Role-Specific Agents Move AI From Prompting Into Financial Services Workflows

Lee Spacagna framed enterprise AI adoption as having developed along two paths: bottom-up use of ChatGPT and Codex by employees, and top-down AI systems built into products, customer service, client advisory, and operational support. The gap, in his view, sits between those two modes: automation at the level of teams and departments.

That is the business consequence OpenAI is attaching to ChatGPT workspace agents. Agents, in this framing, are not chat interfaces for asking questions, but AI systems that can be delegated meaningful tasks and can use the tools employees already rely on: email, calendars, productivity apps, chat, documents, and business systems.

A chatbot answers inside the conversation. A workspace agent, as presented here, is supposed to work across the places where the job is already happening. The example was a “Chief of Staff” agent: a role-specific coworker intended to coordinate work, track priorities, prepare meetings, and keep a team moving.

Spacagna positioned workspace agents as an evolution of custom GPTs. The added layer is a builder that brings applications, skills, instructions, and deployment into one place. The claim is not merely that a user can configure a helpful assistant, but that a business user can build and modify a role-specific agent without code or prompt-engineering expertise.

Useful enterprise work is rarely contained in one system. Many teams do not lack tasks to automate; their work is spread across meetings, documents, email, chat, and business applications, while the decisions depend on context scattered across those systems.

The Chief of Staff demonstration showed the operating model. From ChatGPT, Spacagna selected a Chief of Staff template from a set that also included sales assistant, task management, data analysis, and knowledge search. The template already contained instructions, tools, and starter capabilities. He then connected Microsoft Outlook Calendar, Microsoft Teams, and Outlook Email so the agent could inspect calendars, read relevant communications, and post into a team channel.

The builder used another agent to write and update instructions. The user did not need technical skills. In natural language, Spacagna asked for the agent to run every day at 9:00 AM, review the day’s meetings and connected applications, inspect emails that arrived overnight, and generate a daily brief before meetings began.

The starter prompt “Prepare today’s brief” made the workflow concrete. The visible instruction asked for a concise daily brief using available calendar, email, chat, notes, and documents; it asked the agent to highlight priorities, decisions, blockers, and follow-ups; and it directed the output to the Office of the CFO team, inside the Daily Prep channel.

The permission boundary was part of the product behavior. On its first run, the agent asked for approval before posting to Teams. After permission was granted, the generated “Daily Chief of Staff Brief” appeared in Microsoft Teams in the Daily Prep channel. The pattern is configuration through natural language, access to work systems, scheduled execution, and action in team channels with approval where required. The output appears where the team is expected to use it.

Lee Spacagna extended the Chief of Staff agent from a daily briefing tool into a pre-meeting research assistant. The operational problem was familiar: the team was moving from meeting to meeting without enough time to prepare. The desired capability was for the agent to assemble context before each meeting: who would attend, the latest relevant information, and the meeting’s goal.

To support that work, he connected Microsoft SharePoint for company information and shared notes, and Salesforce for CRM and customer context. The same builder could connect other everyday applications or custom internal applications.

The next layer was “skills.” Spacagna described skills as snippets of information and instructions for critical tasks. Their purpose is to capture tribal knowledge and conventions that otherwise live in people’s heads, then turn them into repeatable workflows.

The Chief of Staff agent already had skills for chief-of-staff behavior and final brief formatting. Spacagna added a meeting-prep skill that specified how the information should be structured, which information mattered, what sources should be used, and where the resulting brief should be posted. He then used natural language to update the agent’s instructions so it would read from Salesforce and SharePoint when needed, generate quick meeting briefs in ChatGPT, and expose the new capability as a starter prompt.

The updated agent showed a third starter prompt: “Prep my next meeting.” When invoked, it gathered context for the next meeting from the connected sources, including Salesforce and SharePoint, and structured the brief according to the meeting-prep skill. The final brief shown in ChatGPT included sections such as “Who’s in the room,” “What this is about,” “Useful context,” “Suggested focus,” “Recommended talking points,” and “Next Steps.”

Spacagna connected the pattern to his own workflow. He said he personally has an agent that runs every day, checks overnight emails, important business updates, commitments he made on Slack or calls, and context from transcripts. The result, he said, is that he gets the first hour of his day back: he arrives to draft emails already prepared with relevant context, reviews them, approves them, and sends them.

Lee Spacagna argued that the Chief of Staff agent was only one instance of a broader pattern. The same approach, he said, can be applied across a financial services organization through role-specific agents.

The larger claim was that the opportunity is not a single automation project. It is a new operating model in which every team can create or deploy agents tailored to its workflows. Those agents take manual work off employees’ plates and help the business move faster by operating with the context, tools, and conventions of the role.

That framing creates the management problem. If one team can create a useful agent, the next question is what happens when an institution has thousands of them. Spacagna used that question to introduce Frontier, OpenAI’s platform for deploying and managing agents at scale.

Lee Spacagna described Frontier as a platform for deploying and managing agents across an enterprise. It connects to systems that are usually siloed, including data warehouses, CRM, internal applications, and systems of record. The goal is to give AI coworkers the same shared context that teams already rely on.

The architecture shown placed interfaces at the top: ChatGPT Enterprise, Codex, and business applications. Beneath them sat an agents layer containing customer-built agents, OpenAI agents, and third-party agents. Frontier formed the platform layer underneath, covering evaluation and optimization, agent execution, business context, and connection to systems of record. A later version of the same diagram populated the agent layer with examples including KYC and onboarding, due diligence, financial analysis and modeling, and memo and presentation agents.

That architecture shifts the product claim from individual agent creation to enterprise management. The agent builder shows how a team might create a role-specific workflow. Frontier is presented as the layer that governs execution, context, evaluation, and optimization when those workflows multiply across the institution.

Spacagna said agents on Frontier can reason over data, run code, use tools, and take actions in a governed environment. He also said the system improves as agents work: they learn from interactions, evaluate performance over time, and get better the more they do, “just like human workers in the business right now.”

Organizations can already build in ChatGPT, Codex, and the API, while OpenAI wants to make it easier to deploy out-of-the-box agents, plugins, and skills designed for financial services workflows. Purpose-built agents, in this account, plug directly into work and handle repeatable processes with less lift and less customization. With those foundations in place, Spacagna said, organizations can delegate more workflows to AI over time.