Six layers. One spine. Evidence from strategy to external assurance.
Six connected operating layers
Each layer answers a distinct question, with distinct owners, cadences and proof. Layer 3 divides into 3a and 3b because technical enforcement inside the system and operational control around it are different jobs. The layers form a loop, not a waterfall — assurance findings feed back into strategy.
Enterprise AI Authority Decisions
What must be true, and what does the organisation authorise? The constitutional decisions that govern the governors: authority and automation boundaries, competence as a precondition for delegated AI authority, protected challenge rights, AI-contribution disclosure obligations, incentive boundaries and workforce principles. Owned by the board and CEO.
Governance
Who decides, on what evidence, with what authority and risk appetite? Owned by the board, CRO, CISO and AI governance.
Operationalisation
How do governance expectations become delivery work, controls, evidence and value? Owned by PMO, product, delivery and DevSecOps.
Technical Runtime Enforcement
Are authority, boundaries and prohibitions enforced in the system itself — permissions, guardrails, tool and agent boundaries? Owned by AI governance and security architecture.
Operational Runtime Control
Can we see, constrain, stop, recover from and learn from live AI behaviour? Owned by AI operations, SOC and incident response.
External Assurance and Regulatory Engagement
Can the organisation prove control to independent reviewers, auditors and regulators? Owned by audit, legal, compliance and assurance.
Layer 0 decides what must be true and what the organisation authorises. Delivery determines how the organisation will make it true. Layer 0 contains the decisions that govern the governors.
One lifecycle for every AI use case
The spine runs across all six layers, giving every AI use case the same end-to-end path from discovery to continuous improvement.
The conversion chain
The core logic beneath every layer, gate and artefact. It prevents organisations from stopping at principles, framework mappings or control statements that have no operational implementation — and enables traceability from any requirement through to the evidence that demonstrates ongoing effectiveness.
Assist, Augment, Automate, Agentic
AURA separates how much decision and action authority is delegated to AI — and insists on the lowest sufficient authority that reliably delivers the outcome. A valuable AI contribution may sit at assist or augment; maximum autonomy is never the default destination.
Assist
AI supports a person's work. The person performs the task; AI accelerates, drafts or informs.
Augment
AI performs substantive work under close human direction and review. Human judgement remains in the loop.
Automate
AI executes defined tasks within set boundaries, with human oversight, monitoring and intervention pathways.
Agentic
AI holds bounded, delegated authority to decide and act — observable, interruptible and recoverable in production.
Risk, maturity and authority are not the same thing
Mixing these concepts creates false confidence and over-engineering. AURA keeps them distinct:
| Scale | What it measures | Question it answers |
|---|---|---|
| Risk tier | Control intensity an individual use case requires | How much control depth does this use case need? |
| Maturity level | How consistently the organisation operates and evidences the method | How reliably do we run AURA across the portfolio? |
| Authority stage | How much decision and action authority is delegated to AI | What may this AI decide and do on its own? |