Build Your Own Finance Agent Inside ChatGPT - Stefano Fabbri & Oliver Steffmann, OpenAI
Use ChatGPT for Financial Services as a starting point, then extend workflows with reusable skills and plugins connected to approved internal tools and data.
- Start with a built-in financial workflow and tailor its skill using your organization’s analysis rules and output templates.
- Use a plugin with MCP tools when a workflow needs authenticated access to internal systems or data.
- Evaluate agent quality with a simple, expert-reviewed test set before optimizing prompts or models.
The workshop’s main lessons are captured in the notes; the demos are setup-dependent. · Best for: Financial-services teams, AI builders, and engineers extending enterprise workflows.
An OpenAI workshop demonstrates a wealth-management workflow using financial data, then explains how developers can extend ChatGPT and Codex with skills, plugins, MCP tools, and customized interfaces. It also covers evaluation, security controls, and ways to scale or audit agent workflows.
Financial-services product and built-in data5:58
- The presenters describe built-in financial data such as SEC filings, delayed market data, news, private-company profiles, and earnings-call transcripts.
- They say the data is prepared with contextual information to help the model interpret it, and describe financial skills and plugins as starting points for common analyses.
Wealth-management workflow and workshop setup14:14
- The example workflow helps an advisor analyze portfolio changes, prepare attribution and client materials, and identify rebalancing actions.
- Participants are directed to a workshop site and ChatGPT for Financial Services; exercises cover portfolio questions, research, and charting.
Skills, plugins, and extensions17:30
- Skills package reusable instructions; plugins add tools, authentication, and integrations, often through MCP.
- Plugin extensions can add custom interface elements to Codex; presenters say this UI customization was then available on desktop, not the web interface.
- The financial-services data access shown is in the web product; the presenters say it was not available through the CLI.
Choosing models and managing reusable workflows40:36
- For repeatable tasks, the presenters recommend building good context and skills, then measuring whether a less costly model meets the quality target.
- Share and maintain a small set of useful skills; too many overlapping skills can be difficult to manage and may slow processing.
- Reference files and historical examples can guide spreadsheet, presentation, and research outputs.
Codex architecture, long-running work, and security47:24
- A plugin can combine skills, an MCP server, tool configuration, and UI components; teams can package and deploy it through enterprise mechanisms.
- Codex policies can restrict local files, network destinations, and plugin use; administrators can configure approval and permission controls.
- The presenters describe compaction as useful for maintaining context during long-running tasks, and the Codex harness as open source and customizable.
Evaluation, research, and writing back to systems1:36:48
- Build evaluations with subject-matter experts; start with simple checks rather than an LLM judge, and prefer yes/no criteria over 1–5 scoring when appropriate.
- For goal-driven iterations, specify a measurable target; the presenters describe Codex goal mode as a way to test and improve performance.
- Plugins can write to local files or remote systems through tools; regulated workflows may store auditable logs linking decisions to sources and logic.
Step by step
- Start with a relevant built-in financial workflow and inspect the prompt, data sources, and generated output.
- Write a reusable skill containing your analysis rules, expected steps, and output format; add templates or historical examples as references.
- If the workflow needs internal data or system access, build a plugin with the required MCP tools and authentication.
- Test the workflow against an expert-reviewed evaluation set; begin with simple pass/fail checks and refine the instructions or tools based on failures.
- Set enterprise permissions for files, network access, plugins, and approvals; add logging where the process must be audited.
- For repeatable workloads, compare models on the evaluation set and use the least costly model that meets the required quality.
Tools & resources mentioned
- ChatGPT for Financial Services — workshop product for financial data and financial-analysis workflows.
- Workshop website — setup instructions and exercises for portfolio questions, research, and charting.
- Codex — environment used in the workshop to build skills, plugins, and extensions.
- MCP — connection mechanism discussed for plugins that need tools or access to external systems.
- OpenAI API — mentioned as an option for custom agent use cases.
- Decision API — described as a yes/no decision tool that can support evaluations.
- Codex goal mode — described as a way to iterate toward a specified measurable target.
- Custom Codex TOML configuration — enterprise settings for permissions and sandbox policies.
Notes are AI-generated from the video's transcript and may miss nuance.