• workshop_dense_100#slide 26

Within RAIDT, uncertainty means more than generic risk. It refers to the practical condition in which organisational users must act despite incomplete knowledge, ambiguous inputs, conflicting signals, or irreducible variability in the environment. The managerial paper aligns this with two forms of uncertainty that matter for AI-supported work: epistemic uncertainty, where the system lacks sufficient knowledge or training coverage, and aleatoric uncertainty, where randomness or noise cannot be eliminated. In managerial settings, both forms matter because deterministic-looking outputs can hide weak evidence and encourage misplaced confidence.

RAIDT treats this problem as a governance issue, not only a modelling issue. Because the run as the unit of governance is where prompts, retrieved context, tools, policies, and oversight meet, uncertainty must be captured in the run-level evidence pack rather than left as an implicit background condition. This is why the framework evaluates a run through the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and expresses the result in a score profile with anchors 1=missing / 3=partial / 5=audit-ready. In practice, uncertainty matters because it determines whether a recommendation should be relied upon, challenged, escalated, or withheld. It also explains why influence methods as governance interventions matter: a retrieval layer may improve Traceability, structured prompting may improve uncertainty communication, and alignment controls may strengthen Responsibility, but only if the evidence for those effects is captured and reviewable.

Consider a public-service eligibility assistant used by a caseworker assessing whether a claimant qualifies for emergency housing support. The situation is uncertain because the application may contain missing documents, inconsistent statements, and time-sensitive facts, while policy clauses can change or be interpreted narrowly. RAIDT would require the run-level evidence pack to preserve the exact prompt, the retrieved policy text and version, the generated recommendation, any uncertainty or limitation statement, and the caseworker?s final decision.

This matters because a seemingly confident answer may rest on partial evidence. A stronger configuration might use retrieval and structured prompting to cite the exact clause, separate facts from assumptions, and flag when escalation is needed. The resulting score profile makes it easier to see whether the run is merely partial or genuinely audit-ready across the five pillars.

• 01-Responsible_AI_for_Managerial_Decision-Making_Under_Uncertainty-V3
• 11-RAIDT_Academic_Logic_M_v11