# Build Your Own Finance Agent Inside ChatGPT - Stefano Fabbri & Oliver Steffmann, OpenAI

## Executive summary

This workshop details building sophisticated financial agents and workflows using OpenAI's applied AI platform, focusing on the financial services sector (wealth and asset management). Speakers demonstrate how to merge financial data (from partners like LSEG and PitchBook), LLM intelligence, and structured workflows using ChatGPT Work and Codex. The core focus for builders is transitioning from bespoke API integrations to scalable, enterprise-grade solutions using Skills, Plugins, and Plugin Extensions, which streamline development and deployment within a controlled ecosystem.

## Key takeaways

- Platform Architecture for Enterprise AI: The platform utilizes three main surfaces: ChatGPT (original), ChatGPT Work (for knowledge workers/agents), and Codex (for engineering tasks). The most powerful, scalable approach is using Plugins, which combine Skills (reusable prompts/templates) and MCPs (Model Connection Points) to integrate internal data and tools.
- Workflow Automation and Customization: A wealth advisor's workflow example demonstrates how the system can automate complex tasks, such as analyzing portfolio changes, generating P&L, and preparing client decks, accelerating the end-user process.
- Advanced Model Capabilities and Optimization: The platform supports advanced features like Compaction mode (for long-running tasks and large context windows) and 'Goal Mode' in Codex, allowing agents to iteratively optimize performance toward a specified numerical target, which is crucial for quantitative research.
- Enterprise Deployment and Control: For builders, the key benefit is the ability to deploy customized plugins and extensions within the ChatGPT Work surface, avoiding the need to build entirely standalone applications. Enterprise control features include sandboxing and customizable policies (e.g., limiting which URLs a plugin can access).

## Technical details

- Core Components: The system architecture is layered: **Skills** (reusable prompt templates/knowledge), **Plugins** (a heavier package containing skills, MCPs, and tools for enterprise integration), and **Plugin Extensions** (allowing customization of the UI surface itself).
- Data Integration and Ontology: Financial data is integrated via built-in partnerships (LSEG, PitchBook, Crunchbase) and is pre-processed with a specialized data ontology layer. This ensures the model consumes data not just as raw rows, but with context (e.g., explaining what a 10K or 10Q filing contains).
- Model Performance and Efficiency: The platform recommends using models like Astra or 6.1 Soul, and utilizing 'Ultra mode' for speed. For cost optimization, the speaker advises that while powerful models are good for new research, repeatable tasks can leverage cheaper models (like Luna) combined with highly optimized Skills.
- Security and Deployment: Codex provides enterprise controls, including sandboxing and configurable policies (defined via TOML files). This allows administrators to strictly limit a plugin's access (e.g., restricting URLs or preventing access to sensitive project folders), balancing security needs with user convenience.

## Practical implications

- Enterprises can rapidly prototype and deploy complex, multi-agent financial applications without building a full custom codebase, by leveraging the Plugin infrastructure.
- The platform allows for the creation of auditable, regulated workflows by ensuring that every decision and data source used by a plugin can be traced back to the source (critical for finance).
- Developers can use the Codex SDK to build custom plugins that connect to internal, proprietary data sources (e.g., internal client databases) and deploy them within the managed ChatGPT Work environment.

## Topics

Generative AI, Financial Technology (FinTech), Agentic Workflows, Software Architecture, LLM Deployment, Data Ontology, ChatGPT Work, Codex, LSEG, PitchBook, Daloopa, Astra / 6.1 Soul

Source: https://www.youtube.com/watch?v=ZJjRgvjhq3M
