GenAI in banking creates fraud gains and governance strain

At a glance

What this is: NHI Mgmt Group's analysis, drawing on Feedzai's blog post: GenAI is now embedded in banking strategy while also expanding fraud, scam and governance risks through synthetic content, agentic automation and data-quality failures.

Why it matters: IAM, fraud and AI governance teams need to treat GenAI as both a business capability and a control problem because the same systems that improve detection can also amplify deception, decision risk and trust failures.

By the numbers:

• 96% of global banks are already implementing GenAI.
• GenAI-related crime could reach $40 billion by 2027, according to Deloitte forecasts cited by Feedzai.

Context

Generative AI has moved from experimentation into production banking workflows, but the security and governance question is whether institutions can control the data, decisions and identity boundaries that these systems now touch. In financial services, that means the risk is not just model accuracy but also fraud amplification, privacy exposure and weak accountability around automated outputs.

Feedzai frames GenAI as both a defensive tool and an offensive enabler: banks can use it to detect scams, investigate alerts and generate synthetic scenarios, while criminals can use the same techniques to scale deception. That dual-use reality makes governance, review and auditability central rather than optional.

Key questions

Q: What breaks when GenAI outputs are treated as trusted evidence in fraud operations?

A: When GenAI outputs are treated as trusted evidence, hallucinations, biased summaries and overconfident recommendations can steer investigators toward the wrong conclusions. The control failure is not just model error. It is the absence of a review layer that checks provenance, confidence and business impact before the output influences a case or customer decision.

Q: Why do endpoint agentic AI tools create more governance risk than chat-only GenAI?

A: Endpoint agentic AI can act inside a user’s session, move data, and trigger downstream actions, which expands the effective privilege boundary. Chat-only tools may still create data risk, but agents can combine access and action in ways that are harder to see and easier to over-scope. That makes workflow control and telemetry more important than simple application approval.

Q: What are the signs that AI fraud controls are failing?

A: Weak controls usually show up as rising scam volume, more difficult-to-spot fraudulent content, higher account takeover rates, and customers abandoning the brand after suspicious interactions. Another warning sign is overreliance on rigid rules that create false positives while missing adaptive attacks. If fraud teams cannot distinguish trusted behavior from AI-assisted manipulation, the detection model is not keeping pace.

Q: How should banks balance GenAI innovation with privacy and accountability requirements?

A: Banks should limit GenAI to governed use cases, keep sensitive data tightly controlled, and maintain clear accountability for every output that affects customers or risk decisions. The practical test is whether the organisation can explain how data was used, who approved the output and what checks were in place before action was taken.

Technical breakdown

How GenAI changes fraud detection workflows

GenAI differs from traditional predictive models because it can generate synthetic examples, summarise complex cases and extract patterns from unstructured data. In fraud operations, that shifts the control surface from simple scoring to assisted decisioning, where the model may create case summaries, recommend next steps or generate rules. That raises a governance requirement: the output becomes part of the control, not just an input to analysis. If the model hallucinates or overstates confidence, investigators can be pushed toward the wrong action path. Practical implication: treat GenAI outputs as governed evidence, not autonomous truth.

Practical implication: Validate how generated summaries, recommendations and synthetic cases are reviewed before they influence fraud decisions.

Why agentic AI creates a new trust boundary

Agentic AI is more operationally risky than standard GenAI because it can act with minimal human supervision, make decisions and execute tasks across workflows. That matters in banking because the model is no longer only generating text or analysis; it may be participating in transaction review, customer engagement or compliance workflows. The security issue is therefore identity and authority, not just model quality. Teams need to know what the agent can access, what it can decide and where the human handoff sits. Practical implication: define explicit authority boundaries for any agentic workflow before deployment. See the AI Infrastructure Workload Identity Guide.

Practical implication: Map every agent action to a clear authority boundary and remove standing access that exceeds the task scope.

How synthetic data can help and hurt

Synthetic data is useful because it can simulate rare fraud events, test controls and reduce pressure on sensitive production datasets. But it also creates a governance trap if teams assume synthetic outputs are automatically safe or representative. The quality of the generated data depends on the training distribution, the prompt or generation method and the validation process. If those are weak, synthetic data can hide edge cases, bias the model or create false confidence in resilience. Practical implication: verify synthetic datasets against real-world fraud patterns before using them to tune controls. See the Generative AI Data Loss Prevention Guide.

Practical implication: Use synthetic data only after testing it against known fraud patterns and documented validation criteria.

NHI Mgmt Group analysis

GenAI is becoming a governance problem before it becomes a technology problem. The article shows that banking adoption is already widespread, but the control question is whether organisations can govern data provenance, decision accountability and output quality at the same pace. In practice, that makes model oversight, auditability and human review part of the security design rather than an afterthought.

Financial crime teams are moving from static detection to assisted decisioning, and that changes the control model. When GenAI generates summaries, recommendations and synthetic examples, the organisation is no longer reviewing only raw alerts. It is also reviewing machine-shaped interpretation, which means errors can be amplified before a human sees the case. Practitioners should treat this as a decision-integrity issue, not just an analytics improvement.

Agentic AI introduces an identity boundary that traditional fraud tooling does not fully address. If an agent can act on behalf of a workflow, then the key governance question becomes what it is allowed to access, decide and execute. That is where identity and privilege controls intersect with AI governance, and where workload identity becomes relevant for the infrastructure behind AI systems.

Trustworthy synthetic content: this is the new control concept exposed by the article. Synthetic examples, summaries and scam simulations are only useful when teams can trust their origin, validation and limits. Practitioners should align GenAI usage with evidence handling, not just model performance targets.

The dual-use nature of GenAI means fraud defence and fraud enablement now share the same tooling category. The same capabilities that help banks simulate attacks or assist investigators can also help criminals scale scams and impersonation. That means governance has to focus on separation of duties, reviewability and dataset discipline. Practitioners should assume dual-use pressure from the outset.

From our research library:

• Generative AI use specifically increased from 33% in 2023 to 79% in 2025, according to McKinsey’s Global Surveys on the State of AI.

What this signals

Decision integrity will become the next major AI governance metric in financial services. As GenAI moves into fraud review, customer communication and compliance workflows, the question is no longer only whether the model is accurate. Practitioners will need to know whether the surrounding process can detect bad outputs, constrain scope and preserve accountability when automation participates in the decision.

Identity boundaries will matter more as banks extend GenAI into agentic workflows. Once an AI system can act on behalf of a process, privilege and authority become part of the risk model. That is why identity governance has to extend into the AI stack, including workload identity, delegated access and lifecycle control for any system that can execute actions.

The next control challenge is not model scarcity but model discipline. Banks already have strong incentives to adopt GenAI, but adoption without validation will widen fraud exposure and compliance friction. The organisations that succeed will be the ones that govern data, review outputs and define who or what is allowed to act on the model's behalf.

For practitioners

• Define governed use cases for GenAI in fraud operations Limit GenAI to use cases with clear review points such as case summarisation, triage support and synthetic scenario generation, then document who approves the output before it affects customer or transaction decisions.
• Set authority boundaries for agentic workflows Specify which systems, records and decisions an AI agent can access, and remove any standing access that allows it to act beyond the intended workflow scope.
• Validate synthetic data against real fraud patterns Compare synthetic examples with known scam and fraud typologies before using them for model tuning, control testing or investigator training.
• Add human review to high-impact AI outputs Require manual validation for outputs that affect fraud decisions, customer communications, regulatory filings or risk scoring, especially where hallucinations could create false confidence.

Key takeaways

• GenAI in banking is now a fraud and governance issue as much as a productivity issue, because the same capabilities that help investigators can also help attackers scale deception.
• The article pairs widespread adoption with a large projected crime impact, which means control design has to keep pace with deployment rather than follow it.
• Banks need tighter review, data discipline and authority boundaries before GenAI outputs are allowed to influence customer, fraud or compliance decisions.

Related NHI Mgmt Group resources: NIST AI RMF Guide · AI Infrastructure Workload Identity Guide · AI Security Platform Buyer's Guide · Generative AI Data Loss Prevention Guide

Key terms

• Generative AI: AI designed to create text, code, images, or other content in response to a prompt. It is usually reactive rather than autonomous, which means the main security concern is output quality, leakage, and misuse of generated content rather than independent action.
• Agentic AI: Autonomous AI systems capable of planning, deciding, and taking actions, including calling APIs, writing code, and orchestrating other agents, with minimal human oversight. Agentic AI introduces new NHI risks as agents must authenticate to external services.
• Synthetic Data: Synthetic data is information generated by a simulation or model rather than collected directly from the real world. It is useful for training and testing, but it still carries governance risk because it can reveal system behaviour, operational patterns, or business logic when broadly accessible.
• Decision Integrity: Decision integrity is the degree to which an organisation’s outputs remain consistent, explainable, and aligned to approved meaning. In data and AI programmes, it depends on controlled definitions, trusted sources, and the ability to prevent different tools from inventing conflicting interpretations.

Deepen your knowledge

Build practical skills in NHI governance, workload identity and machine identity security with the NHI Foundation Level course, the industry's only accredited NHI security programme. It helps identity and security practitioners connect emerging AI workflows to the controls their programmes already depend on.