Module 1 · How a language model works, no math required
A language model, like the ones behind ChatGPT, Claude, or Gemini, is a program that learned to continue text. During training it read an enormous amount of writing and learned which word tends to come next in each context. When you type to it, it doesn't "look up" the answer in a database: it builds, word by word, the continuation that seems most likely.
A useful comparison
Think of the autocomplete on your phone's keyboard, only trained far, far more. If you type "Can I get a large coffee and a…", your keyboard might suggest "bagel." A language model does the same thing with whole paragraphs: it picks up the tone, the topic, and the intent, and replies with something that sounds like what a knowledgeable person would say.
What it does really well
- Drafting and rewriting: a welcome message, a polite reply to an upset customer, a product description.
- Summarizing: turning a long conversation into three bullet points.
- Classifying: telling whether a message is a complaint, a pricing question, or a sign that someone wants to buy.
- Holding a conversation: keeping track of what has already been said in the chat.
What it is NOT
- It's not a database. It doesn't know today's price at your local stationery store or whether they still have a size M, unless you give it that information.
- It doesn't know what happened after its training, unless it's connected to tools that look up current information.
- It doesn't "understand" the way a person does. It produces convincing text, and convincing isn't always true.
Why this matters in sales
An AI agent is a language model that's given instructions, business information (catalog, hours, policies), and tools (check inventory, create an order). How good the agent is depends less on the model and more on what information you give it and what limits you set. That's exactly the skill you'll practice in this course.
Try this today. Open an AI assistant and ask it the price of a product sold by a small shop in your neighborhood. Look at the answer: did it make up a price, say it didn't know, or ask you for more details? Then give it the real price and ask again. Write down the difference: that's what changes when the model has the business's information.
What does a language model do when you type to it?
- It looks up the exact answer in an up-to-date database of the internet
- It generates, word by word, the text that seems most likely
- It copies the answer from someone else who asked the same question
What does an AI agent need to answer correctly about the price of a product in a store?
- A bigger model that already knows market prices
- A customer who asks with good spelling
- The store's catalog with up-to-date prices
Go deeper (these resources are in Spanish): the difference between an AI agent and a chatbot, the glossary of terms, all our guides, how we teach and, once you finish, the Fundamentals and Conversational Seller certifications. See also how much an AI agent costs.
Tutor's question. If a language model only predicts the most likely word, why does it sometimes get things right that it never saw written down? Explain your position with an example, and say what would change your mind.
Module 2 · Prompts that work: role, context, examples, and format
The difference between a generic answer and a useful one almost always comes down to the instruction. The instruction you give a model is usually called a prompt. There are no magic formulas, but there are four pieces that nearly always improve the result.
1. Role: who it should be
"You're the person who answers the text messages for a neighborhood bakery in Chicago. You're friendly and polite, and you never oversell." The role sets the tone and the judgment.
2. Context: what it needs to know
Hours, products, prices, delivery areas, what to do when something is out of stock. Without context, the model fills the gaps with guesses. A simple rule: anything you expect it to know, write it down.
3. Examples: show it what "good" looks like
One or two examples are worth more than ten adjectives. Instead of "answer in a friendly, short way," show it:
Customer: do you have cinnamon rolls? Reply: We sure do! They come out of the oven at 7 a.m. and 4 p.m., $3.50 each. How many should I set aside for you?
4. Format: how you want the output
"Reply in three lines or fewer," "always end with a question," "if it's an order, return it as a list with item, quantity, and total." Format makes the answer usable without having to fix it.
Before and after
Weak prompt: "Answer the bakery's customers."
Prompt that works: "You answer the text messages for Golden Crust Bakery (Pilsen, Chicago). Hours: 6 a.m. to 8 p.m., every day. Products and prices: [list]. Delivery only in Pilsen and Little Village, $4. If someone asks about something that isn't on the list, say you'll check with the manager; never make up prices. Reply in three lines or fewer and end with a question that moves the order forward."
Iterate, don't guess
Write the prompt, test it with tough questions ("can I pay you next week?", "do you have anything gluten-free?", "that's way too expensive"), see where it breaks, and adjust. That loop of testing and fixing is the real work.
Try this today. Write a prompt with all four pieces for a business you know. Test it with five questions: two normal ones, two tough ones, and one the business can't answer. Fix the prompt until all five answers are acceptable, and keep both versions: they're useful for your portfolio.
You want the agent to reply briefly and warmly. What works best?
- Including one or two examples of replies like the ones you expect
- Writing “BE VERY FRIENDLY” in all caps at the start of the prompt
- Asking it to think really hard before answering
The agent gave the wrong business hours. What's the most likely cause?
- The customer asked unclearly and the model understood a different day
- The prompt didn't include the hours and the model guessed
- AI models can't handle times and dates accurately
Tutor's question. You wrote a very long prompt and the agent does worse than with a short one. What's your hypothesis, and what test would you run to confirm or rule it out?
Module 3 · Hallucinations, bias, and ethics: verify before you claim
Hallucinations
A hallucination is an answer that sounds confident but is false or made up: a price that doesn't exist, a return policy the business never had, a citation nobody ever wrote. It happens because the model produces the most likely text, not the true text. In sales, a hallucination is expensive: a customer walks into the store with a price nobody is going to honor.
How to reduce them:
- Give it the source of truth (catalog, policies) and ask it to answer only from that.
- Allow it to say "I don't know" or "let me check with the manager."
- Verify yourself any specific fact before you use it: numbers, dates, rules, names.
Bias
Models learned from text written by people, and they inherit people's biases. They might, for example, assume that whoever asks about power tools is a man, or treat someone differently depending on how they write. In customer service, that turns into unequal treatment. The healthy habit is to review replies to customers with different profiles and fix the prompt when an unfair pattern shows up.
Basic ethics for using AI in sales
- Don't deceive. If a customer asks whether they're talking to a person, don't deny it.
- Don't pressure people with lies: no "only a few left" if it isn't true.
- Protect data. Don't paste customers' personal information (Social Security numbers, home addresses, health information) into a public assistant. In the United States, the FTC can act against unfair or deceptive data practices, and a growing number of states have their own privacy laws.
- Own what you hand in. If you used AI for a piece of work, review all of it and take responsibility for the result; if your school or employer has rules about using AI, follow them.
The golden rule
Verify before you claim. AI helps you write fast; making sure it's true is still your job.
Try this today. Ask an AI assistant for three verifiable facts about a topic you're studying (a date, a number, and a quote). Look up each one in a reliable source. Write down how many were correct and what you did to check them. That habit is what will set you apart most at work.
The assistant gave you a number with total confidence, but you can't find where it comes from. What do you do?
- Use it, because models almost never get numbers wrong
- Round it so it sounds more believable
- Don't use it until you've confirmed it in a reliable source
A customer asks, “Am I talking to a real person?” The agent is a bot. What should it say?
- The truth: that it's an assistant and can connect them with a person
- Yes, so the customer doesn't go to a competitor
- Politely change the subject and offer today's special
Tutor's question. A classmate says that using AI on a college assignment is cheating. Where would you draw the line between using it well and using it badly, and what arguments would you use to defend that line to your professor?
Module 4 · The customer and the job they want done
Clayton Christensen, a professor at Harvard Business School, popularized an idea that changed how people think about customers: people don't buy products, they "hire" products to get a job done in their lives. This idea is known as Jobs to be Done.
The classic example, close to home
Someone who buys a drill doesn't want a drill: they want a hole in the wall. And really, they don't want the hole either: they want to hang the shelf so the apartment looks put together before their friends come over on Saturday. If you understand that job, you can offer something better: maybe a wall-anchor kit plus installation.
Three layers of every job
- Functional: the practical part. "I need lunch to arrive before 1 p.m."
- Emotional: how they want to feel. "I don't want to look bad in front of the team I invited."
- Social: how they want to be seen. "I want them to see I picked a great place."
How to uncover the job in a chat
Ask about the situation, not just the product:
- "What's the occasion?"
- "When do you need it by?"
- "What have you used so far, and what didn't you like?"
Example: someone messages a flower shop, "How much is a bouquet?" A catalog-style reply sends the price list. A reply that understands the job asks: "Is it for a birthday, an anniversary, or to say you're sorry? Do you need it delivered today?" With that, you can recommend the right bouquet, the card, and the delivery time. You sell more because you help better.
Why it matters for an AI agent
An agent that only knows prices answers questions. An agent you've taught the most common customer jobs (with their key questions) guides the purchase. That's why, before designing any agent, you need to know which jobs come in through the chat.
Try this today. Think about the last important purchase you made. Write down the functional, emotional, and social job you wanted done. Then ask two people about a purchase of theirs and do the same. You'll notice the product is almost never the center of the story.
According to Jobs to be Done, what is a customer really buying?
- The cheapest product they can find for what they need
- A way to make progress in a specific situation
- The best-known brand on the market, because it gives them peace of mind
A customer asks, “How much is a bouquet?” Which reply best uncovers the job they want done?
- Sending them the full price list as a PDF so they can choose
- Telling them the price of the cheapest bouquet
- Asking what the occasion is and when they need it
Tutor's question. Think of something you bought this month. What job did you “hire” it to do, and what would have to happen for you to “fire” it and buy something else?
Module 5 · Project: design the sales agent for a local business
Now you'll apply everything in a real project. We'll use the five stages of design thinking, a people-centered design approach popularized by Stanford University's d.school and the design firm IDEO.
1. Empathize
Pick a business near you: the pharmacy, the bakery, the clothing boutique, the auto repair shop. Talk to whoever handles customers (ten minutes is enough) and ask: what do people ask you most by text or WhatsApp? At what times? Which sales slip away? What worries you about using a bot? If you can, watch them serve customers for a while.
2. Define
Sum up the problem in one concrete sentence: "The bakery loses delivery orders after 6 p.m. because nobody answers messages while the register is being closed out." A well-defined problem is worth more than ten ideas.
3. Ideate
List several ways to solve it before you fall in love with one: an automatic message with the hours, an agent that takes orders, a broadcast list with the day's fresh bread, an online form. Choose the one that solves the problem with the least effort for the owner.
4. Prototype
Write the agent's prompt (from Module 2): role, context with real products and prices, examples, format, and limits (what it must never promise, when to hand off to a person). Add the most common customer jobs (Module 4) with the question the agent should ask in each case.
5. Test
Simulate ten conversations in an AI assistant using your prompt: normal orders, questions about things not on the menu, an upset customer, someone asking to pay later. Show the results to the business owner and ask what they'd change. Adjust and test again.
What you hand in
A one- or two-page document with: the business, the defined problem, the agent's final prompt, three test conversations (one that went wrong and how you fixed it), and what the owner said. That document is the centerpiece of your portfolio.
Try this today. Schedule the ten-minute conversation with the business you chose and bring your five questions written down. Ask for permission before using the business's name in your portfolio; if they say no, replace it with a generic one.
According to design thinking, which stage does the project start with?
- Empathize: understand the people and their situation
- Prototype: write the agent's prompt right away
- Test: simulate conversations
Your agent handles the normal questions well. What comes next in the testing stage?
- Call it done, since it already handles the most common questions
- Test it with tough cases and show the results to the owner
- Launch it for all of the business's customers and see what happens live
Tutor's question. The business owner asks you to skip the interview and jump straight to writing the agent. How would you convince them not to, or in what case would you agree with them?
Module 6 · Your portfolio: how to show what you know
When you're looking for your first job, saying "I know how to use AI" doesn't set anyone apart. What sets you apart is showing a real problem you solved. Your Module 5 project is exactly that.
The one-page case study
Turn your project into a short case study with this structure:
- Context: which business and what situation.
- Problem: the sentence you defined.
- What I did: the key decisions (and why you ruled out other ideas).
- Result: what you saw in testing and what the owner said. If it wasn't put into use, say so honestly: a tested prototype is valuable too.
- What I learned: one or two things you'd do differently.
Write it in plain language. Anyone reading it should get it in two minutes.
LinkedIn without exaggerating
- Clear headline: "Business Administration student · AI for Sales."
- Licenses & certifications section: add your Zavora Academy Certificate in AI for Sales with its verification code; anyone can check it on the public verification page.
- A post about your case study: the problem, what you did, and what you learned. Thank the business if they gave you permission to name them.
- Don't inflate: "I designed a prototype and tested it with the owner" beats "I implemented AI at a company" if that's not what happened. They'll ask about it in an interview.
How to talk about the certificate in an interview
The certificate opens the conversation; the case study wins it. Prepare a one-minute answer: what you learned, the project you did, and a mistake you fixed along the way. Showing that you verify before you claim (Module 3) says a lot about how you'll work.
Keep learning
This is the beginner level. If you're interested in selling, the natural next step is the Fundamentals certification in AI-powered WhatsApp sales, followed by Conversational Seller. Both are currently taught in Spanish.
Try this today. Write a draft of your one-page case study and ask someone who doesn't know the topic to read it. If they can't tell you within two minutes what problem you solved, simplify. Once you pass the exam, add the certificate to your LinkedIn profile.
What best shows a recruiter your skill with AI?
- A long list of every AI tool you've tried
- Saying in the interview that you're an expert in AI and sales automation
- A short case study of a real problem you solved, with what you learned
Your agent ended up as a tested prototype, but the business didn't put it into use. How do you present it?
- Honestly: a prototype tested with the owner, plus what you learned
- As if it were already up and running, since it could be
- Better not to mention it, because a project that wasn't put into use costs you points
Tutor's question. A recruiter tells you a free certificate doesn't prove anything. What would you show them to prove otherwise, and which part of their criticism would you accept?