3. What makes RAIDT a mid-range theory rather than a grand theory?

Answer:
RAIDT explains a bounded class of phenomena: how run-level evidence and structured scoring support better governance of generative AI in organisational settings. It does not attempt to explain all technology adoption, all AI behaviour, or all institutional regulation.

Practical example:
RAIDT can explain why documenting retrieval sources and reviewer checks improves contestability in a legal drafting workflow. It does not claim to explain every aspect of legal organisational change.

Link to RAIDT:
This boundedness makes the framework testable in Paper 09 and adaptable in sector playbooks without overstating its scope.

Q4. What problem does the run-level focus solve?

Answer:
The run-level focus solves the governance gap between high-level AI principles and concrete operational decisions. Without a run-based unit, organisations struggle to determine what exactly happened, who was accountable, what context shaped the output, and what should be changed after a failure or near miss.

Practical example:
A claims analyst receives a flawed GenAI summary. If the organisation only knows the model name, it cannot diagnose the issue. If it records the prompt, retrieved files, settings, output, and checks, the failure becomes investigable.

Link to RAIDT:
The run-level evidence pack is the core mechanism through which RAIDT makes failure analysis, scoring, and governance intervention possible.

Q5. What are the main mechanisms proposed by RAIDT?

Answer:
RAIDT proposes mechanisms of visibility, comparability, contestability, and intervention. Evidence capture makes the run visible. Standardised scoring makes runs comparable. Audit lineage and documented checks make outputs contestable. Governance routes then turn those insights into intervention, such as escalation, redesign, retraining, tighter access controls, or prohibition.

Practical example:
Two departments use RAG differently. One logs retrieved documents and reviewer actions; the other does not. The first can explain and contest outputs; the second cannot.

Link to RAIDT:
These mechanisms map directly onto the evidence pack, pillar scoring, and governance interventions that define RAIDT's practical architecture.

Q6. How do the five pillars function in the theory?

Answer:
The pillars are not only reporting categories; they are governance dimensions that operationalise what a good run should demonstrate. Responsibility asks who owns and approves the use. Auditability asks whether the run can be reviewed. Interpretability asks whether there is sufficient explanatory basis for the use context. Dependability asks whether the run is reliable enough for the task. Traceability asks whether the history and lineage of the run can be reconstructed.

Practical example:
A finance drafting run may score well on traceability and auditability because inputs and reviews are logged, but poorly on dependability if outputs vary too much under similar conditions.

Link to RAIDT:
The five pillars are the scoring lens through which run evidence is transformed into a governance judgement.

Q7. Why does this count as Information Systems governance rather than only AI ethics?

Answer:
RAIDT examines how technology, organisational process, human roles, and control structures interact in real work settings. That makes it an Information Systems governance contribution. AI ethics provides normative direction, but RAIDT focuses on how governance is instantiated in systems, workflows, and accountability arrangements.

Practical example:
In a procurement workflow, the issue is not only whether fairness matters in principle. The issue is how prompts, document retrieval, reviewer sign-off, and audit logs are structured in the information system.

Link to RAIDT:
RAIDT connects responsible AI values to operational IS governance through evidence capture, scoring, and intervention design.

Q8. What evidence would show that RAIDT works?

Answer:
Useful evidence would show that RAIDT improves review quality, speeds root-cause analysis, supports clearer accountability, and helps organisations choose proportionate interventions. It may also show improved policy alignment or reduced ambiguity in supervision and audit.

Practical example:
In a pilot study, reviewers using RAIDT might detect missing approval steps or unstable RAG inputs more consistently than reviewers relying on informal notes.

Link to RAIDT:
Paper 09 can test these effects by comparing governance quality before and after adoption of the evidence pack and score profile.

Q9. How does this star connect to standards and policy?

Answer:
This star provides the theory for translating broad policy requirements into run-level operational evidence. Standards and laws often specify what organisations should be able to demonstrate, but they do not always specify how to evidence a particular use episode.

Practical example:
An organisation seeking alignment with ISO/IEC 42001 or the NIST AI RMF can use RAIDT to record controls, review actions, and residual uncertainties for specific high-stakes runs.

Link to RAIDT:
Paper 10 can use this design logic to show how RAIDT becomes a policy pathway, especially when mapping run evidence to control expectations and sector guidance.

Q10. What are the main boundary conditions of the theory?

Answer:
RAIDT is most suitable where organisations use generative AI in identifiable tasks and can capture meaningful evidence about each run. It is less suitable where runs are impossible to reconstruct, where governance authority is absent, or where the primary concern is low-level model internals beyond available evidence.

Practical example:
RAIDT is useful for assisted drafting, triage support, and summarisation workflows with logs and reviewers. It is less sufficient by itself for highly autonomous systems with weak trace capture.

Link to RAIDT:
Boundary conditions help define where scoring is defensible, where additional controls are needed, and where sector playbooks should narrow or extend the framework.

• A local authority uses a GenAI assistant with RAG to draft case summaries. RAIDT captures the retrieved policy documents, the prompt template, reviewer identity, and any escalation. This makes the summary contestable if a citizen challenges the basis of the recommendation.
• A hospital administration team uses a PEFT-adapted model to summarise routine correspondence. RAIDT records the adapter version, the input class, the quality checks, and the approved use boundary. This helps distinguish operational convenience from clinically sensitive decision support.
• A procurement department uses GenAI to flag contractual risks. RAIDT evidences prompt instructions, model version, tool calls, human legal review, and score rationale. Low dependability may trigger mandatory second review rather than prohibition.
• A university research office uses GenAI to draft ethics application guidance. RAIDT logs which internal guidance documents were retrieved, whether the answer was checked by staff, and how the run scored on responsibility and traceability before wider deployment.
• A regulated financial services team uses aligned models for customer communication drafts. RAIDT helps show that RLHF or safety tuning alone is not sufficient; the organisation still needs run-level evidence, reviewer controls, and intervention triggers.

• Run identifier, date, time, business unit, and task category.

• User role, reviewer role, approver, and accountability owner.

• Prompt or instruction text, prompt template version, and any system prompt constraints.

• Model name, model version, provider, configuration settings, and safety or alignment controls.

• Tool configuration, API/tool calls, and external system dependencies.

• RAG status, retrieved documents, retrieval method, knowledge source version, and retrieval timestamp.

• PEFT or LoRA adapter status, adapter version, and intended domain fit where relevant.

• Input type, output produced, and any output transformation or post-processing.

• Human checks, automated checks, pass/fail criteria, and escalation actions.

• Evidence for each RAIDT pillar, score rationale, residual uncertainty, and confidence note.

• Exceptions, incidents, user feedback, override records, and governance intervention taken.

• Policy or standards mapping where relevant, including internal control references.

• Paper 08: foundations and methodological pathways
  This star provides the academic justification for the project by defining RAIDT as a design-science, mechanism-based mid-range theory contribution rooted in Information Systems governance and socio-technical analysis.

• Paper 09: empirical validation
  This star identifies what should be tested empirically: whether run-level evidence improves auditability, whether pillar scoring supports better decisions, and whether mechanisms differ across sectors, risks, and task types.

• Paper 10: policy pathways
  This star explains how abstract obligations from the EU AI Act, ISO/IEC 42001, and the NIST AI RMF can be translated into operational governance through run evidence, traceability, and intervention logic.

• Sector playbooks
  The theory here supports sector-specific guidance by showing which constructs remain stable across sectors and which boundary conditions, thresholds, and intervention routes must be adapted locally.

• RAIDT scoring
  This star gives the conceptual rationale for the five-pillar score profile as a governance device rather than a cosmetic rating exercise.

• RAIDT evidence pack
  This star explains why the evidence pack is central to the framework: it is the mechanism that makes a run inspectable, reviewable, and empirically analysable.

• RAIDT governance interventions
  This star clarifies that the endpoint of RAIDT is action. Scores and evidence should lead to approval, redesign, restriction, escalation, retraining, or refusal where appropriate.

• Design science research in Information Systems.
• Mid-range theory and design theory literature.
• Mechanism-based explanation in organisational and IS research.
• Socio-technical systems and IS governance theory.
• Responsible AI and organisational accountability literature.
• AI uncertainty, managerial decision-making, and risk governance.
• Auditability, traceability, and contestability in AI governance.
• Prompt engineering, RAG, PEFT or LoRA, and alignment control literature as technical grounding.
• Standards and policy sources: EU AI Act, ISO/IEC 42001, NIST AI RMF.
• Empirical studies of AI assurance, documentation, and human oversight in organisational use.

• RAIDT does not claim to explain all of AI governance or all forms of organisational technology governance.
• It does not make opaque models fully transparent at the level of internal parameters or representations.
• It does not replace legal advice, regulatory interpretation, or domain-specific professional judgement.
• It does not guarantee safe or fair outcomes simply because evidence is captured.
• Its usefulness depends on evidence quality, disciplined process adoption, and genuine organisational willingness to intervene.
• Some highly autonomous, weakly logged, or externally controlled systems may require additional assurance methods beyond RAIDT.
• Scores should support judgement, not substitute for judgement.

This star explains the academic core of RAIDT. The argument is that RAIDT is not just a practical checklist for responsible AI, but a design-science contribution to Information Systems governance. Its novelty is to treat the run as the unit of governance. By run, I mean one configured use of a generative AI system for a specific task, at a specific time, in a specific context. That matters because most governance failures do not arise from the model in abstraction; they arise from a concrete use episode shaped by prompts, retrieval context, tools, settings, and checks.

The theory contribution is mid-range and mechanism-based. RAIDT proposes that if organisations capture structured run evidence and convert it into a five-pillar score profile, they improve visibility, auditability, contestability, and intervention capacity. The framework therefore links abstract responsible AI principles to operational governance action. This helps position Paper 08 conceptually, gives Paper 09 testable propositions, and supports Paper 10 by showing how standards and policy expectations can be translated into run-level evidence and governance practice.