Deploying Agentic AI in Financial Services at Scale

nvidia · Finance · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Agentic AI in financial services

Financial institutions are moving from AI assistants to production agentic systems that act on behalf of customers, bankers, analysts, and fraud teams, creating new requirements for compute, proprietary data, accuracy, reliability, and security when agents interact with money and transactions.

Constraints:

Financial agents may handle real-time transactions or financial decisions.
Multi-agent systems require far more inference compute than one-shot generation.
Banks need proprietary data and domain-specific accuracy beyond generic foundation models.
Fraud and AML workflows require explainability and human analyst focus.
The source is an industry interview spanning Capital One, RBC, Visa, and Nvidia examples.

Architecture

The interview describes a financial-services agentic architecture: open-source foundation models such as Llama or Mistral are combined with bank proprietary data, reasoning and post-training, then specialized agents handle domain workflows. Examples include Capital One's Auto Navigator chat concierge, RBC's Aiden research agents, Visa/agentic payments, and Nvidia fraud blueprints that combine graph neural networks, embeddings, and ML models.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  foundation-model-base["Foundation model base"] --> bank-proprietary-data-layer["Bank proprietary data layer"]
  bank-proprietary-data-layer --> specialized-financial-agents["Specialized financial agents"]
  foundation-model-base --> specialized-financial-agents
  specialized-financial-agents --> external-action-systems["External action systems"]
  specialized-financial-agents --> human-financial-workers["Human financial workers"]
  transaction-data["Transaction data"] --> fraud-embedding-pipeline["Fraud embedding pipeline"]
  fraud-embedding-pipeline --> fraud-ml-models["Fraud ML models"]
  fraud-ml-models --> human-financial-workers

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Foundation model base`	Open-source or managed model used as the starting point for bank-specific agents.	`Bank proprietary data layer`, `Specialized financial agents`
`Bank proprietary data layer`	Domain data used for post-training and accuracy improvements.	`Foundation model base`, `Specialized financial agents`
`Specialized financial agents`	Agents for automotive buying, research report updates, fraud, AML, and customer-service workflows.	`Foundation model base`, `Bank proprietary data layer`, `External action systems`, `Human financial workers`
`External action systems`	Dealership scheduling, trading/reporting systems, payments, or fraud workflow tools acted on by agents.	`Specialized financial agents`
`Fraud embedding pipeline`	Graph/embedding pipeline for transaction and cybersecurity fraud detection.	`Transaction data`, `Fraud ML models`
`Transaction data`	Tabular payment or transaction history vectorized for prediction/anomaly comparison.	`Fraud embedding pipeline`
`Fraud ML models`	Models that detect anomalous transactions and suspicious activities.	`Fraud embedding pipeline`, `Human financial workers`
`Human financial workers`	Bankers, analysts, and fraud teams freed to focus on higher-value analysis and judgment.	`Specialized financial agents`, `Fraud ML models`

Data flow

A user, banker, or transaction enters a domain workflow. Agentic systems understand the problem, reason over available frameworks and proprietary bank data, call specialized agents or models, perform actions such as scheduling a dealership visit or updating a research report, and feed outputs into financial workflows. Fraud pipelines vectorize transaction/tabular data and compare new transactions against learned predictions.

“the agentic AI is actually thinking in the background experiencing and uh what prior frameworks and tools it's used to answer similar questions”transcript, stated

“instead of predicting the next word, they can predict the next transaction. And when you can predict the next transaction, you can see the new transaction”transcript, stated

Why this architecture

Financial institutions need proprietary data and specialized agents because generic models are not accurate enough for bank-specific questions, and multi-step financial workflows require reasoning, action, security, and domain context.

“you take an open- source foundation model which might be about 50% accurate to your bank specific needs and and questions.”transcript, stated

“Then you layer in through some post-training your bank proprietary data and that's going to improve accuracy by another 10 to 20 points.”transcript, stated

Models used: Llama · Mistral · Nvidia Neotron · Graph neural networks · Payments foundation models · Bank-proprietary AI models

Scale

Agentic systems in the interview are described as live production systems processing real transactions.
“Not just pilot programs, but actual live systems processing real transactions.”transcript, stated
Agentic AI can generate 100x to 200x more compute than expected.
“all of that thinking is generating 100x 200x more compute than we anticipated if not more.”transcript, stated
RBC's Aiden agents improved research-report generation cycle time by over 60%.
“it's actually improving the cycle time to report generation by over 60%.”transcript, stated
RBC agents analyze 10 times more data than a human can.
“they're able to encapsulate or analyze 10 times more data than a human can.”transcript, stated

Alternatives considered (roads not taken)

Single-shot generative AI

Rejected because: Agentic AI must understand, reason, and act rather than only answer a prompt.

“it's no longer just a oneshot outcome like we're used to experiencing with generative AI where we give it a prompt and we get an answer.”transcript, stated

Managed generic model services alone

Rejected because: Banks moved toward open-source foundations plus proprietary data for improved accuracy.

“the pendulum swing from often times these managed services to leveraging open-source foundation models.”transcript, stated

Only traditional transaction-table fraud models

Rejected because: Vectorized tabular data and payments foundation models can predict expected transactions and flag anomalies.

“you can vectorize tabular data just as you can uh unstructured data.”transcript, stated

Patterns

`multi-agent-orchestration` (architecture)

Capital One's Auto Navigator chat concierge is described as a multi-agent conversational AI for car buying.

“the chat concierge is actually a a multi- aent conversational AI that is designed to enhance the automotive buying experience.”transcript, stated

`tool-use-orchestration` (action)

Agents act in external workflows such as scheduling dealership visits and generating updated research reports.

“it'll actually reach out to the dealerships on behalf of the consumer to schedule test drives and dealership vis visits.”transcript, stated

`fine-tuning-adaptation` (modeling)

Banks combine open-source foundation models with proprietary data and supervised fine-tuning.

“you're going to do some supervised fine-tuning, which means essentially you're taking the bank GPT that you just trained”transcript, stated

`cost-aware-routing` (infrastructure)

The interview frames 100-200x compute growth as an inference-at-scale challenge for AI factories.

“this is really leading into an inference at scale challenge and opportunity for our AI factories”transcript, stated

Tradeoffs

Use agentic workflows in financial services

Chose agents that reason and act over assistive chatbot-only workflows.

Rationale: Financial workflows such as car buying, research updates, and fraud/AML need actions across systems and data streams.

“We're moving beyond AI as a tool that assists humans to AI as an actor that takes action on our behalf.”transcript, stated

Build proprietary bank models from open foundations

Chose open-source model plus bank proprietary data over managed generic model service alone.

Rationale: The interview frames the choice as accuracy-driven for bank-specific needs.

“I I think it goes back to accuracy first and foremost.”transcript, stated

Gotchas (what bit them)

Agentic AI changes the compute profile dramatically.

Root cause: Agents reason and act across multiple steps instead of producing one response.

Resolution: Plan inference factories and infrastructure for 100-200x compute growth.

“these agentic AIs, they uh first focus on uh understanding the problem at hand, then they think or reason through that problem, and then they act. And all of that requires compute.”transcript, stated

Generic model accuracy is too low for bank-specific needs.

Root cause: Open-source foundations may start around 50% accuracy for proprietary financial questions.

Resolution: Layer in bank data, post-training, and supervised fine-tuning.

“open- source foundation model which might be about 50% accurate to your bank specific needs and and questions.”transcript, stated

Evaluation

The interview evaluates financial agent deployments through production status, compute demand, model accuracy, report cycle time, data volume analyzed, fraud detection quality, and analyst productivity.

Metrics: Production deployment · Compute multiplier · Bank-specific accuracy · Report cycle time · Data analyzed · Fraud detection accuracy · Human analyst time

“you're getting again more accurate outputs faster when you're leveraging an agentic AI platform like Royal Bank of Canada is.”transcript, stated

Outcomes

Capital One's Auto Navigator uses a chat concierge for automotive buying support.
RBC's Aiden research agents improved report-generation cycle time by over 60%.
RBC agents analyze 10x more data than a human can.
Bank-specific post-training can improve model accuracy by 10-20 points beyond an open-source baseline.
Nvidia describes fraud blueprints using graph neural networks, embeddings, and production ML.

The source describes production financial systems handling real transactions and workflows, but many metrics are interview-level rather than formally audited.

“Capital 1 is AI agents helping customers buy cars. Visa is letting AI agents spend your money. RBC has agents executing trades”transcript, stated