37 GenAI in Banking & Finance : Governed Hub-and-Spoke
Governed Hub-and-Spoke: The Enterprise Control Plane for Agentic Risk Management
As banks and organizations move from early experiments with agentic AI to large-scale deployments, the nature of the challenge changes. When you are running one or two intelligent agents inside a single team, already previously discussed orchestration patterns on this blog and clever prompts are usually enough. But when that number grows into the dozens—spanning AML, fraud, credit risk, operational risk, cyber, and conduct—the problem is no longer about intelligence alone.
It becomes a problem of governance.
At enterprise scale, agentic AI is not just a technical capability. It is an operational, regulatory, and organizational concern. Leaders need to know what agents exist, what data they touch, who owns them, how they perform, and how they interact across risk domains. Regulators expect transparency. Boards expect control. Risk functions expect defensibility.
This is where the Governed Hub-and-Spoke architecture comes into play. It provides the missing control plane that transforms agentic AI from a collection of tactical prototypes into a strategic, enterprise-grade platform for risk management.
From Agent Patterns to Agent Platforms
Over the past year, many teams have adopted structured multi-agent patterns: hierarchical supervisors, event-driven choreography, or consensus panels. These patterns work extremely well at the workflow level. They help break down complex reasoning into manageable steps and make agent behaviour more explainable.
However, these patterns do not solve a different class of problems that emerge at scale:
-
Who owns each agent?
-
Which data sources can it access?
-
How do you prevent unapproved agents from being used?
-
How do you monitor performance across domains?
-
How do you demonstrate control to regulators?
These are not orchestration problems. They are platform problems.
The Governed Hub-and-Spoke architecture addresses this gap by introducing a central enterprise control plane that governs all agentic activity, regardless of domain, model, or implementation style.
Why Governance Becomes Non-Negotiable at Scale
In large financial institutions, agentic AI inevitably spans multiple domains and organizational boundaries. Consider a typical risk landscape:
-
Financial crime teams running AML and sanctions screening agents
-
Fraud teams deploying real-time payment monitoring agents
-
Credit risk teams using agents for exposure analysis and stress testing
-
Operational risk teams classifying loss events
-
Cyber teams coordinating incident response
-
Conduct risk teams analyzing voice-of-customer data
Each of these teams has different objectives, data entitlements, and regulatory obligations. They also operate across different lines of defense—from 1st LOD business controls to 2nd LOD oversight and 3rd LOD audit.
Without a governing structure, agentic AI in this environment quickly becomes fragmented and risky. Discovery becomes difficult. Security boundaries blur. Observability is inconsistent. Cross-domain scenarios require fragile point-to-point integrations. Worst of all, regulators lose confidence that the organization truly understands and controls its AI estate.
Governance is not a luxury here. It is a prerequisite for trust.
The Core Idea: A Central Hub with Domain Spokes
The Governed Hub-and-Spoke architecture introduces a clear separation of responsibilities.
At the center sits the hub—a shared enterprise service that acts as the control plane for all agentic activity. Surrounding it are spoke agents, each owned by a domain team and responsible for a specific slice of intelligence.
The hub does not attempt to understand financial crime, fraud, or credit risk. Instead, it focuses on governance, security, routing, and observability. The spokes do not worry about authentication, data entitlements, or audit logging. They focus entirely on domain logic.
This division mirrors proven patterns in enterprise architecture, where control planes govern execution planes. It also aligns closely with how regulators expect risk systems to be structured.
What the Hub Is Responsible For
The hub is the single source of truth for the agentic estate. Its responsibilities typically include five core capabilities.
1 Agent registry and discovery.
Every agent must be registered with metadata describing its purpose, owner, version, line of defense, data sources, and service level expectations. This registry becomes the organization's AI inventory—an authoritative answer to the question, “What AI systems do you operate?”
2 Security and policy enforcement.
All requests flow through the hub, which enforces authentication, authorization, data access rules, and policy constraints. This ensures that agents only see the data they are entitled to and that usage aligns with organizational and regulatory requirements.
3 Routing and orchestration.
The hub decides which agents are invoked for a given case and how they are combined. This may involve parallel execution, sequential pipelines, or more complex patterns such as hierarchical supervisor teams. Crucially, orchestration logic is explicit and centrally governed.
4 Observability and monitoring.
The hub provides consistent metrics, tracing, and logging across all agents. Latency, error rates, success rates, and usage volumes are visible in one place. This allows teams to detect degradation, manage capacity, and understand operational risk.
5 Audit and compliance.
Every agent invocation, data access, and decision artifact is logged immutably. This creates a defensible audit trail that can be reviewed by internal audit or regulators without reverse-engineering individual systems.
What the Spokes Are Responsible For
Spoke agents are deliberately simpler in scope. Each one is owned by a specific domain team and encapsulates domain intelligence.
An AML spoke might focus on transaction behavior and sanctions exposure. A fraud spoke might specialize in velocity patterns and anomaly detection. A credit risk spoke might analyze exposures under stress scenarios.
The key is that spokes do not manage their own governance. They expose a well-defined interface and rely on the hub to handle identity, security, and orchestration. This keeps domain teams productive while maintaining enterprise consistency.
Preventing the Shadow AI Problem
One of the most underappreciated risks in large organizations is shadow AI—agents built and deployed without visibility, approval, or oversight. This risk increases dramatically as LLM tooling becomes more accessible.
The governed hub-and-spoke architecture addresses this structurally. If the hub is the only path through which agents can be invoked, unregistered agents simply cannot participate in enterprise workflows. Registration becomes the gateway to production use.
This approach shifts governance from policy documents to architecture. Instead of relying on teams to “do the right thing,” the platform enforces it by design.
Supporting Multiple Lines of Defense
Financial institutions operate under a three-lines-of-defense model for a reason. Business teams, risk and compliance functions, and internal audit all have distinct roles and responsibilities.
A governed hub-and-spoke architecture can reflect this structure directly. Agents can be tagged by line of defense, and the hub can enforce rules about which agents can be invoked in which contexts. For example, a first-line operational agent may not be allowed to override a second-line risk assessment, and audit agents may operate in read-only mode.
This alignment between organizational structure and technical architecture is critical for regulatory credibility.
Enabling Cross-Domain Risk Scenarios
Modern risk events rarely stay within a single domain. A cyber breach may trigger fraudulent transactions, which in turn raise AML concerns and liquidity risks. Handling these scenarios requires coordination across multiple agent types.
Without a hub, cross-domain orchestration becomes difficult to manage. With a hub, it becomes a natural extension of the platform. A single request can trigger multiple domain agents, with the hub aggregating and contextualizing the results.
This capability turns agentic AI into a true enterprise risk fabric rather than a collection of silos.
Observability as a First-Class Capability
In regulated environments, lack of observability is itself a risk. It is not enough to know that agents exist; you must know how they are behaving.
The hub provides a consistent observability layer across all agents. This includes performance metrics, failure rates, and usage patterns, as well as lineage information showing which data sources were accessed for each case.
Over time, this observability enables continuous improvement. Teams can identify underperforming agents, refine prompts or models, and demonstrate to regulators that controls are actively monitored.
Lifecycle Management and Continuous Control
Agentic systems evolve rapidly. Models are updated, prompts change, and new data sources are introduced. Without lifecycle management, this dynamism becomes dangerous.
The governed hub-and-spoke architecture supports explicit agent lifecycles. Agents can move from development to testing, from approval to active use, and eventually to deprecation. The hub can enforce that only approved versions are callable and that deprecated agents are phased out safely.
This mirrors established practices for model risk management and brings agentic AI into alignment with existing governance frameworks.
Deployment at Enterprise Scale
From an operational perspective, the hub-and-spoke pattern aligns well with modern cloud and container platforms. The hub runs as a shared service, while agents are deployed independently by their owning teams. Scaling is straightforward: hot agents can be scaled out without affecting the rest of the system.
Security is enhanced through consistent identity and transport controls between hub and agents. Cost management also improves, as usage can be tracked per agent, per domain, and per case.
Real-World Impact
Organizations that adopt a governed hub-and-spoke approach report tangible benefits. Investigations become faster because cross-domain intelligence is accessible through a single interface. Governance improves because AI inventories and audit trails are readily available. Regulatory interactions become less adversarial because transparency is built in.
Perhaps most importantly, teams regain confidence that agentic AI is something they control, not something that controls them.
When the Hub-and-Spoke Pattern Makes Sense
This architecture is not necessary for every use case. It shines in environments where:
-
There are many agents across multiple domains
-
Multiple teams and lines of defense are involved
-
Regulatory oversight is expected
-
Data governance and auditability are critical
For small, single-team prototypes, lighter patterns are often sufficient. The hub-and-spoke becomes valuable when agentic AI transitions from experimentation to infrastructure.
Sample Python Base implementation using 100% open source model and developed using Antigravity.
This codebase implements a Hub-and-Spoke Multi-Agent Architecture designed for financial services (e.g., AML and Fraud detection). The "Hub" acts as a central coordinator, while specialized "Spokes" act as domain-specific experts.
Here is a breakdown of each file and its role in the framework:
1. common/
(Shared Infrastructure)
common/This directory contains the foundational components used by both the Hub and the Spokes.
- llm_client.py : Provides the OllamaClient class, which handles communication with a local Ollama server. It allows agents to generate text or chat using models like llama3.2.
- models.py: Defines the structured data models using Pydantic.
- Transaction: Represents the input data (amount, sender, receiver, etc.).
- Finding : : The result produced by a Spoke (score, rationale, and severity).
- OrchestrationResponse : : The final aggregated decision produced by the Hub.
- spoke_base.py : An abstract base class (BaseSpoke) that enforces a consistent interface for all specialized agents. Every Spoke must implement process_transaction and get_capabilities.
2. hub/
(Central Orchestration)
hub/The Hub manages the lifecycle, routing, and auditing of the entire system.
- registry.py: Acts as a directory service. Spokes register themselves here with their capabilities (like "AML" or "Fraud"), allowing the Hub to dynamically discover which agents to use for a particular transaction.
- orchestrator.py: The "brain" of the Hub. It receives a transaction, selects the appropriate Spokes, executes them, and then aggregates their findings into a single risk score and final decision (e.g., "APPROVED" or "FLAGGED").
- observability.py : Implements the AuditStore, which logs every event (requests, spoke calls, responses) to a JSONL file (audit_log.jsonl). This ensures full transparency and auditability.
3. spokes/ (Specialized Intelligence)
This is where the domain-specific logic resides.
- implementations.py: Contains the concrete agent implementations:
AMLSpoke : Prompts the LLM to act as an AML Compliance Officer to identify money laundering risks.
4. Root File
- hub_and_spoke_demo.ipynb: A demonstration notebook that wires everything together. It initializes the Hub infrastructure, registers the Spokes, runs a sample transaction through the system, and displays the resulting audit trail and final consensus decision.
Conclusion
Agentic AI is moving quickly from novelty to necessity in risk management. But intelligence without governance is a liability. The governed hub-and-spoke architecture provides the control plane that allows organizations to scale agentic systems safely, transparently, and confidently.
It does not replace existing agent patterns—it elevates them. By embedding governance into the architecture itself, it turns agentic AI from a collection of clever ideas into a platform that regulators can audit, executives can trust, and risk teams can rely on.
✍️ Author’s Note
This blog reflects the author’s personal point of view — shaped by 25+ years of industry experience, along with a deep passion for continuous learning and teaching.
The content has been phrased and structured using Generative AI tools, with the intent to make it engaging, accessible, and insightful for a broader audience.
Comments
Post a Comment