S7.02 - Mid-range_design_theory

S7.02 ? Mid-range design theory

flowchart LR
    A[Abstract principles / grand theory overreach / checklist reductionism] --> B[RAIDT - run-level evidence framework]
    B --> C[[Mid-range design theory - bounded design logic for GenAI runs]]
    H[Healthcare / finance / education / public services / enterprise work] --> C
    C --> D[Run-level evidence pack]
    C --> E[Five-pillar score profile]
    C --> F[Reviewability and contestability]
    D --> G[Governance readiness and organisational learning]
    E --> G
    F --> G

? Star S7 - Academic Theory and Design Logic

Star context: Positions RAIDT within Academic Theory and Design Logic as a bounded, design-science, mechanism-based contribution that explains how run-level evidence can support responsible organisational governance of generative AI.

Academic picture

Definition / background

A mid-range design theory is a bounded explanatory and prescriptive theory that addresses a defined class of practical problems without claiming universal scope. In the RAIDT context, it explains how organisations can govern generative AI more effectively when the unit of analysis is the run rather than the policy statement, model family, or abstract principle. It therefore sits between two unhelpful extremes: a grand theory that is too broad to guide operational governance, and a checklist that is too shallow to explain why particular controls or evidence matter.

Conceptually, the idea draws on the long-standing distinction between theories of broad social order and theories designed for a delimited domain of recurring problems. In design-oriented information systems work, this type of theory is especially useful because it links constructs, mechanisms, artefacts, boundary conditions, and outcomes. RAIDT follows that logic. It does not simply say that responsible AI is desirable; it specifies how governance can be made inspectable through run-level evidence, evidence packs, and a structured five-pillar profile.

This matters in generative AI governance because organisational use is highly contextual. The same model may be acceptable in one task and risky in another depending on prompts, data sensitivity, user role, oversight arrangements, and downstream consequences. A mid-range design theory is therefore appropriate because it is capable of explaining patterned variation across runs without pretending that one universal rule will fit every use case.

Within RAIDT, mid-range design theory belongs at the core of the project because it gives conceptual coherence to the framework's practical outputs. The run-level evidence pack is not just documentation; it is an artefact justified by the theory. The score profile across Responsibility, Auditability, Interpretability, Dependability, and Traceability is not just a dashboard; it is a structured way of representing governance conditions and quality at the level where use actually occurs. In that sense, mid-range design theory is the bridge between RAIDT's academic positioning and its operational usefulness.

Why this concept matters

Mid-range design theory solves a persistent problem in AI governance: many frameworks can state values, but fewer can explain how those values should be translated into inspectable organisational practice. Without a theory at the right level, governance becomes either rhetorical or fragmented. RAIDT uses mid-range design theory to avoid both failures.

The concept also prevents a common confusion between policy aspiration and governance capability. Organisations often believe they are governing AI because they have principles, committees, or approval forms. Yet these devices do not necessarily tell a reviewer what happened in a specific use of a system, what evidence supports that account, or how the quality of that run can be assessed. Mid-range design theory matters because it makes that transition explicit: from broad claims about responsible AI to repeatable run-level governance.

If this concept is missing, the risk is that RAIDT could be mistaken either for a high-level normative manifesto or for a technical compliance checklist. It is neither. Its contribution is a bounded theory of how evidence, review, contestability, and learning can be organised around the run. That is especially important for organisations using generative AI in volatile settings where risks, contexts, and oversight needs change rapidly.

Key idea: Mid-range design theory matters because it gives RAIDT the right level of abstraction to turn responsible AI principles into operational, evidence-based governance of real organisational runs.

What this item explains

Why RAIDT is intentionally bounded to a defined class of organisational generative AI governance problems rather than presented as a universal theory.
How run-level evidence becomes the central explanatory and design anchor for responsible governance.
Why evidence packs and five-pillar score profiles are theoretically grounded artefacts rather than ad hoc reporting tools.
How mechanisms, boundary conditions, and outcomes can be related without collapsing into either abstraction or checklist thinking.
Why RAIDT can travel across sectors while remaining focused on the governance of configured uses of GenAI in context.
How the framework supports reviewability, contestability, audit readiness, and organisational learning through a coherent design logic.

Practical example / likely audience question

Audience question

Why is RAIDT framed as a mid-range design theory rather than as a grand theory of AI governance or simply a practical governance checklist?

Answer

The concern behind the question is usually about legitimacy and scope. A grand theory can appear intellectually ambitious, while a checklist can appear immediately practical. RAIDT rejects both extremes because neither is well suited to governing actual generative AI use in organisations. A grand theory would overclaim by suggesting that one framework can explain all forms of AI governance across all institutional settings. A checklist would underexplain by listing controls without clarifying why they matter, how they interact, and under what conditions they are meaningful.

The direct answer is that RAIDT is framed as a mid-range design theory because its purpose is to explain and support a bounded class of governance situations: organisational GenAI runs. It offers a structured account of the constructs involved, the mechanisms through which better governance is achieved, the artefacts that capture evidence, and the outcomes that matter for responsible use. That gives the framework enough conceptual depth for academic scrutiny and enough practical specificity for implementation.

A practical example helps. Suppose an organisation uses a generative AI tool to draft a briefing for a sensitive procurement decision. A generic AI governance checklist may ask whether the system was approved, whether a human was involved, and whether privacy was considered. RAIDT goes further by focusing on the actual run: which model was used, with what prompt and context, on what data basis, under which role permissions, with what review steps, and with what documented evidence. The framework can then produce an evidence pack and a score profile that a reviewer can interrogate. That is more robust than a generic governance approach because it ties governance quality to a reconstructable instance of use rather than to general organisational claims.

Practical example in RAIDT terms

Consider a healthcare trust using a generative AI assistant to draft a discharge summary from clinician notes. The GenAI use case appears routine, but the governance challenge emerges at the run level: one discharge summary may involve straightforward administrative text, while another may include clinically sensitive interpretations, incomplete source data, or patient-specific ambiguities.

In RAIDT terms, the relevant issue is not merely whether the organisation has an AI policy, but whether this specific run can be governed and reconstructed. The evidence needed would include the task definition, user role, prompt or instruction set, model and version, source data category, review workflow, clinician sign-off, and any flags relating to uncertainty or correction. Responsibility and Auditability are immediately engaged because the organisation must show who initiated and reviewed the run. Interpretability matters because the clinician must understand what the generated text is doing well enough to judge whether it is acceptable. Dependability matters because the discharge summary must be sufficiently accurate and stable for safe operational use. Traceability matters because the trust may later need to reconstruct how the output was produced and approved.

Mid-range design theory improves governance readiness here because it explains why these evidential elements belong together as part of a bounded governance model. RAIDT does not claim to solve every problem in healthcare AI. It provides a theory and artefact structure for governing this class of run-level organisational use in a way that is reviewable, contestable, and auditable.

Detailed link to RAIDT

Mid-range design theory links to RAIDT in four ways.

First, it gives RAIDT its core intellectual shape as a bounded design-science contribution focused on governing real uses of generative AI in organisational settings.
Second, it justifies the choice of the run as the main unit of analysis, because a mid-range theory needs a stable and inspectable governance unit through which mechanisms and outcomes can be examined.
Third, it underpins the evidence pack and score profile as theory-informed artefacts that represent governance quality in a structured and reviewable form.
Fourth, it connects RAIDT to reviewability, contestability, audit readiness, and organisational learning by explaining how evidence from individual runs can support both immediate oversight and longer-term governance improvement.

Mid-range design theory ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

In this chain, the theory defines what kind of governance problem RAIDT addresses, run-level evidence provides the inspectable basis for analysis, the evidence pack organises that basis into a reviewable artefact, the score profile synthesises governance conditions across the five pillars, and governance readiness becomes the practical organisational outcome.

Link to the five RAIDT pillars

Responsibility

Mid-range design theory strengthens Responsibility by clarifying that governance must be attached to identifiable runs, roles, decisions, and oversight arrangements rather than dispersed across vague organisational commitments.