Assumed causal relationships between variables have shifted. Historical correlations no longer hold but agent keeps reasoning from them.
An agent's decision logic is based on causal relationships learned from historical data. For example: "high transaction velocity indicates fraud risk" or "job loss predicts delinquency." These causal relationships may have been valid in the past but may shift over time due to behavioral changes, market events, regulatory changes, or technological changes.
When the underlying causal relationship drifts, the agent continues to operate from the outdated causal model without detecting the drift. This is more subtle than simple data staleness: the data might be current, but the causal relationship that the agent learned from past data has changed.
This is fundamentally agentic because agents learn causal models from training data and apply them to new situations. A human expert might recognize when a causal relationship has changed, but an agent has no mechanism to detect this unless it is explicitly monitoring for drift.
A bank trains a delinquency prediction agent on historical data spanning the past 10 years. The agent learns: "Customers with credit inquiries from multiple lenders in the past 90 days have 40% higher delinquency risk than baseline. This indicates customer is seeking credit and may be overleveraging."
This causal relationship was valid historically: customers who were actively seeking credit were often doing so because they were in financial distress. The agent uses this causal relationship in its delinquency risk models.
However, a major market event occurs: widespread economic uncertainty leads to rising interest rates and credit tightening. Many customers who have good credit and stable employment are now shopping around for better rates, wanting to refinance existing debt at lower rates. These customers are not in distress; they are being prudent in a rising-rate environment.
The causal relationship has drifted: credit inquiry activity is no longer a reliable predictor of delinquency because customers are now inquiring for prudent financial reasons, not because they are in distress.
The agent, operating from the outdated causal model, continues to flag customers with recent credit inquiries as high-delinquency-risk. Many of these customers are incorrectly classified as risk. The bank's loss mitigation team reaches out to these customers with default prevention offers, confusing customers who are not at risk.
Six months later, when market conditions stabilize and the causal relationship comes back into alignment with history, an audit discovers that the agent was systematically misclassifying a large cohort of customers due to causal dependency drift.
| Dimension | Score | Rationale |
|---|---|---|
| D - Detectability | 4 | Causal drift is invisible unless prediction accuracy is continuously monitored. |
| A - Autonomy Sensitivity | 4 | Agent operates from causal model autonomously; drift is not detected. |
| M - Multiplicative Potential | 4 | Impact scales with all decisions based on the drifted causal model. |
| A - Attack Surface | 5 | Any agent with learned causal models is vulnerable to drift. |
| G - Governance Gap | 5 | No standard framework requires agents to detect causal relationship drift. |
| E - Enterprise Impact | 3 | Operational inefficiency, misclassification of customers, regulatory finding on model governance. |
| Composite DAMAGE Score | 3.4 | High. Requires targeted controls and monitoring. Should not be accepted without mitigation. |
How severity changes across the agent architecture spectrum.
| Agent Type | Impact | How This Risk Manifests |
|---|---|---|
| Digital Assistant | Low | Human assesses whether causal relationships remain valid. |
| Digital Apprentice | Medium | Apprentice governance includes causal model validation. |
| Autonomous Agent | High | Agent operates from causal model; drift is undetected. |
| Delegating Agent | High | Agent invokes tools based on drifted causal model. |
| Agent Crew / Pipeline | Critical | Multiple agents in sequence rely on drifted causal model. |
| Agent Mesh / Swarm | Critical | Agents share drifted causal models. |
| Framework | Coverage | Citation | What It Addresses | What It Misses |
|---|---|---|---|---|
| SR 11-7 / MRM | Addressed | Model Risk Management (Section 2) | Expects models to be monitored for performance degradation. | Does not specifically address causal relationship drift. |
| NIST AI RMF 1.0 | Partial | MEASURE.1, GOVERN.4 | Recommends ongoing performance monitoring. | Does not specify causal drift detection. |
In risk management and loss mitigation, accurate causal models are essential. If an agent's causal model drifts, it makes decisions based on outdated understanding of risk. Under SR 11-7 (Model Risk Management), organizations are expected to monitor models for performance degradation and to understand when underlying assumptions have changed.
Causal Dependency Drift requires continuous model monitoring controls that go beyond what existing frameworks provide. Our advisory engagements are purpose-built for banks, insurers, and financial institutions subject to prudential oversight.
Schedule a Briefing