Propensity models are predictive models that estimate the likelihood of a person, account or event matching a target outcome. In AI and security work, they are often used for risk scoring, targeting, fraud detection or prioritisation, based on signals such as behaviour, history or context.
What Propensity Models Are Used For
Propensity models estimate the likelihood that a person, account, or event will match a target outcome. Practitioners use them to rank future interest, prioritise review, score risk, or decide which cases deserve attention first.
They are common in fraud analytics, marketing, security triage, and operational decisioning because they turn many weak signals into a probabilistic estimate rather than a yes-or-no rule. That makes them useful when teams need to compare large populations quickly.
How Propensity Models Work
At a high level, a propensity model learns patterns from historical data, then assigns a score or probability to new observations. The input features usually come from observed behaviour, account history, device or context signals, and other attributes that correlate with the target outcome.
The model itself does not prove that an outcome will happen. It only estimates likelihood based on the training data and the assumptions encoded in the feature set, model class, and thresholding strategy. In practice, that means performance depends heavily on data quality, label quality, and whether the environment that produced the training data still resembles the one being scored.
Because the output is predictive rather than deterministic, propensity models are often paired with policy rules, human review, or downstream controls. A score may trigger step-up verification, case creation, additional monitoring, or a hold for manual approval, but the score alone should not be treated as a source of truth.
Why Propensity Models Matter in Security and AI Work
In security workflows, propensity models help teams focus limited attention on the accounts, events, or sessions most likely to matter. That can improve prioritisation for fraud review, abuse detection, insider-risk triage, and investigation queues, especially when the volume of signals is too high for manual review alone.
They also matter because the score can influence access, escalation, or enforcement decisions. A poorly calibrated model may over-prioritise benign activity, under-prioritise risky behaviour, or create inconsistent treatment across users, accounts, or events. Those failures are especially problematic when the score becomes part of an automated control path rather than a supporting signal.
Model quality is also shaped by the evidence it consumes. If historical labels are biased, stale, incomplete, or easily manipulated, the model may learn the wrong associations and reinforce bad decisions at scale.
Common Failure Modes and Governance Concerns
Propensity models can drift as behaviour, products, threats, and user populations change. They can also be gamed when attackers or other actors learn which signals raise or lower a score. In security settings, that means the model may become less reliable exactly when adversaries begin to adapt to it.
Another common issue is overconfidence in the score. Teams sometimes treat a probability as a decision, when it is only one input to a broader process. That creates operational risk if the threshold, review workflow, or exception handling is poorly governed.
Because these models often work on behavioural and contextual data, they can also carry privacy, fairness, and explainability concerns. Even when the model is technically accurate, an organisation may still need to justify why the scoring logic is appropriate, what data is used, and how decisions are reviewed.
Risk and Threat Considerations
Propensity models create security risk when they are used as a gate for prioritisation, enforcement, or trust decisions without strong validation. They can be manipulated, drift out of date, or produce misleading scores that hide real abuse while flagging benign activity.
Failure mechanism: Adversaries can probe scoring behaviour, alter their activity to stay below thresholds, or exploit weak labels and stale features so the model learns an inaccurate view of risk. Operational teams can also misread the score as certainty instead of probability.
Impact: The result can be missed fraud, delayed incident response, poor access or review decisions, unnecessary friction for legitimate users, and a false sense of confidence in automated prioritisation.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
NIST CSF 2.0, NIST SP 800-53 Rev 5 and NIST AI RMF set the technical controls, while ISO/IEC 27001:2022 defines the regulatory obligations.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST CSF 2.0 | GV.RM-01 — Risk Management Strategy | Propensity models affect security prioritisation and decision risk. |
| ID.RA-03 — Cyber Threats and Vulnerabilities Are Identified and Recorded | Propensity models rely on identified signals, labels, and changing threat conditions. | |
| DE.AE-02 — Potentially Adverse Events Are Analyzed to Support Response Activities | Propensity scores help triage events and prioritise investigation. | |
| Recommendation — Define scoring thresholds and review rules as part of enterprise risk management. Validate model inputs against current threat and abuse patterns. Use propensity scores to rank adverse events for faster response. | ||
| NIST SP 800-53 Rev 5 | AU-6 — Audit Record Review, Analysis, and Reporting | Propensity scoring decisions should be reviewable and explainable in logs. |
| SI-4 — System Monitoring | Monitoring is needed to detect drift, abuse, and unexpected scoring behaviour. | |
| RA-3 — Risk Assessment | Propensity models are risk-estimation tools that need validation and review. | |
| Recommendation — Log score inputs and decisions so analysts can review model-driven outcomes. Monitor model outputs for drift, evasion, and anomalous scoring patterns. Assess model error, bias, and drift before relying on the scores operationally. | ||
| ISO/IEC 27001:2022 | A.5.7 — Threat intelligence | Threat and abuse patterns should inform the signals a propensity model uses. |
| A.8.16 — Monitoring activities | Continuous monitoring is needed to detect scoring drift and misuse. | |
| Recommendation — Feed current threat intelligence into model feature selection and review. Continuously monitor model behaviour and downstream decision quality. | ||
| NIST AI RMF | GV.1 — Govern, map, measure, and manage AI risks | Propensity models are AI-based decision support that need explicit risk governance. |
| Recommendation — Govern propensity models with documented risk ownership and monitoring. | ||
Practitioner Guidance
Why practitioners should care: A propensity model is only as useful as the decision process around it. Use it as a prioritisation signal, not as an unreviewed final authority, especially when the output affects access, fraud handling, or security escalation.
What to watch for: Monitor for drift, threshold instability, label quality problems, and adversarial adaptation. If model outputs start to diverge from observed outcomes, treat that as a control problem, not just a data-science issue.
Practitioner takeaway: The safest deployments combine predictive scoring with governance, monitoring, and a clear override path so the model can improve decisions without silently replacing judgment.