Stefano Fabbri and Oliver Steffmann of OpenAI’s applied AI team lead a 111-minute AI Engineer workshop on financial workflows, with additional Q&A from colleague Dom.

Start with an advisor’s actual workflow

  • The exercise follows a portfolio change through attribution, research, rebalancing discussion, and a client presentation.
  • ChatGPT for financial services supplies financial data, task-specific instructions, and formatted output within the workshop’s enterprise workspace.
  • The description names Daloopa, LSEG, and PitchBook among its data partners.
  • Speakers describe delayed market data, filings, earnings transcripts, news, and private-company information; access remains product- and license-dependent.

Choose the smallest extension that fits

  • A skill packages reusable instructions, examples, and domain knowledge.
  • A plugin can bundle skills, tools, authentication, and MCP connections, the standard links between agents and external systems.
  • A plugin extension adds a custom interface within the host application.
  • Use existing presentation or spreadsheet templates to specify the output your team expects.

Check which surface supports the feature

  • The speakers place the built-in financial-data product in ChatGPT Work’s web interface at recording time.
  • They explicitly say its data access is unavailable through the Codex command line.
  • Their custom interface extensions are desktop-only in this workshop; web support is described as forthcoming.
  • Custom internal-data integrations and built-in financial-data access have different deployment and permission boundaries.

Adapt the context before paying for a bigger model

  • Explain the task, inputs, conventions, and expected artifact inside the skill.
  • Measure whether cheaper workers can handle bounded subtasks while a stronger model coordinates them.
  • Fabbri’s performance percentages are informal observations without a supplied evaluation set.
  • Steffmann warns that too many overlapping skills create maintenance problems and can slow processing.
  • Retest shared skills when models change, and consolidate repeated team workflows into a smaller managed set.

Keep orchestration and records explicit

  • Their research example uses a planner, parallel researchers, and a critic that requests more work when findings conflict.
  • Dom describes moving an existing research application into a plugin so users can inspect and customize the workflow where they already work.
  • Local files can preserve working state; an external application or database can remain the authoritative record.
  • In a proxy-voting example, findings, dates, and sources flow back to an external system through tools.

Keep the enterprise security layers

  • Reusing an approved host can simplify distribution, authentication, and policy management.
  • Administrators still need endpoint and network controls, local workspace restrictions, tool permissions, and approval policies.
  • Steffmann explicitly says developers must protect external tools themselves.
  • Fast model output does not remove laptop compilation time, tool execution, or the need for remote compute.

Evaluate the workflow before optimizing it

  • Fabbri says evaluation requires engineers and subject-matter experts to define success together.
  • Start with simple checks; add model-based judges only when the task calls for them.
  • Decompose vague one-to-five grades into explicit yes/no questions where that produces clearer criteria.
  • His goal-mode example iterates toward a measured target; it does not demonstrate independently validated research quality.
  • Retain source-linked execution records alongside the final artifact when decisions need review.

Caveat: this is a vendor-led workshop with live setup delays and illustrative builds. It does not establish investment performance, production readiness, or automatic regulatory compliance.

“Start simple and then add complexity.” — Stefano Fabbri, 1:40:45