AI & Marketing

Think of one example of AI being used in marketing — something you have seen, read about, or experienced as a customer.

We will build three lists on the board together:

We will come back to these lists throughout the lecture.

No wrong answers. If you are not sure, say what you think and we will figure it out together.

Many marketers and salespeople doubt AI’s ability to boost company revenues or customer satisfaction. Some even believe it adds to their workload, signaling a disconnect between AI adoption and employee confidence.

• Only 39% of marketers and sales professionals in the US and UK are confident that their departments’ use of AI drives revenues
• Nearly half (46%) believe AI only somewhat improves the customer experience or does not improve it at all

Sources

• General Assembly (Aug 2025), AI in Marketing & Sales Report — survey of more than 300 marketing and sales professionals in the US and the UK.

And maybe you do too.

Broad doubts extend beyond organizational performance. Marketers question AI’s impact on their own work. This suggests a gap between executive enthusiasm and employee experience, where leadership may oversell potential while underestimating operational challenges.

• 15% say AI has had little or no impact on day-to-day work
• 18% say it has created more work and pulled them away from strategic priorities
• 21% say their companies have had to hire more people or have become less productive since adopting AI
• 24% of Gen Z employees work against their company’s AI strategy (e.g., don’t disclose, use their own tools) because it adds to workload

Sources

• EMARKETER (2025), Workers wary of AI rely on it daily but often don’t disclose usage

And, unfortunately, lots of exaggeration.

One possible reason for the underwhelming impact is lack of proper training. Many marketers are handed powerful tools with little instruction on how to use them. Without guidance, employees may use AI inefficiently or avoid it due to uncertainty about which tools fit their tasks.

Only 17% report receiving comprehensive, role-specific training that prepared them to use AI effectively.

Sources

• General Assembly (Aug 2025), AI in Marketing & Sales Report — survey of more than 300 marketing and sales professionals in the US and the UK.

Untrained employees be like →

The skepticism is real. But so are the results — in specific applications where AI is well-matched to the task.

The pattern: measurable gains show up where AI handles high-volume, well-defined, data-rich tasks. Gains are harder to find where the task is ambiguous or the data is thin.

• Programmatic advertising — largely AI-driven — grew 18% in 2024, reaching $134.8B

• Marketing and sales is the function where companies most commonly report revenue increases from AI use

• Among AI high performers, revenue lifts of 3–15% and sales ROI improvements of 10–20% are reported

• Content cost reductions of 5–20% and time-to-market cuts of up to two weeks documented in pharma and retail deployments

Sources

• IAB / PwC, Internet Advertising Revenue Report: Full Year 2024
• McKinsey, The State of AI 2025

Generative AI is not just changing how marketing work is done — it is changing which jobs exist.

Using 40 million U.S. job postings before and after the launch of ChatGPT, Luo, Miao, and Sudhir (2025) document a systematic production-to-orchestration shift:

• Work that draws on bounded, retrievable knowledge (writing copy, handling inquiries, running reports) is being automated.
• Work that requires judgment, coordination, and relationships (strategy, sales, account management) is expanding.

The bottleneck is moving — from producing artifacts to deciding what to do with them.

Both things are true at the same time.

AI produces measurable gains in narrow, well-specified tasks. But organizations consistently struggle to identify those tasks, deploy correctly, and build the internal skills to evaluate results.

This is the gap this course is designed to close.

Case 1 · Week 2 HubSpot & Motion AI: Chatbot-Enabled CRM (HBS #518-067; B case #524-088)

To prepare:

• Read the case
• ✏️ In-class case prep quiz (paper & pencil, closed notes)

Case 2 · Week 3 Artea: Designing Targeting Strategies (HBS #521-021)

To prepare:

• Read the case
• Work through the Option 3 Dashboard
• ✏️ In-class case prep quiz
• 📊 Group Assignment 1 due before class

Case 3 · Week 5 PittaRosso: AI-Driven Pricing & Promotion (HBS #522-046)

To prepare:

• Read the case
• Work through the spreadsheet supplement
• ✏️ In-class case prep quiz
• 📊 Group Assignment 2 due before class

Case prep quizzes

• One quiz per case — three total
• Paper and pencil, in-class, closed notes
• No makeups given
• Lowest grade across all participation activities dropped

Peer surveys

Completion of peer assessment surveys counts toward participation.

Class contribution rubric

Negative behaviors (lateness, distraction, disruption) reduce your score.

What it is

The AI Visibility Essentials certification from SEMrush covers how AI is changing search and digital visibility — a core skill for modern marketers.

A SEMrush Director of Organic Search visits in Week 3 to introduce the material.

How to earn full credit

1. Complete the training materials using your Georgetown.edu account
2. Take the exam in-class on April 8 — paper and pencil, no outside assistance, no AI tools
3. Highest score is kept

This is individual work.

• In-person, paper and pencil, closed book
• Short answers, multiple choice, and calculations
• Covers all lectures, cases, and exercises
• Scheduled during the allocated final exam period

Assesses conceptual understanding and applied reasoning across the full course.

What you will design

1. Research question — how does chatbot design choice X affect customer outcome Y?
2. Scenario — a realistic customer situation
3. Manipulation — two versions of the chatbot, one behavioral difference

The post-interaction survey is standardized across all teams. You do not design it.

Same measures. Different manipulations. Comparable results.

Timeline

Artificial intelligence is machines that think.

Artificial intelligence is machines that think.

Artificial intelligence is machines that think, perceive, and act.

We add:

• Perception: acquiring information from the environment.
• Action: doing something that changes the environment.

A smart cat feeder

Recognizes cats. Dispenses food.

A Home Depot return chatbot

Receives a return request. Verifies the order. Checks the product. Issues a return label.

Artificial intelligence is machines that think, perceive, and act.

Artificial intelligence is models of thinking, perception, and action.

We add:

• Model: A simplified, explicit structure that supports explanation, prediction, or control.

Artificial intelligence is models of thinking, perception, and action.

Artificial intelligence is representations that support models of thinking, perception, and action.

We add:

• Representation: The form in which information is encoded so that computation is possible.

Artificial intelligence is models of thinking, perception, and action.

Artificial intelligence is representations that support models of thinking, perception, and action.

We add:

• Representation: The form in which information is encoded so that computation is possible.
• Having good representations that fit your modeling goals is very important.

The goal changes — so must the representation

The first representation answered: is there a cat?

But the product goal is: is this my cat?

cat:        yes
confidence: 0.94

cat_id:     Maurice
confidence: 0.91

Artificial intelligence is representations that support models of thinking, perception, and action.

Artificial intelligence is constraints exposed by representations that support models of thinking, perception, and action.

We add:

• Constraints: rules that limit which states, transitions, or solutions are allowable.
• Once constraints are visible, search replaces guesswork.
• Constraints do not limit intelligence. They make intelligent behavior computable.

Artificial intelligence is constraints exposed by representations that support models of thinking, perception, and action.

Artificial intelligence is algorithms enabled by constraints exposed by representations that model thinking, perception, and action.

We add:

• Algorithm: A procedure that operates over a representation while respecting constraints.

Artificial intelligence is algorithms enabled by constraints exposed by representations that model thinking, perception, and action.

Artificial intelligence is behavior emerging from algorithms, constraints, representations, and models of thinking, perception, and action.

We add:

• Behavior: What the system does over time, including feedback loops between perception, thinking, and action.

How it works

• A domain expert encodes knowledge as IF-THEN rules
• The system applies rules mechanically
• New knowledge requires human intervention

Cat feeder — pressure plate + timer

IF time = 7:00am OR time = 6:00pm
AND plate_weight > threshold
THEN release food
ELSE stay locked

No camera. No model. No inference.

A cat steps on the plate at feeding time → the spring releases → food drops.

The rules are encoded in the physics and the clock.

Any cat, at the right time, gets fed.

Customer service — phone tree

Press 1: Start a return
Press 2: Check order status
Press 3: Speak to an agent

Enter your order number.
Press 1 to confirm.

For damaged items, press 1.
For wrong items, press 2.
For change of mind, press 3.

The rule is encoded in the menu script.

If your problem isn’t on the menu, you’re stuck.

How it works

• Feed the system many examples
• The system detects patterns across examples
• Patterns generalize to new inputs
• Performance improves with more data

Cat feeder

Input: hundreds of photos
       labeled by the owner
       (Maurice, not Maurice)

Output: a model that maps
        new face images
        to the most likely identity

The pressure plate fed every cat.

The learned model feeds only Maurice.

Home Depot chatbot

Input: thousands of customer messages
       labeled with correct intents
       (start_return, check_status,
        damaged_item, wrong_item,
        missing_receipt, escalate)

Output: a model that maps 
        new messages 
        to the most likely intent

The phone tree broke on anything not in the menu.

The learned model handles variation.

The problem

The owner wants the feeder to recognize Maurice — and only Maurice.

We want a system that looks at a photo and decides: is this Maurice?

The function we want

\[f(\text{image}) \rightarrow \{\text{Maurice}, \text{not Maurice}\}\]

We cannot write this function by hand.

We have no rule that reliably identifies Maurice from a photo.

So we learn it from data.

What the training data looks like

Each photo gets a label. The algorithm learns what separates them.

Where do these labels come from?

The problem

A customer uploads a photo of the item they want to return.

We want a system that looks at the photo and decides: is this the right product?

The function we want

\[f(\text{image}) \rightarrow \{\text{match}, \text{no match}\}\]

We cannot write this function by hand.

We have no rule that reliably identifies a product from a photo.

So we learn it from data.

What the training data looks like

Each photo gets a label. The algorithm learns what separates them.

Where do these labels come from?

The annotation task

An annotator is shown a photo and a reference product (e.g., DeWalt 20V MAX Drill, SKU #100672834).

They answer one question:

Does this photo show the same product?

They click Match or No Match.

Then the next photo appears.

For a catalog of 10,000 SKUs with 50 photos each — that is 500,000 individual judgments.

What the annotator sees

Training phase

• Collect labeled examples (images + labels)
• Algorithm finds patterns that separate classes
• Result: a trained model \(\hat{f}\)

Product match example

The model sees thousands of labeled product photos.

It learns which pixel patterns predict “match”.

Four properties

• Representative: covers the range of inputs the model will see
• Labeled accurately: outputs must be correct
• Large enough: more examples generally means better generalization
• Recent enough: old data may not reflect current patterns

The algorithms were not new. The infrastructure was.

The ideas behind neural networks existed in the 1980s. What changed was the ability to feed them data and compute at scale.

The problem

You want the feeder to recognize Maurice — and only Maurice.

You cannot write rules for this. Faces vary by angle, lighting, weight gain, time of day.

So you learn it from examples.

The function

\[f(\text{face image}) \rightarrow \{\text{Maurice}, \text{not Maurice}\}\]

What the training data looks like

Someone — you — sat down and labeled these.

The algorithm learned what separates them.

The model was trained on photos of Maurice, Garfield, and a neighbor’s tabby.

It has never seen this:

What the model sees

A face embedding — a vector of numbers.

Not a cat. Not a mask. Not a costume.

A point in a high-dimensional space.

What happens

cat_id:     Garfield
confidence: 0.97
eligible:   no

The mask changes the pixels.

It does not change the embedding enough to fool the model.

The model generalizes. That is the point of supervised learning.

Structure

• Training data: (input, label) pairs
• Goal: learn \(f\) such that \(f(\text{input}) \approx \text{label}\)
• Prediction: apply \(f\) to new inputs

Two main tasks

• Classification: output is a category
• Regression: output is a number

The setup

The feeder has been running for six months.

Maurice is recognized correctly. Garfield is blocked.

But the access logs show something odd: 47 bowl-approach events that don’t match any registered cat.

You didn’t go looking for this. The data revealed it.

What the logs show

2:14am  Unknown face — blocked
2:31am  Unknown face — blocked
2:14am  Unknown face — blocked
3:02am  Unknown face — blocked
...

A clustering algorithm groups the unknown events by face similarity.

Three clusters emerge:

• Cluster A: Small face, pointy ears, 1–3am → raccoon
• Cluster B: Tabby face, similar to Garfield, daytime → neighbor’s cat
• Cluster C: Looks like Maurice but slightly different → Maurice’s sibling?

You didn’t define these categories. The algorithm found them.

The setup

You built a supervised intent classifier for the return chatbot.

You defined six intents:

• start_return
• check_status
• damaged_item
• wrong_item
• missing_receipt
• escalate

You labeled thousands of messages. You trained the model. You deployed it.

Now you have 20,000 new conversations.

And some messages don’t fit.

The process

Feed all 20,000 conversations — unlabeled — into a clustering algorithm.

The algorithm groups messages by linguistic similarity.

It does not know what the groups mean.

A human inspects each cluster and names it.

Some clusters were expected:

• damaged_item
• wrong_item
• missing_receipt

What this means for the system

The unsupervised analysis revealed five intent categories that were never in the original label set.

Each one represents a real pattern in customer behavior that the supervised classifier was silently mishandling.

The labels came from the data — not from the design team.

Common tasks

• Clustering: group similar examples ← this is what we just did
• Dimensionality reduction: compress representations
• Anomaly detection: find unusual examples

The broader principle

Unsupervised learning is not a fallback when you lack labels.

It is the right tool when you don’t yet know what the labels should be.

The setup

The feeder recognizes Maurice. But when should it dispense food?

You could set a fixed schedule: 7am and 6pm. But Maurice sometimes skips a meal. Sometimes he’s hungry at 3am. Sometimes he’s been fed by a neighbor.

A fixed rule wastes food. A learning system does better.

The agent’s decision at each bowl approach:

Dispense food now — or wait?

Structure

• Agent: the feeder
• Environment: Maurice and his eating patterns
• Action: dispense food, or stay locked
• Reward: Maurice finishes the bowl = positive; food sits uneaten = waste = negative; Maurice goes hungry = negative
• Learning: over time, the feeder learns which conditions predict a finished bowl

Structure — escalation policy

• Agent: the chatbot
• Environment: the customer and their conversation
• Actions: continue automated conversation, or escalate to a live agent
• Reward: issue resolved without escalation = positive; customer abandons chat = negative; unnecessary escalation = wasted agent time = negative
• Learning: over time, the chatbot learns when escalation leads to better outcomes

Pre-training

Self-supervised: predict the next token from billions of text examples. Has the flavor of supervised learning — but labels are generated from the data itself, not provided by humans.

Fine-tuning

Supervised: train on human-curated examples of good responses.

Alignment

Reinforcement learning from human feedback (RLHF): human raters score responses; the model is updated to produce higher-rated outputs.

The framework

A system is doing AI when it:

1. Perceives the world — raw inputs from the environment
2. Represents what it perceives — in a form computation can act on
3. Models patterns in that representation — under constraints
4. Acts on the world — and observes the consequences

Representations, models, constraints, behavior. That is the architecture of every AI system we will study.

What this means for labeling things “AI”

Running a regression on a spreadsheet is not AI.

Clustering customers into segments is not AI — it is a model, but there is no perception loop, no environment, no action.

These are useful tools. They are not AI. Precision matters.

1. Where around you would an AI system actually help?

Not: “where could I add a model?”

But: what perceives, what represents, what acts, and on what?

Think about your workplace, your industry, your daily life. Where is there a perception-action loop waiting to be closed?

What you will do

Design a scenario in which a customer interacts with a Home Depot chatbot.

Change one thing about how the chatbot behaves.

Measure how that change affects customer perceptions.

Teams of 4 or 5.

Teams will be formed before next class (after add/drop).

If you have preferences, indicate them on the Google Sheet on Canvas before next class.

Scenario (pre-purchase)

You are shopping on the Home Depot website and need to replace several bulbs in your living room. You are unsure which type to buy — LED, smart bulb, or standard — and what wattage is appropriate.

You open the shopping assistant chatbot to help you choose.

Your goal: find the right bulb for your space and budget.

Manipulation — one difference

Scenario (post-purchase)

You recently purchased a DeWalt DCD791D2 cordless drill from Home Depot. After a few uses, the chuck has become loose and the drill no longer holds bits securely.

You open the Home Depot customer support chatbot to resolve the issue.

Your goal: determine whether you can return or exchange the drill.

Manipulation — one difference

Form your team.

Teams of 4 or 5.

Indicate preferences on the Google Sheet on Canvas before next class. Preferences are not guaranteed but will be considered.

Add/drop closes before next class — teams are finalized after that.

Read the HubSpot case.

The case is in your course pack.

Read it. Prepare it. Come ready to discuss.

There will be a short case prep quiz at the start of class.

⚠️ Do not be late.