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.

They define how decisions should be governed.

But they do not yet define control at the moment institutional consequence occurs.

The execution gap: control exists before and after execution, but not at the moment it binds.

Where the control challenge remains

Most enterprise control frameworks operate:

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

But consequence occurs at execution:

money moves
contracts activate
records mutate
access is granted

At this moment, most architectures do not resolve whether execution is admissible.

Approval does not establish authority at execution.

From assumed authority to constructed authority at execution.

From governance to execution control

APRA’s direction strengthens governance and oversight.

Execution control extends this by introducing a runtime control boundary.

At this boundary:

authority is resolved at the moment of execution
evidence is bound to the action
context is validated
system state is verified
policy is re-evaluated

Execution is permitted only if these conditions are satisfied.

Authority is not assumed from prior approval.

It is constructed at execution.

Execution control at T=0 (payment example)

Execution control in practice

At the moment a payment is about to execute:

authority must be valid now
supporting evidence must be complete and current
contextual constraints must be satisfied
system state must be verified
execution must be legally and policy compliant

If these conditions cannot be proven:

execution does not occur

This applies equally to AI-enabled decisions across lending, contracts, infrastructure, and access.

Alignment with APRA expectations

This control point makes APRA expectations operational at execution.

CPS 230 — Operational Risk & Resilience

Control must exist at execution to prevent inadmissible actions before consequence occurs.

CPS 234 — Information Security & Control Integrity

Controls must operate effectively at the time of use, including execution.

FAR — Accountability & Traceability

Execution must be attributable and supported by evidence at the moment it occurs.

These expectations converge at the moment execution is allowed to bind.

From AI governance to AI execution control

AI systems can generate decisions at scale.

The next step is ensuring that only admissible actions are allowed to bind institutional consequence.

This introduces a shift:

from decision governance
to execution control

Intelligence may propose.

Only admissible execution may bind.

Further discussion

For organisations exploring execution control in AI-enabled environments: