Governed workflow

How does governed AI produce a decision-grade output?

The method separates early discovery from later assurance. AI may expand possibilities, but reliance is controlled by scope, evidence, challenge, uncertainty and human authority.

Published: 25 July 2026Updated: 25 July 2026Status: Research and Pilot ReleaseAuthor: Walter Shepherd
1

Declare

Define the task

State the question, purpose, audience, consequence level, intended decision and required output.

2

Control

Apply Scope Lock

Set boundaries, jurisdiction, time period, exclusions and what the output is allowed to influence.

3

Register

Register evidence and assumptions

Identify sources, authority, currency, admissibility, transformations, uncertainties and missing material.

4

Discover

Explore widely

Generate options, perspectives, candidate mechanisms and competing explanations without treating them as confirmed.

5

Analyse

Decompose and compare

Break claims into components, compare at the same unit and identify patterns, dependencies and information loss.

6

Challenge

Test alternatives and weak links

Run adversarial review, falsification questions, uncertainty checks and consequence-sensitive stress tests.

7

Integrate

Synthesise without laundering certainty

Combine supported findings, retain unresolved conflict and state what remains unknown.

8

Assure

Assess permission-to-rely

Decide what the output may support, under which conditions, with what confidence and required human review.

9

Operationalise

Act, monitor and improve

Translate the supported output into roles, decisions, implementation steps, monitoring and review.

Reliance checkpoint

Loading a framework does not make the session governed.

A session is governed only when the declared controls are actually applied, missing material is disclosed, the audit trail is preserved and an authorised person accepts responsibility for use.

Discovery Potential is not Reliance Class

A novel possibility can be valuable while still being unsuitable for decision use.

Convergence is not proof

Agreement among models may reveal a pattern, but shared training data or assumptions may explain it.

Weakest link controls the claim

The overall conclusion cannot be stronger than its necessary unsupported component.

Permission is purpose-specific

An output suitable for exploration may not be suitable for policy, regulation or irreversible action.

Apply the workflow

Choose an application domain.

See how the common method adapts to policy, research, strategy, investment and assurance tasks.

Explore applications

See it in practice

Review a worked case.

Case studies show the question, evidence, method, weak links, uncertainty and permission-to-rely.

Open case studies