How Agentic AI is Transforming Wall Street

Ben Lorica

Josh Pantony on Agentic AI, LLM Orchestration, and the Future of Financial Analysis.

Subscribe: Apple • Spotify • Overcast • Pocket Casts • AntennaPod • Podcast Addict • Amazon •  RSS.

Josh Pantony, CEO of Boosted AI, discusses Alfa, an agentic AI platform that creates persistent AI workers for finance professionals, moving beyond traditional prompt-response AI to proactive, autonomous systems. The conversation covers Alfa’s technical architecture using multiple LLMs, data strategy combining premium partnerships with user-generated datasets, and real-world applications in financial analysis. Pantony explains how Alfa competes with incumbents like Bloomberg and envisions a future where AI companions provide continuous, conversational assistance to financial professionals.

Subscribe to the Gradient Flow Newsletter

Interview highlights – key sections from the video version:

Jump to transcript

Related content:

Support our work by subscribing to our newsletter📩

Transcript

Below is a heavily edited excerpt, in Question & Answer format.

Product Overview and Agentic AI

What is Alfa and how does it differ from traditional AI tools?

Alfa is Boosted AI’s flagship platform—an AI-powered terminal specifically built for finance professionals including portfolio managers, analysts, and hedge funds. Unlike traditional generative AI where you ask a question and get a one-time answer, Alfa functions as an agentic AI platform that creates persistent, autonomous “workers” that learn specific tasks and perform them repeatedly over time.

The key difference is moving from a Google-like experience (prompt and response) to a YouTube or TikTok-like experience where the system proactively surfaces valuable work based on understanding your needs. Once you’ve expressed ongoing interest in a topic or workflow, Alfa continuously monitors and delivers insights without repeated input. You can teach it your specific workflows and needs, and it will then proactively learn, build work for you, and enhance your entire process.

How does your definition of “agentic AI” differ from the common understanding?

Many people think of agents as a system where you ask a question, it does more work, and you get a longer, more detailed answer back—like an advanced search query. We think of agents differently. For us, an agent is like a dedicated worker that learns a specific task and then performs that task for you repeatedly and forever.

These agents can: summarize every new earnings call in your particular format; notify you every time an email about a stock on your watchlist arrives; extract specific financial metrics using your unique methodology automatically every time; monitor map data for new store locations daily. The key is that they perform long-range, repeatable work consistently. Users typically start with one or two high-value agents but power users can have hundreds of these workers operating simultaneously, consuming billions of tokens daily.

Technical Architecture and Implementation

How does Alfa’s architecture work under the hood?

Alfa uses what we call a “large language model choir”—a collection of different LLMs working together, including both standard models from providers like Anthropic and OpenAI, as well as internally tuned models. The architecture has three main components:

  1. Task Decomposition Model: A proprietary internal reasoning model that breaks down complex user requests into smaller sub-tasks. For example, “build a financial model” would be decomposed into: learn about the company, gather all relevant data, extract key numbers, build the model structure, and link all the pieces.
  2. Specialist Models: Once decomposed, the system routes each sub-task to the best-performing LLM for that specific job based on internal benchmarks. This could be a third-party model or one of our fine-tuned models. For example, one model might excel at mathematical operations while another handles Chinese data better.
  3. Authenticator Model: After work completion, another proprietary model acts as a verifier, reviewing all conclusions against source facts and ensuring information presented to users is sufficiently backed by evidence.

How do you handle model selection and optimization?

We do not rely on public benchmarks. While they can be a good initial filter, nothing beats testing models against your own specific data and use cases. We’ve developed a suite of proprietary benchmarks for tasks most important in finance, including:

  • Tensor manipulation and time-series data analysis
  • Foreign-language data processing
  • Complex table information extraction
  • Financial metric extraction with nuanced variations

Every time a new model is released, we run it through our benchmarks to determine its strengths and weaknesses. This allows our task decomposition model to intelligently route each sub-task to the most capable and efficient LLM for that specific job.

How do you ensure consistency and auditability in a probabilistic system?

This is a critical challenge in regulated finance. Our system is designed to be much more deterministic than standard LLM-based systems through several approaches:

  1. Testing Frameworks: Similar to traditional software development, we use unit tests for local code blocks and integration tests for the entire system. We define specific conditions at both the local subtask level and globally across the entire workflow.
  2. Behavioral Constraints: While the intelligence comes from the LLM, we use rules and heuristics to sandbox its behavior and ensure it consistently performs specific actions. This creates more deterministic behavior while preserving the underlying intelligence.
  3. Complete Audit Trails: The system provides the exact primary source for its information, the methodology used, and its reasoning process in a way users can understand. If the system analyzes portfolio exposure to China today and runs again tomorrow, it will use the exact same methodology.

Data Strategy and Management

What’s your approach to data acquisition and management?

Our approach is two-pronged:

  1. Premium Data Partnerships: We work with about 20 different top-tier data vendors including S&P and specialized providers like Quarter for earnings presentations. This gives us access to over 60,000 news sources and 7,000 data series covering everything from short interest to analyst estimates.
  2. User-Generated Custom Datasets: We empower users to generate their own custom datasets through our web agents that can scrape and interact with websites. For example, users can build agents to track daily product prices on Target’s website or monitor Domino’s expansion in China through map data, creating unique datasets that feed into revenue predictions.

This hybrid approach means users aren’t completely dependent on any single data source and can combine premium third-party data with proprietary data generation for powerful, flexible analysis.

How do you handle data dependency risks with external providers?

By maintaining partnerships with multiple top-tier data providers for core financial data while also providing the capability to generate custom data through web agents, we ensure users have alternatives. If a data source disappears, users can often recreate similar datasets through our agents, maintaining continuity in their workflows.

Practical Applications and Use Cases

What are the most impactful use cases you’re seeing in production?

Three standout use cases demonstrate the platform’s capabilities:

  1. Key Financial Metric Extraction: Analysts typically spend up to 40% of their time manually extracting financial metrics from presentations and filings. They do this manually because of idiosyncratic definitions and the need to normalize data across companies. Alfa agents understand exact definitions and automatically extract these numbers with full audit trails, matching specific formatting requirements.
  2. Email Monitoring and Workflow Automation: Portfolio managers receiving hundreds of emails daily can deploy agents to scan inboxes, identify all analyst and sell-side updates related to watchlist stocks, and provide concise summaries. This extends to automating morning note creation by synthesizing all relevant overnight information.
  3. “Superpower” Deep Analysis: When news breaks (like Nvidia’s Blackwell chip delay), users can task agents to scan 3,000+ companies’ earnings calls, summarizing which companies remain excited about Nvidia, which seek alternatives, and potential disruptions. This ability to analyze thousands of contexts simultaneously was fundamentally impossible before.

How do teams typically adopt and scale their usage?

New users usually start with one or two high-value agents—perhaps one for building financial models and another for email summarization. Once they see time savings, they typically expand. Power users can have hundreds of agents running simultaneously, processing billions of tokens daily. There’s no known upper bound on usage yet. About 70% of users are on the buy side, with 30% on the sell side. Most are non-technical fundamental analysts who benefit from the interface without needing to code.

Reasoning, Latency, and Multimodality

How do you balance quality, latency, and cost for reasoning-intensive tasks?

Our primary focus is quality, as finance clients are less price-sensitive if the tool provides genuine edge. The system dynamically determines user intent and task complexity. Simple queries like “Who is the CEO of Nvidia?” return in seconds. Complex requests like “What is the compensation history for the CEO of Nvidia compared to peers?” require extensive work—analyzing years of SEC filings, normalizing data, and producing tables.

The system communicates task complexity and expected completion time rather than offering separate “fast” and “deep” modes. This intelligent adaptation provides better user experience while maintaining quality.

How are you incorporating multimodality and what are its applications?

Multimodality is important in two key ways:

  1. Unlocking Non-Text Information: In finance, crucial context exists in charts, graphs, and images within investor presentations. Combining visual and text analysis provides richer insights. This extends to audio and video content like podcasts or interviews containing Alfa-generating insights not yet in analyst reports.
  2. Transforming User Interaction: We envision conversational AI interactions—getting verbal portfolio summaries during commutes, having the system listen to calls with experts or IR teams, and synthesizing private dynamic context with public data. This richer interaction allows agents to produce higher-quality, more personalized work.

Competitive Landscape

Why can’t incumbents like Bloomberg just build what you’ve built?

Traditional incumbents face several disadvantages:

  1. Data Moat Erosion: They rely on thousands of humans to manually extract and authenticate data. With modern AI, we can create data at a fraction of the cost with 99.9% accuracy.
  2. Technical Debt: Bloomberg has about 15,000 hand-coded screens costing billions to maintain. We estimate reproducing all these screens costs about $1 million in compute. Soon, our platform will generate custom UI on the fly based on user requests.
  3. Static vs. Dynamic: Their one-size-fits-all approach contrasts with our hyper-personalized, proactive system that learns individual preferences.

Their remaining advantage is distribution and the social network (Bloomberg chat), but that erodes as disruptive platforms reach scale.

How do you compete with horizontal AI players and internal bank builds?

While many banks attempt internal builds, most struggle to match a focused AI company’s development pace. The bigger threat comes from horizontal players like Anthropic and Perplexity—they have concentrated talent and move faster than incumbents. However, they lack deep focus on finance workflows. When building purpose-built AI for a specific industry, domain expertise and workflow optimization create significant advantages over horizontal players serving every industry.

Future Developments and Innovation

What developments do you most want to see in foundation models?

Three key areas:

  • Smaller, faster, cheaper models maintaining high fidelity
  • Continued improvement in structured reasoning capabilities
  • Support for persistent, embodied AI experiences (like glasses-based assistants providing real-time multimodal context)

More capable models running at lower cost and latency would unlock even more sophisticated workflows.

What innovations from other products have influenced your approach?

Several products have provided inspiration:

  • Cognition Labs’ Devin: One of the first to show truly magical UI for agents with visible “thought processes”
  • Developer tools (Replit, Vercel, Cursor): Their ability to create and share functional software pieces
  • Perplexity: In-line citations becoming industry standard
  • Social media (TikTok, Instagram, WhatsApp): Interaction patterns informing how AI agents should communicate with humans

How do you see AI changing finance workflows in the next few years?

We’re moving toward AI companions that are always present, consuming context from emails, watchlists, earnings calls, and conversations throughout the day. Instead of prompting systems, they’ll proactively surface insights based on comprehensive understanding of interests and workflows. Interactions will become more conversational and multimodal—talking to AI assistants during commutes for comprehensive briefings. The biggest unlock might be agents participating in day-to-day work, creating truly collaborative AI partnerships.

Access and Availability

Is Alfa accessible to individual investors or only institutions?

While originally built for hedge funds and financial institutions with institutional-grade capabilities, Alfa is now open to everyone. Anyone can sign up at boosted.ai for free access. General retail users now get direct access to the same powerful technology that top-tier hedge funds use, democratizing advanced AI capabilities in finance.