What Should Transaction Monitoring Do When AI Scoring Is Unavailable?

Transaction Monitoring  ·  November 2026  ·  Australia & International

Why transaction monitoring needs a defined, tested fallback when an AI risk score cannot be produced.

When an AI component becomes unavailable, a transaction monitoring team faces a practical question: should the underlying monitoring process stop, or should it continue under a defined set of controls?

For systems involved in financial crime risk, resilience is part of decision quality. An intelligent score can help inform investigations, but a critical process needs a predictable outcome when that score cannot be produced.

transaction monitoring visual showing active rules, score status, fallback plan, incident playbook and recovery checks when AI scoring is unavailable

A resilient monitoring design defines what continues when an AI score cannot be produced, how the missing score is recorded and how recovery is validated. Illustrative only.

Quick answer

Monitoring should continue under defined controls. In FraudShield, AI scoring does not block transaction processing: if the AI Engine is unavailable, transactions continue against the active rules and the AI assessment becomes unavailable for that transaction. An unscored transaction is not used for AI-score-based automatic case creation, and it should never be represented as having received a low score.

Understand the Difference Between Rules and Scoring

Many transaction monitoring programmes use configurable rules to identify specified patterns or threshold breaches. An AI-supported approach may add another signal that helps investigators notice activity worth reviewing.

Those functions should not be confused. A rule can trigger based on documented conditions. An AI score may add context about other patterns or, in FraudShield’s documented implementation, how close a transaction came to breaching the configured rule plan.

Interpret scores carefully

An AI score is not automatically a probability that fraud occurred. Interpreting it as one could lead to unsupported conclusions.

Define the Fallback Before an Incident

A resilient design should specify what happens if scoring is disabled, fails or times out. At minimum, teams should know whether transactions continue through rules, how the failed scoring attempt is recorded and whether any alert-generation functions depend specifically on an AI score.

FraudShield’s AI risk-scoring documentation states that AI scoring does not block transaction processing. If the AI Engine is unavailable, transactions continue against the active rules; the AI assessment becomes unavailable for that transaction.

Not scored is not low risk

This is an important operational distinction. A transaction that lacks an AI score should not silently be represented as having received a low score.

What Happens to AI-Triggered Cases?

Where an organisation configures automatic case creation based on AI score bands, an unavailable score creates a specific exception. It cannot be evaluated against the score threshold in the usual way.

According to FraudShield’s current guidance, a transaction without a score is not used for AI-score-based automatic case creation. Rule-triggered processing continues independently.

Score status What happens Case creation
AI score produced Rules run, and the score can be evaluated against any enabled AI-score case-creation ranges. Rule and AI-score cases
AI score unavailable Rules continue to run. The transaction is not used for AI-score-based automatic case creation. Rule-triggered cases only

Operational teams should understand the consequence for alert coverage and determine whether additional review or monitoring is warranted during an extended scoring outage. The appropriate response depends on their own risk framework.

Build an Incident Decision Playbook

A useful playbook assigns owners to five questions:

Detection

How will we know scoring is unavailable or disabled?

Continuity

Which monitoring and case creation functions remain operational?

Coverage

Which types of transaction might no longer receive an additional signal?

Recovery

How will service restoration be validated?

Review

Do transactions processed during the interruption need follow-up analysis?

These decisions should be documented before an outage. Waiting for an incident makes it harder to distinguish a measured response from improvisation.

Test the Fallback, Not Just the Model

A test plan should simulate missing scores, disabled models and scoring errors while reviewing the expected rule outcomes and API or dashboard statuses.

Where the platform supports separate Test and Production environments, organisations can validate monitoring behaviour without treating a live environment as an experiment. FraudShield documents distinct Test and Production processing contexts, together with explicit model activation on rule plans.

Also review operator-facing wording. ‘Not scored’ must be distinguishable from ‘low risk’, and support teams need to know how to interpret both states.

Decision AI Needs Clear Limits

The broader lesson for Decision AI is that an intelligent component should have a defined scope of authority. If it cannot provide a result, downstream processes need an honest, controlled way to handle uncertainty.

For financial crime investigators, reliable status information may be more valuable than a system that always displays a number. Clear fallback rules protect the integrity of the decision record.

Reliability Builds Trust

An AI-enhanced monitoring programme is not credible because it promises uninterrupted intelligence. It is credible because it clearly states what continues to operate, which evidence is missing and how the organisation responds.

Take the Next Step

Explore FraudShield transaction monitoring and its AI risk-scoring documentation to understand how rule processing and AI signals work together.

Explore FraudShield AI risk scoring


Frequently Asked Questions

No. Its documentation states that transactions continue to be processed using the active rules.
No. It indicates that the score was not produced and should be handled as a separate status.

Talk to Nexiant about resilient monitoring

Speak with our team about how FraudShield rule processing and AI signals work together.

Contact Nexiant

This article was accurate at the time of publication in November 2026 and is intended for general informational purposes only. It does not constitute legal, regulatory or compliance advice. Organisations should seek qualified professional guidance in relation to their specific obligations.