Financial institutions are moving fast on GenAI, but not always forward. While the promise is clear (faster investigations, smarter automation, better service), scaling GenAI responsibly remains elusive. Banks face unique constraints: legacy infrastructure, fragmented AI stacks, rising model complexity, and a regulatory environment where oversight is not optional.
So how do leaders move beyond proof-of-concept to production at scale? How do they build GenAI systems that are accurate, transparent, cost-efficient, and deployable across critical workflows like fraud detection, anti-money laundering (AML), and Know Your Customer (KYC)?
Keep reading to see what true GenAI scalability looks like in practice, including a closer look at a production-ready architecture that bridges strategy and engineering, keeps humans in the loop, and adapts to diverse banking environments without sacrificing compliance or control.
The New GenAI Landscape in Banking
Opportunities and Risks
Banks have been at the forefront of tech adoption for decades, from ATMs to mobile banking. But GenAI marks a shift in capability and risk.
We’re moving from a tool that analyzes data to systems that reason independently and can take action.
- Malcolm DeMayo, Global VP of Financial Services at NVIDIA
This evolution is transforming customer engagement, decision-making, and business models. Capital One’s “Chat Concierge,” for instance, exemplifies AI stepping beyond transactional support to orchestrate full customer journeys. But with that potential comes the pressure to move fast, and without compromising trust.
AI Factories and the Rise of Sovereign AI
Referencing NVIDIA’s launch at GTC Paris, Malcolm introduced the concept of “AI factories”: a shift from static data centers to intelligent systems that actively generate value. But doing so in financial services requires careful navigation of sovereignty, compliance, and security.
As highlighted by BNP Paribas CTO Jean-Michel Garcia, responsible and sovereign AI infrastructure should be built on three foundational pillars:
- Independent and trustworthy infrastructure: Control over infrastructure is critical to maintain sovereignty and resilience. Institutions must be able to choose where and how their AI runs, whether in public cloud, private cloud, or dedicated environments, without overreliance on a single vendor or jurisdiction.
- End-to-end data protection across public, private, and dedicated cloud environments: With AI expected to “live everywhere,” data must be protected consistently from ingestion through output. This requires encryption, strict access controls, and monitoring across all environments to uphold privacy, security, and compliance.
- Reliable models aligned with regulatory and ethical standards: Models must meet the accuracy demands of production use while operating within financial regulations and ethical AI principles. This includes bias detection, explainability, and traceability so that every decision can be understood, validated, and audited.
Aligning Strategy and Execution: Where Dataiku Fits In
Next, a critical question facing financial institutions today is: How do banks turn GenAI vision into real systems? The answer lies in architectural alignment, connecting business priorities to engineering execution through governed, flexible platforms.
You’re really trying to strike a difficult balancing act… You need a consistent and governed space, but one that still connects to legacy systems, diverse model vendors, and hybrid environments.
- John McCambridge, Global Solutions Director for Financial Services at Dataiku
The takeaway? GenAI doesn’t erase long-standing enterprise complexity; it rather amplifies the need for adaptable, scalable design.
While GenAI feels new, the underlying enterprise tension is familiar. The fundamental nature of these problems, if you boil them down, are the same ones banks have been dealing with for decades. The difference now? Scale, complexity, and the pace of model evolution.
A Concrete Solution: The AML Agent Assistant Blueprint
A production-ready solution that brings these ideas together is a GenAI-powered AML agent assistant. Co-developed by Dataiku and NVIDIA, the AML assistant shows how agentic systems can boost human investigation workflows without sacrificing auditability or control. It’s built using:
- NVIDIA NIM™ to host optimized foundation models.
- The Dataiku LLM Mesh to orchestrate agent flows across tasks and tools.
- Visual and code-based interfaces to empower cross-functional teams to build, monitor, and extend solutions.
Crucially, this agent-based system doesn’t replace investigators. As John noted, "The human agent remains and is massively more effective, much more efficient, and much more capable of making the critical decisions that they are ultimately there to make."
How It Works
The AML Agent Assistant is designed to fit seamlessly into an investigator’s workflow while dramatically reducing the time and effort needed to assess alerts. The process begins when the system receives an incoming AML alert, either triggered by a specific condition or on a scheduled basis.
A master agent orchestrates the investigation by delegating to multiple specialized sub-agents, each focused on a specific investigative task. These sub-agents, powered by NVIDIA NIM microservices for the latest AI foundation models and orchestrated through the Dataiku LLM Mesh, can:
- Search across internal datasets linked to the alert, such as customer profiles, transaction histories, and prior case notes.
- Conduct graph-based entity exploration, mapping relationships between accounts, entities, and transactions to highlight potentially suspicious links.
- Generate detailed risk summaries, including unusual findings, red flags, and suggest next steps for deeper investigation.
- Draft sections of a Suspicious Activity Report (SAR) that investigators can review and finalize.
All outputs are consolidated into a structured investigator dashboard within Dataiku. This dashboard includes:
- A high-level summary for each alert, generated automatically.
- Associated risk scores and key identifiers such as account numbers and transaction IDs.
- Interactive graph visualizations that let investigators explore entity connections in context.
- A contextual chat interface connected to the same agent logic, enabling follow-up questions and additional analysis on demand.
By automating the most time-intensive aspects of information gathering and structuring, the system frees investigators to focus on interpretation, decision-making, and case resolution.
Why It Matters
Deploying GenAI in a regulated environment requires balancing speed with trust, transparency, and alignment to real-world workflows. This blueprint addresses several critical needs for GenAI in financial services:
- Human-on-the-loop design ensures accountability and compliance. The system augments rather than replaces human investigators. Every output is reviewable, editable, and auditable, ensuring compliance requirements are met without slowing investigative throughput.
- Agentic logic that reflects established investigative processes mirrors how AML teams already operate: gathering information from different sources, cross-referencing it, and iterating toward a conclusion. This familiar structure reduces adoption barriers and makes the system intuitive to use.
- Flexible, cross-functional development and governance enabled by Dataiku’s visual and code-based tools, allowing both technical teams and business analysts to build, monitor, and refine the system together.
- Transparent AI decision-making via Dataiku’s Trace Explorer, which captures every step an agent takes (including which tools were called, how long each step took, and the associated costs), providing a clear, auditable record for governance, optimization, and regulator-ready reporting.
The result is a scalable, compliant, and adaptable framework for bringing GenAI into mission-critical financial workflows, starting with AML but extendable to fraud detection, KYC, and beyond.
Beyond AML: A Reusable, Scalable Architecture
This AML assistant is just the beginning. "These NIM microservices are lego blocks that can quickly be redeployed for other blueprint automations," Malcolm said. And that’s the point: what makes this blueprint powerful isn’t just the AML use case, it’s the architectural design.
The same modular approach, the Dataiku LLM Mesh, NVIDIA NIM, agent orchestration, hybrid model routing, visual prompts, can support fraud detection, KYC, client onboarding, claims triage, and more.
If you’re asking how to deploy GenAI in financial services responsibly, efficiently, and at scale, this blueprint is a real answer. With Dataiku and NVIDIA, you keep humans in the loop, scale with confidence, and turn GenAI from vision into value.