Session 1 of 5
The AI Business Mindset
How to spot problems AI can actually solve
with Elizabeth Tran · Tuesday, July 21, 2026
Hi, I'm Elizabeth
AI Agent Builder & Automation Expert
- Technology Entrepreneurship @ Babson College
- Partnerships Lead at The Generator (Babson's AI Lab)
- I build production AI agents for paying clients
- I've helped 80+ small businesses adopt AI
- I've worked on partnerships with Anthropic, OpenAI, Cursor & more
Real agents I've shipped
AI Sales Agent — automatically follows up after discovery calls, qualifies leads, and drafts personalized emails while keeping context across every conversation.
Financial Dashboard Agent — connects 6+ bank accounts, calculates cash on hand, and pushes a daily report to Slack. Fully hands-off.
Newsletter Curation Agent — monitors sources, extracts insights, and drafts executive-ready summaries — saving hours every week.
$2,000+ early revenue
Real paying clients
80+ businesses consulted
5M+ TikTok views
🎮 Quick Warm-up
Chatbot or Agent?
You text a tool: "Book me a haircut for Saturday afternoon." It checks the salon's calendar, finds a 2pm slot, books it, and texts you a confirmation. Which is it?
It's an Agent. 🤖 It didn't just talk about booking — it took action in the real world (checked a calendar, made a booking, sent a confirmation). That's the whole difference: agents do things.
The core difference
💬 Chatbot
• Only responds to what you type
• No memory of past chats
• Can't do anything but talk
Ask: "What's our Q2 revenue?"
It says: "I don't have access to that."
🤖 AI Agent
• Understands a goal, not just a message
• Remembers context
• Uses tools — databases, email, APIs
Ask: "What's our Q2 revenue?"
It does: queries the database, formats it, posts it to your dashboard.
One-line memory trick: a chatbot talks, an agent acts.
🎮 Spot the Agent
Which one is the agent?
Same request, two tools. Tap the one that's behaving like a real agent.
B is the agent. A just hands you words — you still do all the work. B reasons, uses tools, and takes action to actually finish the job. When you sell to a business owner, B is what they'll pay for.
How an agent "thinks"
Every agent loops through the same three steps
🧠 Thought
"What do I need to do next?" The agent reasons about the goal.
→
⚡ Action
It picks a tool and uses it — "search the database for Q2 revenue."
→
👀 Observation
It reads the real result — "$150,000" — and decides what's next.
It repeats this loop — Thought → Action → Observation — until the job is done. That loop is the "agent."
🎮 Your Turn
What's the missing step?
An agent is restocking an online store:
🧠 Thought: "I should check which products are low."
⚡ Action: "Query inventory for items under 10 units."
❓ ???: "Returns: 3 products low."
🧠 Thought: "I'll reorder those 3."
🧠 Thought: "I should check which products are low."
⚡ Action: "Query inventory for items under 10 units."
❓ ???: "Returns: 3 products low."
🧠 Thought: "I'll reorder those 3."
It's an Observation. 👀 Reading the real result of an action ("3 products low") is what lets the agent decide its next move. No observation = the agent is just guessing.
Now the real skill: spotting the opportunity
Which business problems are worth building an agent for?
✅ Great fit
• Repetitive — happens over and over
• High-volume — many times a week
• Rule-based — follows a clear pattern
• Costs real time or money by hand
❌ Poor fit
• One-off or rare
• Needs deep human judgment or empathy
• Rules keep changing
• High-stakes with no room for error
The money question: "What boring, repetitive task is eating this business's time every single day?"
Where the opportunities hide
Real problems, by the kind of business
🏠
Real estate: 20+ inquiry emails a day asking the same questions → an agent answers FAQs and flags the serious buyers.
🍽️
Restaurants: phone rings all day for reservations & hours → a voice/chat agent books tables and answers 24/7.
💇
Salons & clinics: no-shows and manual scheduling → an agent books, confirms, and sends reminders.
📊
Agencies: hours spent chasing leads and logging them → an agent researches prospects and drafts outreach.
🛒
E-commerce: repetitive "where's my order?" tickets → an agent handles support and escalates only the hard ones.
🎮 Consultant Mode
Is this an agent opportunity?
A local gym owner says: "Every morning I spend an hour copying new sign-ups from 3 different forms into one spreadsheet."
Yes — textbook fit. ✅ Repetitive ✓ daily ✓ rule-based ✓ wastes real time ✓. An agent pulls from all 3 forms and auto-fills the sheet. That's a problem an owner will happily pay to make disappear.
🎮 Consultant Mode
And this one?
The same gym owner says: "I'm trying to decide whether to open a second location downtown."
No — and knowing this makes you a real consultant. 🧠 This is a big, one-off, judgment-heavy decision. An agent could gather data to help, but the call itself needs a human. Great consultants know what NOT to automate.
By Session 5, you'll be able to…
- Build a real AI agent from scratch
- Spot the business problems agents can solve
- Consult a business owner and scope their needs
- Deploy your agent live with a real tool
- Pitch and sell it to an actual business owner
The edge most people never get: you'll learn to build AND sell. That's what turns a cool skill into income.
Next up: we start building
Session 2 — Thursday, July 23 · 1:30pm
• The anatomy of an agent: brain, memory & tools
• How API keys and MCPs connect agents to the real world
• I'll show you a conversational restaurant agent I built in Claude Code — with a real voice 🎙️
• Live n8n demo + a library of workflows to explore
• You'll pick one — and getting it running is your first mission 🚀
• How API keys and MCPs connect agents to the real world
• I'll show you a conversational restaurant agent I built in Claude Code — with a real voice 🎙️
• Live n8n demo + a library of workflows to explore
• You'll pick one — and getting it running is your first mission 🚀
Set these up before Thursday (free, ~10 min):
Claude.ai · n8n
Claude.ai · n8n
Let's build something
Questions?
Then we get our hands dirty.
Session 2 of 5
Give Your Site a Brain
Agents, API keys & MCPs
Thursday, July 23, 2026 · 1:30–3:30pm
🎮 Warm-up · Recap
Quick check from Tuesday
A tool reads your inbox every morning, figures out which emails are urgent, drafts replies, and flags the tricky ones for you. What is it?
An agent. 🤖 It has a goal, it uses tools (your inbox), and it acts — it doesn't just talk about your email. Remember the rule: a chatbot talks, an agent acts.
Your site is empty. Let's fix that.
Not with filler — with something real.
You haven't built anything yet. That's completely normal — you're 15, you just started.
So today, you build project #1 — and it goes on your site.
By the time you leave, your portfolio doesn't say "here's a kid who likes stuff."
It says "here's something I built that works — go try it."
It says "here's something I built that works — go try it."
🎯 You choose
What are you actually into?
You might think you don't have expertise yet. Wrong. Whatever you're into — a game, a sport, music, cars — you know way more about it than most adults do. That's the raw material.
🎁 The Recommender
Someone says what they already like — your agent recommends what to try next.
🧠 The Quiz Master
Quizzes visitors on your topic, keeps score, roasts them a little when they're wrong.
💡 The Explainer
Answers beginner questions about your topic — like you'd explain it to a friend.
The real test isn't "did it run." It's: can you send it to your friends and do they think it's cool?
Concept 1
API — how software talks to software
You're at a restaurant. You don't walk into the kitchen and cook your own food. You tell the waiter what you want — the waiter takes it to the kitchen and brings the food back.
🧑 You
Your website makes a request
→
🧾 The waiter
The API carries the message
→
👨🍳 The kitchen
Claude does the thinking
→
🍽️ Your food
The answer comes back
An API is just a messenger between two programs. That's it.
Concept 2 ⭐
API key — your agent's ID badge
Right now you're signed into Claude Code on this laptop. That works because you're sitting here.
But your website isn't you. It runs at 3am while you're asleep, for a stranger who found your portfolio. It can't open a browser and log in.
So it needs its own ID badge. That's an API key.
sk-ant-api03-••••••••••••••••••••••••••••••
⚠️ Non-negotiable
Three rules for keys
People break these and it costs them real money.
🚫
Never paste it in your code
It goes in a .env file — never typed directly into a file you'll share.
🙈
Never commit it to GitHub
Add .env to .gitignore. Bots scrape public repos for keys within minutes.
💳
It spends real money
Anyone holding the key can run up charges. Treat it like a debit card number.
🎮 Spot the Mistake
Which one gets you in trouble?
You just finished your project and you're about to share it. Which move is the dangerous one?
B — and it's the most common beginner mistake there is. 💥 The moment that file hits GitHub, automated bots find the key and start spending. Developers have woken up to hundreds of dollars in charges this way. Keys live in .env. Always.
Concept 3
MCP — the universal adapter
Your agent can talk. But how does it do things — read a calendar, check a database, send an email?
😩 The old way
A developer writes custom code for every single tool.
Google Calendar? Custom code.
Gmail? More custom code.
Slack? More.
It never ends.
🔌 MCP
Model Context Protocol — one standard way for AI to plug into tools.
Think USB-C. Before it, every device had its own weird cable. Now one plug fits everything.
MCP is USB-C for AI.
👀 Watch this
Let's see it use a tool
I'm going to connect Claude Code to a tool and ask it something it can't know on its own.
🧠 Thought
"I need real information I don't have."
→
⚡ Action
Reaches for the tool through MCP
→
👀 Observation
Reads the real result, decides what's next
Same loop from Tuesday. That loop, plus tools, is what makes it an agent.
🎙️ Demo time
The restaurant agent
The one I promised you Tuesday — you can talk to it out loud.
Go ahead: ask it about the menu. Try to book a table. Try to confuse it.
A restaurant owner would pay real money for this — no more missed calls during dinner rush. That's the business side we get into in Session 5.
Nothing magic — just today's four pieces
- API key — let my code talk to Claude and to the voice service
- System prompt — told it "you're a host at this restaurant, here are the hours and the menu"
- Tool use — let it actually check availability and book the table
- Agent loop — thought → action → observation, running until the job was done
Every impressive agent you'll ever see is these same pieces, stacked. Now you build one.
🛠️ Your build
Let's build the one you picked
- Set up the key — create .env, add it to .gitignore, and tell me why
- Wire up the call — get your site talking to the API
- Give it its brain — this is where your knowledge of the topic goes in
- Test it and break it — find out where it falls apart
Step 3 is the real work. Anyone can wire up an API — not everyone knows what makes a good answer in your topic. That's your edge.
You're driving. I'm coaching. Errors are part of the plan.
🧠 Important
When your agent makes things up
Ask it something it doesn't actually know about you. Watch what happens.
It will often invent an answer — confidently. That's called a hallucination, and it's the #1 way agents embarrass people in front of clients.
We fix this in Session 5 with guardrails. For now, just notice it happens.
Knowing your agent's weak spots is what separates someone who ships from someone whose stuff breaks in front of a paying customer.
⭐ The payoff
Now put it on your site
Add a section called Projects. Write three lines:
- What it is
- What you built it with
- A button to try it
This morning your site was a page about a kid.
Now it's a page with something you built that a stranger can use.
Now it's a page with something you built that a stranger can use.
Everyone else your age has a résumé that says "hard worker." You have a thing people can click.
🏠 Your mission
Homework
Required:
- Get it genuinely working end to end
- .env stays out of GitHub — this part is graded
- Show it to a friend and watch them use it. Write down what confused them.
Stretch — pick one:
- Make it remember earlier messages
- Give it a real tool it can use
- Style it to match your site
- Make its personality so specific your friends can tell it's yours
The friend test is the real assignment. Watching someone else use your thing teaches you more in 2 minutes than an hour of testing it yourself. That's literally what I do with paying clients.
Monday: from personal project to paid work
Session 3 — Monday, July 27 · 1:30pm
You just built an agent for yourself. Monday we build one for a business — the kind of thing someone actually pays for.
Here's the thing you'll notice: it's the same shape as what you built today. Same key, same loop, same tools — pointed at a problem that costs a business money.
Nice work today
Questions?
You just shipped an agent. That's real.
Session 3 of 5
Building the Restaurant Agent
Same skills. A real business problem.
Monday, July 27, 2026 · 1:30–3:30pm
👋 Recap
Show me the game recommender
Is it still working the way you left it? Did you show it to a friend yet — what happened, what confused them, what did they think was cool?
Whatever broke, whatever confused them — hold onto that. It's exactly the kind of thing a business owner would tell you about their agent, too.
🔧 Quick fix, live
Let's make it a little better
What's one thing that bugged you when you tested it — a recommendation that felt off, something it didn't know, a tone thing?
We're going to fix it right now — not by rewriting code, but by rewriting the instructions.
That's most of what "improving an agent" actually looks like.
⭐ Today's shift
Same machine. Different owner.
Your game recommender: someone says what they want, it thinks for a second, it gives a good answer.
A restaurant has the exact same shape of problem. Someone calls, wants a table, and somebody has to answer, check availability, and book it.
Swap "games" for "reservations." Swap "recommend" for "book." Same machine.
🎮 Games → 🍽️ Reservations
Recommend → Book
Nobody pays you for game recommendations. A restaurant owner pays real money to stop missing calls during the dinner rush.
📋 Quick word
A "spec" is just a plan
That's really it. Deciding what the thing should do before you build it, instead of figuring it out as you go.
Building without a plan means building the wrong thing — then redoing it. A plan first is just faster.
And here's the good part: you don't have to write it yourself. You're going to talk it through with Claude — out loud, like explaining an idea to a friend.
💬 New move
Let Claude interview you
Open a new chat. Tell Claude what you're building — just talk normally.
"I want to build an AI agent for a restaurant. It should answer questions about the menu and hours, and help book a table. Can you ask me some questions so we figure out exactly what it needs to do?"
Claude will ask you things like:
- What's the restaurant called, what's the vibe?
- What should it answer — hours, menu, dietary questions?
- Should it actually book a table, or just say if one's free?
- What should it do when it doesn't know something?
That back-and-forth is the spec. It's just alive instead of written down.
✍️ The payoff
Now have Claude write the prompt
"Great — can you write the system prompt for this agent based on everything we just talked about?"
Claude turns your whole conversation into the actual instructions your agent will run on.
This is genuinely how I do it with real clients. I don't write a spec alone in a room — I talk it through with the business owner and use Claude to turn that conversation into working instructions. You just did the real thing.
🎙️ Same demo, new eyes
Watch it again — as your blueprint
You heard this Thursday. Today, listen differently — every piece you're about to build is happening right in front of you.
💬 Answers a question
From its system prompt — the menu & hours you gave it
→
⚡ Books something
That's tool use — it's not just talking anymore
→
🔁 The loop
Thought → action → observation, same as always
Everything in your spec from ten minutes ago — you're about to build every piece of it yourself.
🛠️ Your build
Let's build it
- Set up the project — .env + .gitignore, this time without being reminded
- Drop in the prompt Claude just wrote — read it together, check it matches what you said
- Wire up the conversation — get it answering hours, menu, "table for 4 at 7pm?"
- Add one real action — check a small availability list, actually book against it
You're driving. I'm coaching. Invent whatever restaurant sounds fun to you.
Make it yours
This isn't a real restaurant — you're inventing it. What's it called? What kind of food? What's the vibe?
🍽️ Example: what a filled-in spec looks like
Name: Mo's Late Night Ramen
Hours: 5pm – 1am, closed Mondays
Menu highlights: Tonkotsu ramen, spicy miso, gyoza
Vibe: Casual, a little chaotic, host has personality
Never: Never confirms a table without checking availability first
Fill in your own version of this. This is the real work of today — anyone can wire up an API, but a good spec is what makes the agent actually useful.
🧠 Consultant move
Absorbing someone else's expertise
Notice what you're doing right now. With your game agent, you used your own knowledge. Today you're inventing — and encoding — someone else's business knowledge.
That's exactly what a consultant does: sit with a business owner, absorb what they know, and turn it into something that works without them in the room.
✅ Test it
Test it against your own spec
Scroll back up to your conversation with Claude. Does it actually do what you said it should — every single thing?
Now the real question: "If you were the restaurant owner and I showed you this right now — what would you ask me to fix before you trusted it with real customers?"
That question — noticing gaps from the customer's point of view, not just the builder's — is the actual client-facing skill.
🏠 Your mission
Homework
Required:
- Get the agent handling everything from today's spec correctly
- Test it on a friend or family member — have them try to order something weird
- Write down one thing it got wrong
Stretch — pick one:
- Add the ElevenLabs voice layer so it talks out loud
- Make it handle a party size it can't accommodate — gracefully, not by just saying yes
- Add a second menu or location to see if your prompt generalizes
Wednesday: the other half of the business
Session 4 — Wednesday, July 29 · 1:30pm
Today's agent handles people who come to you. Wednesday's agent goes out and finds people worth reaching out to — a lead-gen and outreach agent.
Two different halves of running a business: answering the people who show up, and finding the people who haven't yet.
Great work today
Questions?
You just built something a real business would pay for.
Session 4 of 5
Outreach + Building Your Portfolio
No new code today — a different kind of skill.
Wednesday, July 29, 2026 · 1:30–3:30pm
👋 Recap
Two real things, built and working
Your game recommender. Your restaurant agent. Both actually work.
Today is different — no new agent to build. Two halves instead: how I actually find real people to reach out to, and turning what you've already built into something real people can see.
💡 New idea
Not every problem needs building
Sessions 2 and 3 you built things that live on a website. That's one real skill.
Today's a different one: using Claude directly — no code — to actually get work done.
🛠️ Build
Something needs to run on its own, for other people, without you there. That's when you write an agent.
💬 Use
You need an answer or a draft right now, for yourself. That's when you just talk to Claude directly.
Knowing which one a problem calls for is part of the business mindset from Session 1.
👀 Watch this
How I actually find people to reach out to
No custom tool. I just tell Claude who I'm looking for.
- Describe the target audience — plain language
- Ask for a list of real or representative prospects
- Pick one and draft a short outreach message together
Let's do it live — right now.
What makes a message actually work
Not the tool. The message itself.
- Short — nobody reads a paragraph from a stranger
- Specific to them — proof you're not copy-pasting
- One clear ask — not five options, one next step
Your turn — pick the next prospect and draft it yourself. I'll coach.
🎯 Second half
Two things you built. Just... sitting there.
Right now, your game recommender and your restaurant agent exist. But a stranger clicking your site wouldn't necessarily notice them, or get why they matter.
Let's turn them into a real "Things I've Built" section — the kind of thing that goes on a resume or a college app.
Every project gets three lines
- What it is — one sentence, plain English
- What you built it with — real tool names (Claude Code, ElevenLabs)
- The resume sentence — the version you'd actually say out loud
"Built and deployed an AI agent that handles restaurant reservations using Claude and ElevenLabs."
You write it. You say it out loud. Then it's done.
🖥️ The section
Things I've Built
🎮 Game Recommender
Lives right on the page — visitors can try it themselves. Lowest-risk, most interactive.
Live & playable
🍽️ Restaurant Agent
Shown as a short demo video — always works, always looks its best, no live-chat risk.
Video demo
One interactive, one video — that contrast actually shows range.
🎬 Before recording
Real lines, not "just talk naturally"
We write the actual sentences together first. No improvising on camera.
- One-line intro — what it is
- The ask — a real reservation request
- The moment it actually checks & books
- One-line close — what it just proved
Rehearse once or twice. If camera's not your thing, record it silently and narrate over it after — same result, less pressure.
⭐ The payoff
Look at your site now
This morning it was two separate things you built.
Now it's a real portfolio.
Now it's a real portfolio.
One link. A stranger clicks it, scrolls, and immediately gets that you build real things.
🏠 Before Friday
Homework + what's next
Homework:
- Polish anything that felt rushed today
- Show the finished site to one more friend — what do they click first?
Friday — Session 5: we harden both agents, add guardrails, and I teach you how to actually pitch this to a business owner who'd pay for it.
Nice work today
Questions?
Two builds, one real portfolio.
Session 5 of 5 — Final Session
Harden, Deploy & Scope a Real Client
From "what's an agent?" to knowing exactly who to sell to
Friday, July 31, 2026 · 1:30–3:30pm
👋 Look what you built
Five sessions ago, you didn't know what an agent was
Now you have a live portfolio site with two real projects on it. Let's look at it together before we finish it off.
S1
Chatbot vs. agent
S2
Built the game recommender
S3
Built the restaurant agent
S4
Turned it into a portfolio
S5
Harden, deploy, scope a client
🔧 First up
Let's fix your real bugs
Not made-up examples — the actual things that broke while you were building. That hallucination in Session 2. Whatever went sideways in Session 3's testing.
Real bugs are more convincing than hypothetical ones. Let's go fix them properly.
🛡️ Concept
Guardrails — for real this time
You've heard the idea since Session 2. Today we write the actual rules into the prompt.
- Never confirm something it didn't actually check
- Say "I don't know" instead of guessing when unsure
- Escalate to a human for anything it's not confident about
Then we test that they actually hold — try to break it on purpose.
✅ Final pass
Ready for a stranger to click
- .env never exposed — check both repos
- Both projects work cold, for someone who's never seen them
- A real README on each — what it is, what it's built with, live link
This is the difference between "a project on my laptop" and something real people can actually use.
🧭 The real skill
Let's find your ideal client
Not a performance. Not a test. Just the two of us figuring out something real: who would this restaurant agent actually be perfect for?
Knowing exactly who to approach is just as valuable as knowing what to say — and most people never even stop to think about it.
💭 Question 1
What size business, what revenue?
Too small, and missed calls barely matter. Too big, and they already have a system — or a team.
A single independent restaurant, doing somewhere around $500K–$2M a year, is often the sweet spot: enough volume that a missed reservation actually costs real money, but no in-house tech team already solving it.
What do you think? Let's land on a range together.
💭 Question 2
What kind of owner?
Picture the actual person. Not a company — a human being.
- Busy, hands-on, probably answering the phone themselves
- Not especially technical — doesn't want to "manage software"
- Feels the pain personally — missed calls, double-bookings, chaos during rush
- Would feel real relief, not need convincing that AI is legitimate
This is the person who says "finally" — not the person you have to talk into it.
💭 Question 3
What's the tell?
How would you actually spot a restaurant like this from the outside?
Back to Session 1's rubric: repetitive, high-volume, rule-based, costly. What would you notice about a place that's losing bookings to missed calls?
Let's write down the actual signs together.
📝 Make it real
Write it down as a real profile
4-5 lines. This becomes an actual document — not just a conversation that disappears.
- Business size & revenue range
- What the owner is like
- What tells you they need this
- What they'd care about most when you approach them
This is called an ideal customer profile — real consultants define this before they ever reach out to anyone.
🎭 Simulation
Now I'm the business owner
I match the profile you just wrote. But here's the thing — your job isn't to sell me anything.
Your job is to figure out what I actually need — and what working together would look like.
Three things to walk away with: the problem, the workflow, and the contract structure.
Step 1
Ask me questions first
Don't guess what I need. Find out. Same rubric from Session 1:
"What's the most repetitive part of your day?"
"How often does that happen?"
"What does it cost you when it goes wrong?"
"What are you doing about it right now?"
Step 2
Now define the actual workflow
Not "an AI agent." Specifically — what does it do, step by step?
- What triggers it? (a call, a form, a message)
- What does it do? (check, book, confirm)
- What happens when it's unsure? (escalate — your guardrails)
- What does the owner see? (a text? a dashboard? nothing unless it breaks?)
Say it back as a sequence. That sequence is the scope.
Step 3 ⭐ New
What kind of deal makes sense?
Four common ways to structure work like this:
💵 One-time build
Flat fee, you build it, hand it over. Good when it's simple and well-defined.
🔁 Build + monthly
Setup fee, then a smaller recurring fee for monitoring and fixes. Most common for agents.
🧪 Paid pilot first
Small fee for a trial, then a bigger contract. Good when they're hesitant.
📅 Retainer
Ongoing monthly for continued work. Good when they keep needing changes.
🤔 Your call
Which one fits — and why?
There's no single right answer here. The reasoning is the whole point.
Think about: How much will this need fixing after it's built? Does the owner seem hesitant? Is this one project, or the start of more?
Let me show you a real one — the actual contract from this course, and how it's structured.
🎉 Look back
What you actually did in 5 sessions
- Learned what makes something an agent, not just a chatbot
- Built two real, working AI agents from scratch
- Learned API keys, MCPs, system prompts, and guardrails
- Turned your work into a real, live portfolio
- Defined a real ideal client profile — who to approach, and why
- Scoped a real workflow and picked a contract structure for it
🚀 What's next
This doesn't have to stop here
Keep building. Keep showing people. Keep the Claude subscription going if you want to keep going.
If a real business matching the profile you defined today ever comes up — a family friend, someone you know — you're ready for it. No pressure, just an option.
Proud of you
You did it.
From "what's an agent?" to knowing exactly who you'd build for next. Five sessions.