Stefano Fabbri and Oliver Steffmann, both of OpenAI's applied AI engineering team, ran a two-hour finance build session carried by AI Engineer. The demo ran inside ChatGPT Work for Financial Services and the Codex desktop app, with no outside attacker and entry gated to an enterprise workspace.
Can a wealth team do portfolio prep without shipping data to a one-off app? The short answer from the session is yes, if data, tuned behavior and document output live in the same surface.
ChatGPT Work for Financial Services is OpenAI's enterprise build for knowledge workers, distinct from base ChatGPT for chat and Codex for engineering tasks. Fabbri said it rests on three parts: piped third-party data, models tuned for jobs like M&A and equity research, and artifacts that draft presentations and spreadsheets from reference templates. The data list he named included SEC filings, news, private company profiles and earnings call transcripts, plus market data delayed by 15 minutes, Fabbri said.
A skill, which is reusable instructions that push the model toward one finance job, gives advisors a starting baseline. Fabbri described a plugin as a heavier bundle of those skills plus MCP connectors, where an MCP server is an open connector that lets the model pull data and take actions in outside systems. He said OpenAI built the investment banking and equity research plugins with practitioners, then tells teams to treat the result as 80 percent done and add the last 20 percent for house style. Steffmann, who said he once built a similar SEC ingest inside an asset manager, warned the hard part is serving fresh filings at speed with a context layer the model can actually use.
The wealth demo walked from a portfolio change to profit and loss attribution to a client deck that explains the rebalance. Steffmann framed the build choice for internal developers as path A, a custom app on an API key with its own upkeep, or path B, a skill turned into a plugin and then a plugin extension that changes the ChatGPT and Codex interface for that workflow. A Human Mode guide describes the smaller DIY split many teams know, where the model reads messy merchant names while Power Automate runs the repeatable moves.
That leaves the budget call with wealth operations and data owners, who must fund curation of private data and template standards before any agent touches client paper. Watch whether advisors accept the 15-minute delay and enterprise-only gate, and whether plugin extensions reduce one-off app work or just move it inside ChatGPT.
