AICoP: Agentic AI with OpenAI

The September session closed with the promise of an industry partner, and October delivered one. Parixit Dave, Senior Director for AI and Emerging Technologies, welcomed two guests from OpenAI to what he called the community’s pre-Halloween edition. Chad Pasha works with Columbia as part of OpenAI’s education team. Hasun Khan, an Applied AI Architect, joined from the West Coast. Chad explained that OpenAI’s education team exists because so many people log in to ChatGPT to learn and build skills, and he noted the team had been on campus the day before with the SPS AI Lab.

What followed was less a product tour than a working method. Hasun spent the hour on one question: how do you get from asking a chatbot questions to delegating real, repeatable work to an agent, without losing control of the result?

Context Is the Slow Part

Hasun opened with a poll. Which takes longer: gathering the context needed to make a decision, or making the decision itself? The room’s answer matched what he said OpenAI has found internally. People are generally good at weighing information and deciding. The slow, messy part is collecting what they need from email threads, documents, calendars, and data that only exists in visual form. That makes context gathering the best place to apply AI. The decision, he stressed, should stay with a person.

Five Parts of an Agent

From there he broke every agent into five components. The first is a goal and instructions: a one-sentence statement of why the agent exists, plus the steps it takes to get there. The second is context, which should include only what the task actually needs. The third is the model, which determines how the agent thinks. Hasun contrasted a more capable model that questions its own scope and catches discrepancies with a faster one that keeps the work moving. The fourth is tools, the applications the agent can call. The fifth is human review.

That last part was the thread running through the entire session. “In most instances, you’ll find the AI tool gets you 80, 90 percent of the way,” Hasun said. Human review is how you get the last 10 percent. He recommended three checks on anything an agent produces: whether the evidence matches the sources provided, where the agent flagged uncertainty, and whether the actions it proposes stay within the original scope. He also shared a rule borrowed from another campus: use AI with others the way you would want it used with you, and don’t add AI slop to the world.

Two Demos, One Pattern

The first demo was meeting prep, the task Hasun said nearly everyone starts with. Working in draft-only mode, ChatGPT combined a calendar invite, a course project brief, and a student team’s notes into a brief for an advisor check-in. The brief did what a good assistant should. It flagged new questions and a proposed feature in the team’s notes that fell outside the agreed scope, and it caught a date conflict between the brief and the notes. It left both for the human to resolve.

The second demo moved to data. Using sample financials from a fictional company OpenAI uses for demos, ChatGPT reconciled a monthly forecast against actual results. It built an Excel workbook with a bridge, recovery evidence, and release gates, then turned the same analysis into a slide readout. An unsigned item was flagged for follow-up with its owner, and the output stated plainly that a recommendation did not authorize release. Hasun said campuses he works with apply the same pattern to enrollment trends, space allocation, and procurement spending, including staff who had never done data analysis before.

From Workflow to Skill to Schedule

The more durable lesson was how to spot work worth automating. Hasun walked through the five components again, this time asking which parts repeat each time. An advisor meeting with dozens of student teams sees the same kinds of briefs, notes, and calendar slots every time. A monthly reconciliation follows the same method with new numbers. He also counts how often he does a task by hand each week or month. Not everything qualifies. When students ask him for career advice, he said, “I’m not going to throw that to an agent,” because the task depends on personal judgment and experience.

Once a pattern is clear, it can become a skill: a reusable set of instructions that ChatGPT stores and manages, so the user only has to supply fresh inputs. Hasun built one live through conversation, without writing code. He then opened an advising-request triage skill developed with another campus and updated it with a few sentences of prompting to add color-coded urgency. Skills can be exported and shared with colleagues. The next step up is scheduling a skill to run on its own, such as every morning against an inbox, with the results delivered by email.

What’s Coming

Hasun closed with a preview of Dots, a personalized agent that runs continuously inside ChatGPT and handles background work like meeting briefs and draft follow-ups. He said OpenAI uses Dots heavily internally. Dots is not available at Columbia yet, and neither are plugins, which remain under review. Hasun was clear about this up front and said OpenAI is working with CUIT to bring these capabilities to campus. The presentation itself was a demonstration of where things are headed: an interactive website that Hasun, a data engineer by trade, built entirely by prompting in ChatGPT.

The Q&A

Questions centered on practical concerns. How long does setup take? About four to five minutes for a first version, roughly an hour to harden a workflow for daily use, and one to two weeks to test something across a department. How do you keep source checking from eating all the time AI saves? Tell the agent its job is to gather, not decide. Have it list every assumption and identify who published each external source. Ask it to produce a source sheet that flags the weakest sources. And what is the difference between ChatGPT’s Chat and Work modes? Chat is for questions and quick answers. Work is for reaching external resources and building things like decks, code, and applications.

Takeaway

The fall AI Literacy program is built on the idea that fluency comes from seeing tools do real work. October’s session showed what comes after fluency. The step from using a tool to delegating to it doesn’t depend on a new product. It depends on a habit: name the goal, scope the context, choose the tools, and keep a person at the checkpoint. “The idea is to create a tool that works alongside you,” Hasun said. That framing fits the team’s approach to bringing capabilities like Dots to Columbia: carefully, with review, and shaped by what the community says it needs.

Hasun’s interactive presentation will stay online. Here is the link. Questions, use cases, and consultation requests go to [email protected], and requests for new features and connectors should come to the AI & Emerging Technologies team rather than to vendor reps.