S8.01 - Manual_implementation

S8.01 ? Manual implementation

flowchart LR
    A[Fragmented prompts and informal AI use] --> B[RAIDT - run-level evidence framework]
    A2[Policy aspirations without operational proof] --> B
    B --> C[[Manual implementation]]
    H[Controlled spreadsheets
Reviewer forms
Saved prompts and outputs
Scoring sheets
Timestamps] --> C
    C --> D[Run-level evidence pack]
    C --> E[RAIDT score profile]
    C --> I[Reviewer reconstruction]
    D --> F[Reviewability and contestability]
    E --> G[Governance readiness]
    I --> J[Organisational learning and policy alignment]

? Star S8 - Implementation and Operations

Star context: Shows how RAIDT can be adopted manually, semi-automatically or through orchestration, and how it becomes part of real governance routines. Manual implementation is the entry point that proves RAIDT can operate as a disciplined governance method before dedicated tooling is introduced.

Academic picture

Definition / background

Manual implementation refers to the use of human-operated but structured processes to apply RAIDT without requiring a dedicated software platform. In practice, this means using controlled spreadsheets, saved prompt and output records, reviewer forms, scoring sheets, naming conventions, and simple evidence folders to document each run of a generative AI system. The concept matters because many organisations need a credible starting point for governance before they can justify more advanced infrastructure.

Conceptually, manual implementation sits at the operational end of RAIDT's governance model. RAIDT treats the run as the unit of governance, so the question is not whether a workflow is high-tech or low-tech, but whether each run is captured in a way that supports reviewability, reconstruction, and scoring. Manual implementation therefore differs from ad hoc manual working. A user casually saving a screenshot or copying a prompt into a notebook is not enough. Manual implementation in RAIDT requires structured evidence fields and repeatable handling rules.

This item belongs in RAIDT because it demonstrates that run-level evidence can be produced even when automation is absent. That matters for pilots, supervision workshops, proof-of-concept studies, small teams, and early-stage institutional adoption. It also clarifies the relationship between implementation mode and governance output: the evidence pack and the five-pillar score profile do not depend on sophisticated tooling in principle, but they do depend on disciplined capture of the right evidence at the right level.

Manual implementation is therefore best understood as a transitional but valid governance mode. It enables organisations to establish evidence habits, test reviewer workflows, refine rubric interpretation, and identify where later semi-automation or orchestration would add value. In RAIDT terms, it is not a weaker idea of governance; it is the minimum operational configuration that still takes evidence seriously.

Why this concept matters

Manual implementation solves a practical adoption problem. Organisations often agree that generative AI use should be governed, but they delay implementation because they assume governance only becomes real after dashboards, wrappers, logging pipelines, or enterprise platforms are available. RAIDT avoids that deadlock by showing that governance can begin with disciplined manual processes, provided that runs are documented systematically.

The concept also prevents a common confusion between principles and operations. A team may say that it values accountability, transparency, and oversight, yet still have no reproducible record of how a specific AI-assisted decision or output was produced. Manual implementation closes that gap by creating a tangible route from policy aspiration to evidence capture. It gives teams a manageable first step while preserving the core logic of RAIDT.

If this concept is missing, organisations risk treating governance as a future technical project rather than a present operational responsibility. That creates blind spots around contested outputs, reviewer disagreements, weak audit trails, and inconsistent scoring. Manual implementation matters because it makes governance start now, not after an ideal technical architecture appears.

Key idea: Manual implementation matters because RAIDT can begin as a disciplined evidence practice before it becomes a software-supported governance system.

What this item enables

Adoption of RAIDT in pilots, workshops, early-stage studies, and low-resource organisational settings.
Structured capture of prompts, outputs, reviewer judgements, timestamps, and contextual notes for each run.
Construction of a run-level evidence pack without requiring a bespoke technical platform.
Production of a five-pillar score profile from manually assembled but controlled evidence.
Human review, contestation, and reconstruction of runs when questions arise later.
Comparison of runs across teams or use cases using common templates and scoring sheets.
Identification of where semi-automation or orchestration would reduce burden or improve consistency.

Practical example / likely audience question

Audience question

Can RAIDT start without software, or does manual implementation make the framework too weak for serious governance?

Answer

The concern behind this question is understandable: many people assume that if governance is manual, it is informal, fragile, or impossible to scale. The direct answer is that RAIDT can start manually without losing its core governance value, so long as the evidence fields are captured reliably and the review process is controlled. Manual implementation is not equivalent to improvised note-taking; it is a structured method for documenting runs before platform-level support exists.

A practical example would be a research team trialling a generative AI assistant for drafting internal policy summaries. The team may not yet have a wrapper application or integrated logging layer, but it can still require each run to be recorded in a spreadsheet, save the prompt and output in a versioned folder, use a reviewer form to assess quality and risk, and complete a RAIDT scoring sheet. That creates a usable evidence trail for later inspection.

RAIDT handles this better than a generic AI governance approach because it specifies what needs to be captured at run level and why. A generic policy might say that outputs should be reviewed by a human. RAIDT goes further by making that review evidential: which prompt was used, which output was produced, who reviewed it, what concerns were raised, how the run scored against the five pillars, and what follow-up action was taken.

Practical example in RAIDT terms

Consider a public-sector benefits team using a generative AI tool to draft first-pass response letters to citizens. The use case is administrative productivity, not autonomous decision-making, but each run still affects communication quality, fairness, and accountability. In a manual implementation of RAIDT, the officer records the task context, prompt, model used, date and time, source material, generated draft, and reviewer comments in a controlled evidence template.

The run-level issue is that a disputed or misleading letter may later need to be reconstructed. Without structured capture, the team may know that AI was used but be unable to show how the wording emerged or whether the draft was properly checked. Manual implementation addresses this by preserving the prompt-output pair, the reviewer sign-off, any edits made before issue, and the reasoning used for the RAIDT score.

The evidence needed includes the task description, prompt text, output text, reviewer identity, revision notes, applicable policy constraints, scoring sheet, and any escalation decision. Responsibility is affected because a human owner must remain accountable for the run. Auditability and Traceability are affected because the sequence of actions must be reconstructable. Interpretability matters because reviewers must explain why the output was acceptable or problematic. Dependability matters because repeated use should show stable review practice rather than arbitrary handling. The result is improved governance readiness even though the implementation remains manual.

Detailed link to RAIDT

Manual implementation links to RAIDT in four ways.

First, it links to RAIDT's core idea by showing that governance begins with evidence discipline around a specific run rather than with general statements about AI principles.
Second, it links to the run because manual implementation determines how the inputs, outputs, context, reviewers, and decisions for that run are captured and preserved.
Third, it links to the evidence pack and score profile because a manually assembled record can still support structured scoring across Responsibility, Auditability, Interpretability, Dependability, and Traceability.
Fourth, it links to reviewability, contestability, audit readiness, and organisational learning because the preserved record allows later inspection, challenge, comparison, and improvement.

Manual implementation ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

In other words, manual implementation is the operational bridge that turns RAIDT from a conceptual framework into a reproducible governance routine.

Link to the five RAIDT pillars

Responsibility

Manual implementation strengthens Responsibility by making clear who initiated the run, who reviewed it, and who accepted the final output for use. It prevents the diffusion of accountability that often occurs when AI assistance is treated as an informal background tool.