Press enter or click to view image in full size
The founder of https://www.openbb.co/ —
wrote this really interesting article on the Evolution of AI Agents in Finance https://didierlopes.beehiiv.com/p/the-evolution-of-ai-agents-in-finance
From Models to Workflows: The Real Future of AI in Finance
His phased analysis outlines where the challenges for both FinTech challengers and incumbent institutions evolve dramatically, touching everything from model strategy to the fundamental user experience.
A Personal Perspective on Complexity and Coherence
This topic is personal for me. I’ve built high-frequency trading businesses and models across multiple financial instruments. I’ve witnessed firsthand how seismic events, coupled with poor telemetry and risk management, shattered firms like Bear Stearns and Lehman Brothers. I saw a FinTech startup I founded on the 78th floor of the World Trade Center nearly destroyed by a seismic act of terror. I have two software patents as well in the space. I have seen events unfold quicker than financial models could ever predict before.
My experience has taught me that the world operates with high variance, irreducible to simple models. This is precisely why Large Language Models are so compelling: they offer a new way to bring coherence to a complex world and synthesize more resilient insights, particularly when used in conjunction with smaller, traditional financial models within a much more powerful workflow.
I’ve also had a front-row seat to the incumbent response, having worked for two of the largest market data firms in this space. It was a revealing contrast: the two giants were radically different in how they perceived their strategic position and their ability to execute in this new era.
The AI Model Paradox & The Flawed Path
The first harsh reality of this new era is that building a proprietary foundation model is a strategic dead end for most firms. The frontier models from labs like OpenAI possess a profound advantage due to transfer learning. Their reasoning power comes from being trained on the entire spectrum of human knowledge — from satellite images to the music of Bach. This allows them to perceive a complex confluence of patterns across modalities that humans can’t, an ability that a narrowly trained model cannot replicate. A model that understands a protein-folding paper, for example, can transfer that scientific reasoning to better analyze a biotech investment. This makes over-investment in building a new model from scratch a folly when better performance comes from augmenting an existing one with specialized data.
This reality makes it clear that the future isn’t about owning a model, but about creating intelligent workflows that leverage secure, open standards. This brings us to the current landscape. As Didier notes, protocols like MCP were a pivotal “Phase 4” stepping stone that commoditized connections, but they proved to be a double-edged sword. They’ve led to a supply inundation of simple, brittle connectors, masking a low value proposition.
For serious, cross-enterprise workflows, today’s agent protocols are non-starters. Vulnerabilities like prompt-injection, tool-poisoning, and agent-hijacking are fundamental blockers, making these early standards unfit for high-stakes commercial use. MCP is a great step however in a larger unfolding story.
The Next Frontier: A New Interface for Thought
The truly revolutionary leap happens at the user-facing layer, where we redefine the interaction between the analyst and the machine, moving beyond insecure connectors to a truly integrated experience.
The Data Consumption Reversal: Serving the Bots First
Historically, financial data has been optimized for human consumption. As Andrej Karpathy predicts, this is being inverted. Nearly all optimization efforts will soon focus on making content digestible for LLMs, not humans. The next-gen platform’s primary user is the AI agent itself, which means moving from human-readable PDFs to richly tagged, structured, API-first data feeds.
The Interface Becomes the Answer: The Generative Canvas
This bot-first reality powers the new financial desktop: an ephemeral, intelligent canvas generated on-the-fly for the exact task at hand. It’s a decision surface that moves beyond static visualization to dynamic, situational foresight.
For example, a portfolio manager reassessing NVDA’s exposure to Taiwan doesn’t scroll through reports. They summon a canvas where the AI has already synthesized:
- A Geopolitical Risk Map showing risks like tariffs or shipping bottlenecks.
- A Filtered News Feed touching NVDA’s specific suppliers.
- A Weighted Dependency Graph linking factories, shipping routes, and downstream customers.
- A Predictive Cost Chart that responds to new headlines.
The workspace moves from static data retrieval to dynamic, situational foresight. In this world, rigid, intent-based standards like FDC3, designed for a previous era of static applications, become less relevant as the AI itself orchestrates the components on the canvas.
The Language of Orchestration: “Vibe Coding”
The language we use to command this canvas is what Andrej Karpathy has called “vibe coding.” It’s not just for one-off commands; it’s the language used to orchestrate inference. It operates on two levels:
- Creating Tools: An analyst can generate an ad-hoc analysis like a backtest or, more powerfully, create reusable components for the canvas itself: “Build me a ‘Supplier Fragility Score’ that weighs political stability indices against single-sourcing risk.”
- Orchestrating Insight: The user then uses natural language to direct the AI on how to connect the components and what to infer from them. The prompt is no longer just asking for a dashboard; it’s asking for an answer:
“Summon the NVDA Taiwan canvas and integrate my ‘Supplier Fragility Score’. Now, cross-reference the news feed with the dependency graph. Highlight any supplier with a high fragility score and recent negative news sentiment, and model the potential impact on our predictive cost chart.”
The system turns that intent into a governed, compliant workflow. The risk of creating fragile, “footgun-riddled” software is mitigated by the “Iron Man suit” paradigm — an analogy from Karpathy — where the workspace provides the necessary security and data governance guardrails.
The Synthesis
Ultimately, the future of financial AI isn’t a monolithic “Finance-GPT,” but a programmable, composable, bot-first environment where analysts and agents co-create insights. This new world is built on a foundation of intelligent workflows and operates under the paradigm of human-in-the-loop augmentation. Its users interact through a dynamic generative canvas, speaking the creative language of vibe coding to orchestrate inference.
The game is no longer about owning the model; it’s about orchestrating intelligence. In Financial Services, the future belongs not to the hoarders of data or the builders of bigger models, but to the architects of actionable insight within workflows.