The next challenge is more operational: What happens when multiple AI agents need to work together across the same fraud and AML lifecycle?
An agent may help develop a fraud strategy. Another may support investigations. Another may prepare reporting. Individually, each can produce useful results. But the value of agentic AI at scale depends on what happens between those agents: how context moves, how decisions are governed, where humans intervene, and whether the entire chain can be trusted.
That was the focus of DataVisor’s recent webinar, Trust at Scale: Multi-Agent Orchestration for Fraud, Investigation, and Reporting.
The discussion explored how financial institutions can move from isolated AI agents toward coordinated, governed workflows spanning detection, investigation, reporting, and continuous optimization.
Why Multi-Agent Orchestration Matters for Fraud and AML
Fraud and AML operations are inherently connected.
A suspicious signal may trigger a detection strategy. That strategy produces a decision or alert. An investigator then needs to understand the entities, transactions, devices, relationships, and prior activity behind it. Findings may need to move into regulatory reporting. Finally, investigation outcomes should inform future strategy optimization.
The workflow looks something like this:
Signals → Detection → Decisions & Alerts → Investigation → Reporting → Optimization
Introducing AI at only one point in this lifecycle can accelerate an individual task. But it does not necessarily improve the end-to-end process.
This is where multi-agent orchestration becomes important.
Rather than treating each AI agent as an isolated copilot, an orchestrated system coordinates specialized agents across the workflow while preserving the context, evidence, permissions, and human oversight required at each stage.
For fraud and AML teams, the objective should not be to deploy as many agents as possible. It should be to create a system in which specialized agents can reliably work together.
The Most Important Part of a Multi-Agent System: The Handoff
A multi-agent system is only as strong as the handoffs between its agents.
Imagine a detection agent identifies a potential synthetic identity fraud ring involving several linked accounts. It may have access to shared devices, IP addresses, triggered rules, behavioral signals, prior alerts, and transaction activity.
An investigation agent receiving that case needs the relevant context to understand why those accounts were flagged.
If only the alert itself is transferred, valuable information may be lost. If too much unstructured information is transferred, the investigation agent may struggle to determine what matters. If permissions are unclear, the downstream agent may take an action it was never intended to take.
The same problem appears when investigation findings move into reporting.
A reporting agent should not have to reconstruct the investigation from scratch. It should receive structured, approved findings and the supporting evidence necessary to produce the required narrative or documentation.
That makes the handoff contract a practical foundation for multi-agent governance.
For every agent-to-agent handoff, fraud and AML teams should be able to answer five questions:
- Context: What information is being transferred?
- Evidence: What supports the agent’s output?
- Authority: What can the receiving agent recommend or execute?
- Approval: Where must a human review, edit, approve, or reject the result?
- Auditability: What is recorded for future review, governance, and learning?
Explicitly defining these elements makes an agentic workflow easier to understand, test, audit, and improve.
From AI Agents to an AI Control Plane
As organizations deploy more agents, individual governance controls are no longer enough. They need a shared operating model for coordinating the system.
This is the role of the AI control plane.
A governed multi-agent architecture can be viewed across four layers:
Data layer → Agent layer → Orchestration layer → Human and regulatory interface
The data layer provides the signals, events, entity information, policies, and other context agents need.
The agent layer contains specialized AI agents responsible for defined tasks.
The orchestration layer manages how those agents interact, including context routing, handoffs, permissions, policies, and workflow state.
And the human and regulatory layer preserves accountability for consequential decisions while providing the oversight required for fraud, AML, model risk, compliance, and audit teams.
Across those layers, a mature control plane should address five areas.
1. Identity and Access
Who can invoke, configure, approve, or publish an agent?
Role-based access controls should extend beyond who can simply use an AI capability. Institutions also need to define who can modify it, approve its outputs, and authorize changes to production workflows.
2. Context Management
What information is available to each agent at each stage?
Context continuity becomes increasingly important as the number of agents grows. Each downstream agent should receive the relevant signals, evidence, prior decisions, and workflow history without forcing teams to reconstruct context manually.
3. Policy Enforcement
Which actions can an agent recommend, and which can it execute?
These boundaries should be explicit.
An agent might be permitted to recommend a new fraud rule, for example, while deployment requires human approval. Similarly, an investigation agent may propose a disposition without having authority to close the case or trigger a consequential action.
4. Observability
Can teams see how the system is performing?
Organizations should monitor measures such as latency, failures, output quality, drift, low-confidence responses, and exception rates across the agentic workflow.
A multi-agent system introduces dependencies. One slow, unavailable, outdated, or low-quality component can affect everything downstream.
5. Change Management
Can changes be tested, traced, audited, and reversed?
Agent configurations, prompts, workflows, integrations, and versions should follow controlled change processes. Teams should know what changed, who approved it, what evidence supported the decision, and how to roll back the change if necessary.
The Hidden Risk: Orchestration Sprawl
Deploying more AI agents does not automatically create a more sophisticated AI operation.
Without governance, it can create orchestration sprawl.
Orchestration sprawl occurs when agents, prompts, integrations, queues, and handoffs multiply without a common operating model.
Warning signs can include:
- Duplicate agents performing similar tasks
- Different agents following inconsistent policies or thresholds
- Integrations governed differently across teams
- Limited end-to-end monitoring
- Unclear ownership of the overall workflow
- Agents receiving incomplete or conflicting context
This distinction matters because a successful agent pilot can hide problems that emerge only when agents begin interacting.
An individual agent may perform well while the overall workflow fails because of context loss, conflicting decisions, permission drift, evidence gaps, or downstream latency.
The production question therefore changes from “Does this agent work?” to “Can we trust the system when one part of the chain does not?”
What Multi-Agent AI Can Look Like in a Fraud Investigation
During the webinar, DataVisor demonstrated how its AI agent, Vera, can support a workflow extending from fraud strategy into investigation and reporting.
Consider a synthetic identity fraud use case.
A strategy-focused agent can analyze production and labeled data, evaluate existing fraud detection strategies, identify gaps, and propose new features or rules.
Once the strategy is established, an investigation-focused agent can use that context to help design the downstream investigation process, including:
- Alert queues and routing
- Triage and escalation processes
- Investigator-facing information and evidence
- Investigation checklists
- Risk scorecards
- Service-level expectations
- Knowledge Graph views
- Operational dashboards and reporting
A reporting-focused agent can then build from the approved investigation structure and outcomes to help surface the metrics and reporting needed by relevant stakeholders.
The important point is not simply that AI can assist with each task.
It is that the context generated at one stage can move into the next stage through a governed handoff, reducing the need for each team or agent to start over.
Human reviewers remain part of the process. They can review specifications, examine evidence and test results, request changes, approve or reject recommendations, and determine whether changes should move into production.
Prompts, outputs, supporting evidence, approvals, and timestamps can also be logged to create an audit trail around how the workflow evolved.
Five Questions to Ask Before Scaling an AI Agent
Before adding an agent to a production fraud or AML workflow, teams should evaluate more than its standalone accuracy or capabilities.
A useful readiness assessment covers five dimensions.
- Scope
Is the agent responsible for a clearly defined task? - Context
Does it receive the information necessary to perform that task, including relevant upstream decisions and evidence? - Authority
Are the boundaries between recommendations and executable actions clearly defined? - Accountability
Can a human reviewer edit, approve, reject, or stop the workflow at consequential decision points? - Resilience
What happens when the agent fails, times out, receives outdated information, or produces a low-confidence result?
If these questions cannot be answered clearly, adding another agent may increase operational complexity faster than it increases value.
Where Should Human Approval Remain Mandatory?
Human-in-the-loop controls become more important, not less, as agentic workflows expand.
The appropriate approval architecture will depend on the institution, workflow, and risk involved, but human review should remain explicit around consequential decisions and actions that affect production systems.
For example, an agent might identify a strategy gap and recommend a change. A human can evaluate whether the recommendation is grounded in appropriate evidence before allowing the workflow to proceed.
The same principle applies across investigation and reporting.
Humans should be able to intervene when context appears incomplete, supporting evidence is weak, recommendations are inconsistent with policy, or the system is moving toward an executable action with meaningful consequences.
The purpose is not to insert manual friction into every automated step. It is to preserve clear accountability where judgment matters.
Where Should Financial Institutions Start With Multi-Agent AI?
Scaling does not have to begin with a large fleet of agents.
In fact, a more practical approach is often to start with a small number of agents around a well-understood workflow, standardize the handoffs, instrument the process, and expand as operational needs become clear.
Investigation and reporting can be useful starting points because these functions often involve substantial manual work around evidence gathering, summarization, documentation, and workflow design.
From there, institutions can progressively connect additional capabilities across detection and strategy optimization.
The sequence might evolve from:
Single agent → Coordinated workflow → Governed orchestration → Continuous optimization
At each stage, teams should establish shared context, evidence requirements, permissions, approval points, and monitoring before increasing the number of agents.
The goal is not agent proliferation. Each additional agent should solve a defined operational problem and fit within the governance architecture already in place.
Trust the Chain, Not Just the Agent
Agentic AI creates an opportunity to rethink how fraud and AML operations work across the entire lifecycle.
But as financial institutions move from individual AI assistants toward multi-agent systems, the measure of success changes.
The question is no longer simply whether an individual agent can generate a useful recommendation.
Can the detection agent pass the right evidence to investigation? Can investigation preserve that context through disposition and reporting? Can humans understand and control consequential decisions? Can every action be traced? And can the system respond safely when one component fails?
That is what turns a collection of AI agents into a dependable operating model.
The winning question isn't how many agents you have. It's whether you can trust the entire chain.
To see these concepts in practice, watch DataVisor’s webinar, Trust at Scale: Multi-Agent Orchestration for Fraud, Investigation, and Reporting, and learn how Vera supports governed AI across fraud and AML workflows.
