# Build Your Own Finance Agent Inside ChatGPT

Stefano Fabbri & Oliver Steffmann, OpenAI | AI Engineer NYC 2026 | 1:50:32

Source: https://www.youtube.com/watch?v=ZJjRgvjhq3M
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/build-your-own-finance-agent-inside-chatgpt
Published: 2026-11-07
Tags: deployment, enterprise, mcp, workflows

## TL;DR
- ChatGPT for financial services combines financial data, domain-specific workflows, and artifact generation in one web interface.
- Finance plugins package reusable skills, MCP connections, tools, and interface components so teams can start from an industry workflow and add their own data and logic.
- Codex lets engineers turn prototypes into enterprise-managed plugins, with configurable interfaces, deployment options, and sandbox policies.

## Summary
Stefano Fabbri and Oliver Steffmann present a workshop on building financial workflows inside ChatGPT Work and Codex. They begin with ChatGPT for financial services, which includes market data, SEC filings, private-company profiles, news, and earnings-call transcripts from data partners. Its finance plugins combine skills and MCP connections for tasks such as equity research, valuation, earnings analysis, and portfolio work. The example workflow follows a wealth advisor from detecting a portfolio change through attribution, client explanation, and presentation preparation. The second half focuses on extending the product. Engineers can add internal data through MCP, package skills and tools into plugins, and customize the interface through plugin extensions. The speakers compare skills with plugins, explain how Codex can support local and hosted development, and describe enterprise controls for authentication, deployment, network access, and sandboxing. They also discuss model choice, evaluation design, reusable templates, and the tradeoff between standalone applications and workflows that remain inside ChatGPT.

## Key ideas
### ChatGPT for financial services combines data access with financial workflows and artifacts
[04:57](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=297s)
Stefano describes three parts of ChatGPT for financial services: built-in and partner data, training for financial work such as M&A and equity research, and artifact creation. The product can build presentations and spreadsheets and use historical templates as references for new analysis. Oliver adds that the presentation of information matters too. Users can ask for charts, verification, and annotations showing where data came from. The goal is to give financial users a product that understands both the source material and the expected form of the output.

### The financial data layer adds context instead of exposing raw files
[07:51](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=471s)
The workshop includes market data, SEC filings, news, private-company profiles, and earnings-call transcripts. The speakers name LSEG, PitchBook, Daloopa, and Crunchbase as data partners or sources. Market data is delayed, but it can still be used inside the product. Stefano explains that the data is processed so the model can access the relevant meaning of a dataset, rather than receiving an undifferentiated raw file. For a 10-K or 10-Q, this includes explaining what parts of the filing contain relevant information. He says the system uses customized data access, similar to a tool call, rather than custom embeddings.

### Finance plugins give teams a starting point built from industry workflows
[10:06](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=606s)
The speakers define a plugin as a library of skills plus MCP connections. OpenAI worked with investment banks, asset managers, and equity researchers to build finance workflows that can be used as starting points. Examples include pitchbook preparation, earnings analysis, valuation, and LBO modeling. Stefano recommends taking an existing skill, using it to create a repeatable pipeline, and then adding the firm's own logic and preferences. He describes the benefit as starting from an industry-informed baseline rather than writing the whole workflow from zero.

### A wealth advisor workflow can run from portfolio change to client presentation
[14:25](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=865s)
Oliver uses a wealth management example for the hands-on work. A customer portfolio changes, and the advisor needs to understand what happened, prepare portfolio attribution or P&L, explain the changes, and decide what would bring the portfolio back into balance. The workflow can then produce a client-facing deck. The intended user is the wealth advisor team, which may have different working styles, so the prompts and workflow are meant to be customizable. The workshop provides sections for portfolio questions, research, charting, and the wider customer example.

### Plugins extend ChatGPT Work with internal data, tools, and custom interfaces
[17:02](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=1022s)
Oliver presents two ways to extend the product. A company can build a separate application with an API key and its own web interface, or it can keep the workflow inside the ChatGPT and Codex environment that employees already use. In the second approach, a skill describes reusable behavior, an MCP supplies connections and tools, and a plugin packages the heavier integration. A plugin can run on the web or desktop surfaces. Plugin extensions add interface components that can appear in response to a conversation, file type, or business workflow.

### The right skill can reduce model cost for repeatable work
[42:14](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2534s)
Stefano says the most capable model is useful for new research and new codebases, but repeatable tasks can use less expensive models when the context is designed well. A skill can explain the expected work in enough detail for a smaller model to perform well. In a multi-agent workflow, the orchestrator may need the strongest model while individual subagents can use smaller models. He connects this approach to evaluations, because teams need to measure whether the skill and model combination meets the required performance.

### Teams should limit and maintain their skills instead of creating one for every task
[1:08:09](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4089s)
Oliver warns that skills can become difficult to manage when every person or team creates many of them. He recommends defining core capabilities, such as portfolio analysis or trading analysis, rather than turning every step of a workflow into a separate skill. Teams also need to retest skills as models change. He says a plugin may be a better choice when a builder needs packaging and deployment control. A plugin can include skills, code, authentication, MCP connections, and interface components in one package.

### Codex supports a path from local prototype to managed enterprise workflow
[48:20](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2900s)
The architecture depends on the surface where the workflow will run. Codex can run locally and connect to local MCP servers, internal databases, client systems, or Salesforce. A plugin can then connect the workflow to ChatGPT Work or the desktop application. Oliver describes enterprise deployment through managed plugin folders and existing deployment mechanisms. The Codex harness is open source, and the speakers say it can also be moved to the server side through an SDK or App Runner. This gives teams a way to reuse work that began as a local prototype.

### Codex policies control what plugins can access
[1:28:47](https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=5327s)
Oliver explains that enterprise administrators can use TOML configuration files to set Codex policies. These policies can restrict websites, local workspace folders, network access, plugins, and other operations. The controls apply when the Codex client runs code or a plugin makes a tool call. He also describes a guardian mechanism that can help distinguish sensitive actions from less sensitive ones, so users do not need to approve every ordinary operation manually. The speakers present this as a way to combine enterprise security requirements with a workflow that remains usable.

## Notable quotes
- Oliver Steffmann: "The first part is really the use case. You bring this product, ChatGPT for financial service, to your wealth advisor teams and they can basically accelerate your workflows." (15:20)
- Stefano Fabbri: "A plugin is kind of like a library of skills plus MCPs." (10:24)
- Stefano Fabbri: "When you build something like a repeating task, you can get top-level model performance if you build the right context." (42:34)
- Oliver Steffmann: "The plugin is where I need some code, I need some logic, and I need to have some way of getting my credentials for authentication." (55:33)
- Oliver Steffmann: "The Codex provides policies, the config files TOML, that are managed on an administrator side as an enterprise." (1:29:27)

## Tools & references mentioned
- OpenAI
- ChatGPT
- ChatGPT Work
- ChatGPT for financial services
- Codex
- LSEG
- PitchBook
- Daloopa
- Crunchbase
- MCP
- App Runner
- Salesforce
- Figma
- Adobe
- SEC filings
- Codex SDK
- Decision API

## Who should watch
- You are building financial workflows for wealth management, asset management, investment banking, or equity research and want data access and artifacts in the same user-facing product.
- Your company already uses ChatGPT Work or Codex and you want to add internal data, tools, authentication, or custom interface components without starting with a separate application.
- You need to decide whether a reusable skill, a full plugin, or a plugin extension is the right packaging for an enterprise workflow.

## Related talks

- [How OpenAI Built ChatGPT for Bankers and Analysts](https://aietalks.com/talks/how-openai-built-chatgpt-for-bankers-and-analysts) (Joseph Kim, OpenAI, 18:59)
- [New Ways of Working with Codex](https://aietalks.com/talks/new-ways-of-working-with-codex) (Dominik Kundel & Charlie Guo, OpenAI, 58:03)
- [How Financial Planners With Zero Coding Experience Built Their Own Agents](https://aietalks.com/talks/how-financial-planners-with-zero-coding-experience-built-their-own-agents) (Matt Ragone, Range, 16:09)
- [OpenAI Codex Masterclass](https://aietalks.com/talks/openai-codex-masterclass) (Vaibhav Srivastav & Katia Gil Guzman, OpenAI, 1:01:59)
- [How to Build Financial Agents That Work Across the Web](https://aietalks.com/talks/how-to-build-financial-agents-that-work-across-the-web) (Jay Sahnan & Arik Bird, Browserbase, 19:38)
