Back to Diagnoses
COA DiagnosisDelegated Cognition Failure

AI Credit Lending: Silent Authority

When Advisory Became Authoritative Without Anyone Deciding

What is a COA Diagnosis?

Cognitive Operating Architecture (COA) governs how institutions delegate cognition to machine systems. Unlike AI safety (which governs model behaviour) or AI ethics (which governs design choices), COA governs the institutional conditions under which delegation occurs—how authority emerges, how accountability persists, and how learning asymmetry is managed.

A regional bank deployed a machine learning model to support credit lending decisions. The model was explicitly positioned as advisory—loan officers retained final authority and could override recommendations. All governance requirements were satisfied.

Two years later, a regulatory inquiry revealed systematic lending disparities. When investigators asked how decisions were made, the institution discovered it could not explain its own processes. The model had become the de facto decision-maker, but the accountability structures still assumed human judgment.

The AI system did not malfunction. The missing layer was Cognitive Operating Architecture— the institutional infrastructure that should govern delegation boundaries, monitor authority accumulation, manage learning asymmetry, and maintain accountability continuity.

“The model was always 'advisory.' But when no one overrides advice, and performance rewards alignment with advice, 'advisory' becomes 'authoritative' without anyone deciding it should.”

The authority laundering this case reveals

The Four COA Failures

All four functions of Cognitive Operating Architecture failed in this case. Click each function to see how it failed.

Authority Drift Timeline

Month 0: ML model deployed as 'advisory decision support'
COA not considered
Month 6: Override rate drops to 8%; model influence grows
No authority threshold monitoring
Month 12: Training materials updated to emphasise model alignment
Institutional learning favours model
Month 18: Override rate at 2%; model is functionally authoritative
Authority laundering complete
Month 24: Regulatory inquiry into lending disparities
Accountability chain doesn't exist
Month 30: Institution cannot explain its own decisions
Cognitive dependence irreversible

COA Failure Modes Present

Institutionalised Automation Bias

Human judgment deferred to model recommendations systematically

Ritualised Oversight

Compliance reviews occurred but examined the wrong version of the system

Authority Laundering

Model decisions presented as human decisions for compliance purposes

Responsibility Diffusion

No single human accountable for AI-mediated outcomes

Irreversible Delegation

Officers lost capacity to make decisions without model input

What COA Would Have Provided

Explicit Delegation Thresholds

Defined criteria for when a system transitions from advisory to authoritative

Authority Accumulation Metrics

Tracking override rates, confidence clustering, and deference patterns

Learning Asymmetry Alerts

Triggers when system learning outpaces institutional oversight cycles

Accountability Mapping

Clear chains from outcome to responsible human for every decision type

Trust Recalibration Schedules

Mandatory intervals to verify human judgment remains independent

Sunset Conditions

Pre-defined criteria for reverting delegation if thresholds are crossed

COA Assessment for Your Organisation

Use these four questions to assess whether your organisation has cognitive operating architecture for AI systems.

1

Are delegation boundaries explicitly defined with measurable thresholds for advisory vs authoritative?

2

Is authority accumulation tracked—do you know when deference patterns indicate silent authority transfer?

3

Is trust recalibration scheduled—are there mandatory intervals to verify human judgment independence?

4

Is accountability mapped—can you trace any AI-mediated outcome to a responsible human?

Related Resources

Is Your AI Governance Actually Governing?

AI safety governs models. AI ethics governs design. COA governs the institutional conditions under which delegation occurs. Most organisations have the first two but not the third.