Artificial Intelligence and Execution Control

APRA · AI RISK MANAGEMENT

Artificial Intelligence and Execution Control

Response to APRA’s letter to industry on artificial intelligence (AI)

APRA’s letter to industry on artificial intelligence reinforces the importance of governance, accountability, model risk management, and human oversight.

These are necessary foundations.

But they do not yet address a critical control point:

the moment an AI-driven decision becomes institutional action.

Where current controls stop

Most enterprise control frameworks operate:

before execution — policy, approval, model validation
after execution — monitoring, audit, incident response

But institutional consequence occurs at execution.

At this moment, capital is committed and legal obligation is created.

money moves
contracts activate
infrastructure changes
regulatory records are created

Most architectures do not control this moment.

Most control frameworks do not operate at the moment of execution.

Execution control model showing post-hoc governance and the absence of authority resolution at execution

This diagram illustrates the execution gap and the introduction of a non-bypassable boundary where admissibility is resolved at T=0.

The execution control gap

When an AI system proposes an action, the critical question is not only:

“Was this decision well governed?”

It is:

“Is this action allowed to execute right now?”

In most environments, this question is not resolved at execution.

As a result:

actions may execute under outdated authority
conditions may no longer be valid
evidence may be incomplete
system state may have changed

Execution proceeds regardless.

In most environments, accountability cannot be reconstructed at the exact moment of execution.

Execution admissibility

The missing layer is a control mechanism at the point of execution.

A boundary where authority, context, state, and evidence are resolved at the moment of commitment (T=0).

Execution is allowed only if those conditions are satisfied.

No action binds without admissibility at commit.

Operational resilience, accountability, and AI

This control point directly supports:

Operational resilience

Preventing inadmissible actions before consequence occurs

Accountability

Establishing who had authority at the exact moment of execution

Model risk management

Ensuring model outputs do not translate into unauthorised actions

Audit and assurance

Providing replayable evidence of admissibility at execution

Extending current control frameworks

APRA’s direction strengthens governance and oversight.

Execution admissibility extends this by introducing:

a runtime control boundary
pre-execution validation of authority and conditions
deterministic, evidence-bound execution decisions

This complements — rather than replaces — existing governance frameworks.

This introduces a control point that is not currently defined within standard governance or model risk frameworks.

From governance to control

AI systems can now generate decisions at scale.

The next step is ensuring that only admissible actions are allowed to bind institutional consequence — at the moment execution occurs.

Intelligence may propose.

Only admissible execution may bind.

Further discussion

For organisations exploring execution control in AI-enabled environments: