S1.10 - High-impact_decisions

S1.10 ? High-impact decisions

flowchart LR
    A[Consequential GenAI use
rights services safety records compliance] --> B[RAIDT
Run-level evidence framework]
    B --> C[[High-impact decisions
Threshold for stronger governance]]
    C --> D[Run-level evidence pack]
    C --> E[Five-pillar score profile]
    C --> F[Reviewer reconstruction]
    D --> G[Reviewability and contestability]
    E --> H[Governance readiness]
    F --> I[Organisational learning]
    J[Healthcare finance education public services law enterprise workflows] --> C

? Star S1 - Origins, Background and History

Star context: Explains why RAIDT emerged where responsible AI concerns, managerial uncertainty, governance obligations, audit traditions and GenAI deployment pressure meet situations in which outputs can shape consequential organisational action.

Academic picture

Definition / background

High-impact decisions are decisions, or decision-support processes, in which a GenAI output can materially affect a person's rights, opportunities, treatment, access, obligations, safety, or the integrity of important organisational records and actions. The concept matters because not all GenAI use is governance-critical. A casual brainstorming exchange and a run that helps determine eligibility, triage, compliance action, disciplinary response, or professional advice do not carry the same consequences.

In RAIDT, this item helps distinguish ordinary low-consequence use from contexts where evidence quality must rise. The concept therefore functions as a threshold concept for proportionate governance. It does not mean that every important use must be banned or fully automated under heavy compliance machinery; rather, it means that when consequences are meaningful, organisations need clearer run reconstruction, stronger justification, and better evidence about what was configured, produced, reviewed, and acted upon.

This differs from broad responsible AI language that often speaks about risk in general terms. RAIDT makes the issue operational by treating the run as the unit of governance. A run that contributes to a high-impact decision should generate a richer run-level evidence pack and should be assessed more carefully through the five-pillar score profile. In that sense, high-impact decisions belong inside RAIDT because they help specify when the framework becomes especially necessary.

The item is also conceptually linked to managerial uncertainty and audit traditions. Where consequences are serious, managers cannot rely on generic policy statements alone. They need evidence that a particular GenAI use was appropriate for its purpose, intelligible enough to review, dependable enough to rely on, and traceable enough to contest or improve later.

Why this concept matters

This concept matters because organisations often misunderstand where governance effort should be concentrated. If all GenAI use is treated as equally risky, governance becomes impractical and is likely to be ignored. If high-impact use is treated as just another productivity aid, organisations create exposure without adequate evidence, oversight, or contestability.

High-impact decisions solve a scoping problem inside RAIDT. They help identify which runs deserve deeper scrutiny, stronger documentation, and more robust review. This avoids the confusion between general AI use and consequential AI-supported organisational action.

If this concept is missing, an organisation may deploy GenAI into workflows that influence benefits, compliance, admissions, case handling, financial judgement, or staff decisions without being able to explain how outputs were generated, what constraints existed, or whether a reviewer could reconstruct what happened. That weakens accountability precisely where accountability is most needed.

Key idea: High-impact decisions matter in RAIDT because they indicate when GenAI use must move from convenience-based trust to evidence-based governance at the level of the individual run.

What this item explains

It explains when a GenAI-supported activity becomes consequential enough to require stronger governance.
It explains why RAIDT is risk-proportionate rather than universally heavy-handed.
It explains how decision significance affects evidence expectations for a run.
It explains why run-level documentation is especially important when outputs may influence rights, services, safety, or compliance.
It explains how a practical governance threshold can connect ethical concern to operational review.

Practical example / likely audience question

Audience question

Is RAIDT needed for every casual chat?

Answer

No. The concern behind the question is that governance frameworks can become overextended and make everyday experimentation unworkable. RAIDT addresses this by being proportionate. It does not assume that every prompt to a general-purpose model requires the same level of documentation or review.

The direct answer is that RAIDT becomes most valuable where reviewability matters and where a run can influence an important record, judgement, service outcome, or decision with meaningful consequences. A casual ideation chat may still benefit from good practice, but it does not usually require the same evidential depth as a run that supports a student misconduct review, a benefits eligibility decision, a financial risk summary, or a clinical administrative recommendation.

RAIDT handles this better than a generic AI governance approach because it does not stop at broad policy categories. It asks what happened in this run, under this configuration, for this task, in this context, and with what downstream consequence. That makes it possible to scale governance effort according to impact rather than applying either too little or too much control.

Practical example in RAIDT terms

Consider a public service team using GenAI to draft case summaries for housing-support applications. The model does not make the final decision, but its summary shapes which cases are escalated quickly, which applicants appear high priority, and which supporting facts receive attention.

The run-level issue is that a single summary may materially influence a high-impact judgement about access to assistance. The evidence needed would include the prompt, relevant input sources, configuration settings, date and time of the run, reviewer identity, the generated output, edits made by staff, and the rationale for accepting or rejecting the recommendation. The most affected RAIDT pillars are Responsibility, Auditability, and Traceability, with Interpretability and Dependability also important because reviewers need to understand and trust the basis of the summary. Treating this as a high-impact decision improves governance readiness because the organisation can reconstruct the run, challenge the output if needed, and show that consequential use was not left to undocumented discretion.

Detailed link to RAIDT

High-impact decisions links to RAIDT in four ways.

First, it connects to RAIDT's core idea that governance should attach to real uses of GenAI rather than abstract principles alone.
Second, it sharpens the importance of the run, because consequence depends on the specific task, context, timing, configuration, and downstream use of one run.
Third, it strengthens the need for the evidence pack and the score profile, since consequential runs need richer documentation and more defensible scoring across the five pillars.
Fourth, it supports reviewability, contestability, audit readiness, and organisational learning by ensuring that influential runs can later be examined, challenged, and improved.

High-impact decisions ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

In this chain, the item acts as a governance trigger. It indicates when RAIDT should move from general good practice to stronger evidential discipline.

Link to the five RAIDT pillars

Responsibility

High-impact decisions intensify the need to show who authorised the use, who reviewed the output, and who remained accountable for the outcome.