IESE AI Club · Applied AI Series

From chat to work.

Intro to Claude & the OpenAI ecosystem.

2 hoursHands-onCowork · ChatGPT · CodexNo coding required
IESE AI Club1 of 4 · what it is

Building leaders fluent in AI.

The club, in one line

IESE's student-led community for people who would rather build with AI than just talk about it. Practical AI for MBAs and business leaders.

Open to every IESE MBA and Executive Education student. The learning materials, playbooks and ecosystem map are public and free; membership and workshops are for IESE students.

Four pillars
  • Learn & explore · AI strategy, machine learning, analytics workshops.
  • Connect & network · industry leaders and alumni.
  • Build & apply · case competitions, consulting projects, prototyping.
  • Impact & ethics · responsible adoption and what it does to people.
Notes: Rajas presents slides 2 to 5, five minutes total. Every line on these four is from ieseai.com as of 13 Sep 2026.
IESE AI Club2 of 4 · what we do

Build, test, ship.

Learn
  • The learning path · eight steps, from how a model predicts to shipping a real app. Three doors in: Curious, Operator, Builder.
  • The Gym · skill modules with provable artifacts: prompting, context, agents, memory, evals.
Do
  • Seven applied workshops · this is the first. Hands-on, not lectures.
  • A December hackathon.
  • The Industry Fellowship · September to mid-December. Next slide but one.
Ship
  • Nine role playbooks · product, sales, marketing, VC, operations, consulting, founder and more.
  • Member projects · an elective-bidding planner, an anti-AI writing skill, the club's design system.
  • The ecosystem map · who makes the models, who builds on them.
Notes: Point at the room: today is workshop one of seven. The projects are all on the site under Projects, built by people on the next slide.
IESE AI Club3 of 4 · who we are

Eleven MBA candidates. Most of us aren't engineers.

Rajas Bhagat · President
Emi Takeda · Corporate Relations & Finance
Naisarg Gohil · Chief of Staff
Khyati S. Kumar · Marketing & Operations
Jose Cimadevilla · Marketing & Operations
Hernan Espinosa · Community Relations
Hanif Ramadhan · Community Relations
Muhammad Maula Arif · Ecosystem Management
Radu Pană · Ecosystem Management
Josh Clements · Knowledge & Skills
Zachery Ballenger · Knowledge & Skills
Notes: Names and roles from the About page. Rajas names whoever is in the room; the bios are on the site, not here.
IESE AI Club4 of 4 · the Industry Fellowship

Pick an industry. Build for it.

How it works
  1. 5+ coffee chats with people already using AI in your industry.
  2. 3 use cases, judged on value and feasibility.
  3. 1 working MVP for the one that holds.
  4. Publish it for the next cohort.

September to mid-December. No technical background needed.

Data & AI Club Kick-Off
QR: Data & AI Club Kick-Off

Scan to register.

Apply · Class of 2028
QR: fellowship application

Five minutes to fill in. Intake is open now.

Join the club
QR: club sign-up

groups.iese.edu · DataAI

Notes: Leave this up while people scan. What fellows walk away with, if asked: a portfolio project, industry-specific AI expertise, a wider network, an interview-ready story, a published piece.
Data & AI Club · Workshop 1

Anthropic, OpenAI, and You.

What they are, and how to make them work for you.

Where I started

I opened Claude.ai for the first time in March. I'm not a developer, tech guru, or computer science major.

My takeaways
  • Trial
  • Error
  • Skepticism
  • Curiosity
Notes: Warm, not a victory lap. Set the contract in one line: every fifteen minutes you touch a tool. Say the plan-truth early and plainly: today runs on the desktop app and a paid plan, because the tools are the point. Anyone on the free plan does Task 1 in a claude.ai Project and watches the rest; say it once, don't apologize for it.
About Me

Zach Ballenger. Second-year MBA.

Helicopter cockpit over the sea off Okinawa
Before IESE
  • From the US
  • 12 years military
  • Pilot / diplomatic affairs
  • Japan, Bahrain, BCN
The cohort at the end of the New York term
Here
  • Responsible Business Concentration
  • Co-Pres VBC
  • Data & AI Skills
With AI
BSP app
BSP app
Habla Ya
Habla Ya
Pocket Coach
Pocket Coach
Reading Room
Reading Room
Study guides
Study guides
Lorien
Lorien
Notes: Thirty seconds, no more. The point is "I'm one of you, I just went through the door first". Every tile on the right was built in chat or Cowork-style delegation, no computer-science degree involved; say that in one breath and move on.
Ask the roomhands up

AI in the Room

1 · Who regularly uses AI?

What do you use it for?

2 · What do you wish it could do for you?

What is your biggest hurdle?

3 · How often do you check the output?

How do you validate it?

How I started

Testing the waters

1 · A trip

A couple of years ago, I asked Google's chatbot to plan a month of travel. Great general info, terrible planning. Multiple - "Plan an itinerary for 29 days" "Sure, here is your 29-day itinerary." (it was never 29 days...)

2 · A technical note

I dropped an operations note in. It gave me a summary and a working page that ran the formulas. Opening it now.

3 · A BSP App

I wanted to make a flashcards or a study tool for BSP.
BSP-APP.netlify.app
Invite Code: IESEenBSP

Notes: Open the queueing calculator live here (assets/queueing-calculator.html next to this deck; the original was in Downloads). Source note: "Queueing: Waiting times due to random variability" (OITN-6-E), Operations Management session 12; the chat was 2 Mar 2026, one sentence plus the PDF, and the tool came back with the note's own airport example already loaded. The point for the room: the prompt was one line; the note carried the formulas. Not a screenshot. Card 3 names the BSP app by choice; the invite code is on screen, so anyone in the room can use it. Seventy confident mistakes in that app is the pivot into the checking habit.
About ten minutesget running

Getting set up.

Do this
  1. Install the Claude desktop app (claude.com/download), sign in. Paid plan: you'll see three tabs, Chat · Cowork · Code.
  2. Make a folder on your desktop called Career Fit. Put your CV (PDF) and the packet file in it. Every task today lives in that folder.
  3. Have a ChatGPT login ready, and be signed into LinkedIn (Task 2 reads your job alerts).
Having Trouble?
  • Ask a helper, or the person next to you.
  • Or ask Claude or ChatGPT how to do that exact step: paste the step and what happened. That's not cheating, that's the skill.
  • Still not running: use claude.ai in the browser for Task 1. We fix the install in the break.
product steps verified 7 Sep 2026 · re-check before the session
Notes: Relaxed. It takes the time it takes. You and the helpers walk the room; anyone already running helps the next person. The first hiccup is the first lesson, say that out loud: ask the AI how to do the step. Free plan in the room: Chat and Code tabs only, Task 1 in a claude.ai Project, then watch. Say it once.
Key ideas1 of 5

It predicts. It doesn't know.

The whole trick
  • A model is a very good guesser of what words come next. That's it, scaled up.
  • It has no separate part that checks the guess against the world. So a right answer and a wrong one sound identical.
  • Which is the whole day in one line: reality is the judge. Your job is to check.
Two words you'll hear
  • Context is what it can see right now: your message, the files you gave it, the tools you connected. Nothing else.
  • Tokens are the meter. Everything it reads and writes costs some.
Notes: Two minutes, no diagrams. The point of this slide is Context, which sets up the next one. Skip anything about training, parameters, or vendors.
Key ideas2 of 5

It only knows what you put in front of it.

What it can see
  • What you type or paste into the chat.
  • A file you upload, or a folder you connect.
  • An app you plug in: your email, your Drive, your calendar.
  • A website, only when it has a browser.

Nothing else. Not your laptop, not your inbox, not the internet, unless you handed it over.

The catch

It won't say "I can't see that". It answers anyway, from what it guesses the thing looks like, in the same confident voice.

So when it tells you about your file, your inbox or a web page, first ask: did I actually give it that?

Notes: Story if there's time: I asked where a settings button was. It told me, confidently, five times. The button was never there, because it had never seen the page. The version that could see the page found it in one go.
Key ideas3 of 5 · the map

Two ecosystems. Five tools. Defined by what they do.

Anthropic · Claude
  • Chat · you drive. Drafts, ideas, quick answers.
  • Cowork · you delegate. Reports, research, files, scheduled work.
  • Code · you program the autopilot. Real automations, in a folder.

One desktop app, three tabs.

OpenAI
  • ChatGPT · the chat most of you already use. Strong daily driver.
  • Codex · the coding agent, a mode inside the ChatGPT desktop app. The Code-tab equivalent.

Chat is enough for most things. I built my first app in ordinary chat.

Notes: One question picks the tab: how often does this need to happen? Once, a bigger one-off job, or every week. Once → chat. A job → Cowork. Every week → Cowork scheduled, or Code if it needs your local files.
Key ideas4 of 5 · the loop

You direct. It executes. You verify.

1 · You direct

Say the outcome, the scope, and what done looks like. The clearer the brief, the less it guesses.

2 · It executes

Let it run, on something you can afford to lose. It asks when it should; answer short and keep going.

3 · You verify

Open the result. The AI saying it worked is not evidence. What you find here becomes the next brief.

Three steps, every time, at every size. Installing the app is the first rep of it. Building a company on it is the same three steps.

Notes: Task 1 and Task 2 are direct and execute. Task 3 is verify, done properly. Say that mapping out loud so the next ninety minutes have a shape. The four reference pages that follow are a glance, not a lecture. The loop stays true through any change in a project; the tools change, the three steps don't.
Key ideas5 of 5 · six words

Six words you'll hear today.

Skill · a recipe card. Teaches it how you like something done: your tone, a format, a checklist.
Connector (MCP) · a power outlet. Plugs it into Gmail, Drive, a calendar. MCP is a phone line between the AI and your data.
Plugin · a starter kit. Skills and connectors bundled, installed at once.
Scheduled task · a job that runs on its own, on a schedule. In the cloud, or on your laptop when it needs your files or your logins.
Project · a chat with a memory: files you upload once and it keeps reading.
Agent · the AI pointed at a job, with a role and an end state you describe. That's all it is.
Notes: Six cards, four seconds each. The four reference pages after the loop carry the tools and the rest of the words; don't define anything beyond those today.
Toolsthe connectors I actually use · each signs in on its own page, never in the chat

Connectors I'd vouch for.

Your day
  • Gmail · read, search, draft. Never sends without you.
  • Google Calendar · your week, in the answer.
  • Google Drive · the files you already have.
  • Gemini · Google's model as a connector, for a quick second opinion.
Your work
  • NotebookLM · your readings, answered from the readings only.
  • Obsidian · your notes vault, connected so it reads and writes your own markdown.
  • Canva · Figma · a deck or a design, from a brief.
  • The browser · built into Cowork. Most sites you can log into, it can read. Banking and other high-risk sites are blocked.
  • Web search · the internet, when you switch it on.
Your build
  • Codex · OpenAI's coding agent, connected to work in your folder.
  • GitHub · where the code lives and its history. Workshop 5.
  • Netlify · puts the page online.
  • Supabase · the app's database and logins.
Notes: Thirty seconds. Every name here is one you've used yourself; nothing on this slide is an endorsement by hearsay. The one rule: a connector opens the service's own sign-in page; if anything asks for a password or a key inside the chat, stop.
Toolswhat I built with · none of it needed today

Oh, the websites you'll go!

NotebookLM · Google's study notebook. Upload the readings, it answers only from them, and it can turn them into a podcast. How I did case prep in year one.
Netlify · puts a web page online from a folder, free. The BSP app and Habla Ya live there. Drag the folder in, get a link.
Supabase · a database in the cloud with a login system. It's the app's memory: who you are, what you answered, your streak.
GitHub · cloud storage for repos, a website that holds project folders online so you can back up and share them. Private repos are private. A backup, not a billboard.
ElevenLabs · turns a script into a natural-sounding voice — realistic narration and voice clones from plain text. The audio layer when a project needs to talk.
UptimeRobot · watches whether a site you shipped is still up and pings you the second it goes down. Free, and you hear it from a robot before a user does.
Notes: Thirty seconds, four seconds a card. These are just names you'll see around; the point is they stop being scary, not that you learn them today. (The password/key-in-a-chat rule still lands on the Tools slide and the "API" card, so it stays covered.)
Filesthe ones you'll meet inside a project

Oh, the files you'll meet.

.claude / .codex · a folder, not a file: the tool's backpack for a project. Settings, skills, rules and memory live inside.
.md · a plain-text document with light formatting. The portable memory you hand to Claude. Opens in Notepad.
.html · a web page, the file a browser turns into something you can see and click. The calculator and this deck are one file each.
CLAUDE.md · standing orders. A text file Claude reads at the start of every session so you stop re-explaining your rules.
Repo · a project folder with a built-in history, every change saved as a point you can go back to. Your undo button for the whole project.
.gitignore · a list of files a repo should never save or share: secrets, junk, anything private. If a file's name is in here, it stays on your machine.

How you hand it a file: in chat, upload it. On your laptop, right-click the file, Copy as path (Mac: hold Option, Copy as Pathname), paste that into the chat. I found that by accident.

Notes: Read the bold, one line, move. If someone asks "do I need to learn these": no. You need to recognise them when Claude names them, so you can open the right one and look. The copy-as-path line is the most useful thirty seconds on the slide.
Wordsyou'll hear these, so here they are

Six words, plain English.

Model vs app · Claude is the app. The model is the engine inside it, and engines have version names.
Prompt · what you type. A good one is a brief: role, context, task, what done looks like.
Hallucination · a confident guess delivered as fact. It doesn't feel different from the inside.
Context window · how much it can hold at once. Long chats forget their own beginning.
API · a service window where one program asks another for something. An API key is the password for that window. Never in a chat.
Diff · a side-by-side of what's about to change, old next to new. You read it before you say yes.
Notes: Six cards, four seconds each. "Diff" sets up the Code tab at the end; "hallucination" is the what-it-can-see slide with a name on it.
Task 1 · 20 mineveryone · Cowork on your folder

Find out what fits you. Upload, answer, read the verdict.

Set it up (5 min)
  1. Cowork → point it at your Career Fit folder. It sees this, nothing else.
  2. Type: "Read the Career Vault Builder file and my CV. Give me the questionnaire, short version."
  3. Answer one line each. "Don't know" is an answer. Skip anything you don't want on a screen in this room.
Read, check, keep (15 min)
  1. It reflects you back, names your gaps, and proposes two or three paths you didn't say.
  2. Check one thing. One fact it stated about an organisation: location, entry route, pay band. Look it up. Did it hold?
  3. Keep it. "Write my Career Fit profile as one page. Save it as career-fit.md in this folder." Watch it ask before it writes. Task 2 reads that file.

Fast finishers: same folder, in the Codex app. Which one asked the harder question?

Notes: Say the privacy line once: the folder is on your laptop and your CV goes to your own account, nowhere else. Free plan: same steps in a claude.ai Project, and copy the profile into a file yourself. Walk the room. Bring two things to the debrief: one honest gap it named that stung, and one fact that didn't hold. At home the same folder builds the whole Career Vault; the debrief's task-for-home carries it.
Debriefwhat went wrong, and the moves

The Good, the Bad, and the Ugly

The moves
  • Sniff test First glance, be skeptical: does it look remotely correct?
  • Did it read your answers and CV correctly? Anything completely off, or worded wrong?
  • Ask Why: what made it draw that conclusion?
  • Clarify. Say what's wrong, what the issue is, what it should look like.
  • Tool Audits: Claude v. Codex, Claude v. Claude, Codex v. Gemini...
Task for home

Take the project and continue to add data, build your Career Vault. Ask AI to walk you through the process. Find a way to make the data useful to you.

Early AI success is often learning how not to, as much as it is learning how to.

Notes: The cost story if you need one more: I asked for a cost estimate and it was wrong. Fifteen separate times. Everything I stacked on it was wrong too. Not "off by fifteen times": wrong fifteen times.
Break

Break. Ten minutes.

Not running yet? Find a helper now.

Notes: Use the break to get Cowork open on your own screen with the practice folder already connected. Nothing should be set up live after the break except what the room is meant to watch.
Cowork, livewatch, then you do it

Cowork: you delegate. It asks first.

The folder: it asks first
  1. Same folder as Task 1. The folder is the permission slip: it sees this, nothing else.
  2. "Move my CV and raw notes into a subfolder called Inputs — but leave career-fit.md where it is." It asks before each move. I approve the moves; when it offers to move career-fit.md too, I refuse — that file stays put, the radar reads it here.
The internship radar
  1. Its own browser. I signed into my mail and LinkedIn once; it remembers. Never paste a password or a key into any chat.
  2. Scheduled → Create with Claude. Every Monday: open my inbox, find the newest LinkedIn job-alert email (or my CDC Symplicity digest), open and read every posting, rank them against career-fit.md. Runs on this laptop: the postings need my login.
  3. Already running since the break. Watch it open a posting and read it, then the finished table on the Scheduled page.
verified 13 Sep 2026 · re-check before the session
Notes: Start the radar run before the break so the finished table is ready; on stage, show it opening one posting live, then cut to the table. If the wifi or the sign-in sags, the recording is the fallback: name the file and where it sits once it exists. Say plainly that you work mostly in Code and this is Cowork through your eyes. The password line comes from doing exactly that once and rotating everything after.
Task 2 · 20 mineveryone · three doors

Build one thing that works while you don't.

The internship radar

Set a LinkedIn job alert from your profile's roles (the card walks you). Then Cowork, same folder → Scheduled → Create with Claude, paste the radar prompt, sign into LinkedIn once when its browser asks. It reads the alert emails and ranks them against career-fit.md.

A skill

Cowork → Customize → Skills → new. A tone, a format, a checklist you use weekly. Describe it in plain words, or record yourself doing it once.

The Code tab

Same folder, Code tab. "Build the internship radar as a page I run with one click." It shows a plan: approve it. It edits: read the diff before you say yes.

Pick the one you'd use this month. Two minutes stuck on a sign-in: switch doors. Task for home: whichever you built, run it Monday — the alerts keep coming on their own.

Notes: LinkedIn job alerts come from the same sender for everyone, so the prompt filters on that address; the alert may not have arrived within the session, so the thing they BUILD is the standing automation and the live run is on whatever digest exists (yours for the demo). The radar takes minutes to open every posting; Task 3 reads the result, however far it got. The Code door is for the curious: the plan and the diff, not a finished page. Anyone not hunting takes the skill door.
While it runswhat to do with the wait

It's going to ask, stall, or wander. All three are normal.

The moves
  • It asks a question. Answer it. Short is fine. That's the intake working, not a bug.
  • It stalls on a step. Paste the step and what happened, into any chat, and ask how to do that exact step.
  • It wanders off. Don't restart. Say what's wrong, what it should look like, and carry on.
  • It says "done". Don't believe it yet. That's the next twelve minutes.
Meanwhile

Open the first thing it produced and read it while the rest runs. The first mistake shows up early, and it's cheaper to catch now than after it's built on top of it.

The first time I let something run unwatched, it built forty minutes of work on one wrong number.

Notes: Walk the room while this is up. The radar takes a few minutes to open every posting; the free door is faster. Anyone stuck on the sign-in step: two minutes, then switch doors; the abandon rule is on the Task 2 slide.
Task 3 · 12 mineveryone · prove it works

Now prove it. Not "it said so". Prove it.

The moves
  • Sniff test the table. Do the top three look like jobs you'd want? If not, fix the profile, not the tool.
  • Open the top posting yourself. Does it say what the two lines say?
  • Ask why. "What put this one first?" If the answer is vague, it guessed.
  • Change one line. A city, a sector. Run again on the top three only. Did the ranking move the way it should?
  • Skill or Code door: run it on something fresh. Find the line that isn't yours. Change it, run again.
Task for home

Let it run every Monday. Each week, add one line to your profile from something you learned: a company you liked, a role you'd never take. Keep a short list of what it got wrong.

What would you have believed if you hadn't looked?

Notes: This is the whole session in twelve minutes. Collect two examples for the close: one that passed, one that didn't. Skill door, if anyone took it: run the skill on a fresh email, find the line that isn't your tone, change one line, run again. A LinkedIn alert email can take up to an hour, so a student may not have their own digest in class — that's fine: they verify on the demo run now and their own run later. What matters is the automation is built and proven, not that their inbox fired in the room. If a radar run hasn't finished, the check works on one posting.
Five minuteswhere this leads

Same intelligence, three amounts of rope.

Chat · you drive

One question, one answer. The queueing calculator from earlier was one sentence and a PDF in ordinary chat. Most days, this is enough.

Cowork · you delegate

A folder, a browser, a schedule. What you just built. It asks before it acts, and it runs while you're in class.

Code · you program the autopilot

The Code tab. It plans, you approve, it edits, you see exactly what changed. Every app on my bio slide came from here. Codex is the same door on OpenAI's side.

The folder you built today is where all three run. Same loop, more rope each step.

Notes: Demo only if time allows, pre-opened: the Code tab, plan, approve, diff, open. Drop it if Q&A needs the room. On Windows the Code tab may ask for Git for Windows for some features; if it does, install it, don't fight it here.
If you remember one thing

The AI sounds confident whether it's right or wrong. Reality is the judge. Your job is to check.

Clear the context

New chat. Paste only the output, not the conversation. Ask what's wrong with it. No memory of how it got there, no loyalty to it.

Internal audit

Same tool, opposite job: "argue against this answer". If it folds fast, it never had the goods.

External audit

A different AI. Paste answer A into tool B and ask what's off. Two confident voices that disagree is your cue to go look.

Notes: Slow down. This is the sentence they keep. Task 3 is this sentence made physical. Say the why once: everything today ends with one act of checking, not because the tool is bad, but because it is fast and literal and will never tell you which of its sentences to doubt. Then the three audits, one breath each.
Closethe one habit, at every size

You direct. It executes. You verify.

You direct

Say the outcome, and what done looks like.

It executes

Let it run on something you can afford to lose.

You verify

Open the result. "Done" is not evidence.

Notes: The last thing they see is the loop. Thank them, point at the take-home guide, open Q&A. Nothing here is homework.
IESE AI CLUB← → move · S speaker notes · F fullscreen · print for a PDF 1 / 19