AI in Marketing, Sales, and Customer Experience

Introduction and Context

Marketing, sales, and customer experience (CX) have traditionally depended on segmentation, market research, sales judgment, campaign experimentation, and service processes. Artificial intelligence (AI) extends these activities by finding patterns in large datasets, predicting likely outcomes, generating content, recommending actions, and supporting conversations at scale. Its value, however, does not arise from automation alone. Entrepreneurs must connect AI applications to a clearly defined customer problem, a measurable business objective, reliable data, and responsible governance.

AI can help an enterprise answer questions such as:

• Which customer groups are most valuable or most likely to respond?
• What message, offer, product, or channel is appropriate for a particular customer?
• Which leads deserve immediate sales attention?
• What demand should the firm expect next month?
• Which customers are at risk of leaving?
• How can service be made faster without making it less trustworthy?

The entrepreneurial challenge is to use AI as a decision-support and value-creation capability rather than as a collection of disconnected software features.

Learning Objectives

By the end of this chapter, students should be able to:

1. Explain how AI supports segmentation, personalization, forecasting, content generation, recommendation, lead scoring, conversational commerce, and customer service.
2. Distinguish descriptive, predictive, and prescriptive marketing analytics.
3. Select suitable metrics for acquisition, retention, customer lifetime value, and marketing return.
4. Interpret a simple AI marketing model and evaluate its assumptions.
5. Design a responsible AI-enabled customer journey.
6. Identify risks involving privacy, fairness, transparency, security, and human oversight.

1. A Business Framework for AI-Enabled Marketing

A useful marketing AI system has five connected layers:

1. Data: first-party behavioral, transactional, demographic, service, and consent data; legally obtained external data may also be used.
2. Models: statistical or machine-learning models that classify, predict, rank, generate, or recommend.
3. Decisions: campaign selection, offer selection, sales prioritization, inventory planning, or service routing.
4. Interactions: advertisements, email, websites, mobile applications, sales conversations, chat, and human service.
5. Measurement and governance: outcome measurement, experimentation, monitoring, privacy controls, and accountability.

The layers form a feedback loop. Customer interactions generate new data; new data can improve a model; model outputs affect future interactions. This loop can create learning advantages, but it can also amplify errors. For example, if a recommendation system repeatedly promotes already popular products, it may reduce exposure for new or minority products and make its own historical pattern appear increasingly correct.

AI marketing applications can also be classified by analytical purpose:

Prediction is not causation. A model may identify customers who are likely to purchase without showing that a particular advertisement caused the purchase. Causal questions require controlled experiments, credible quasi-experimental designs, or other methods that address confounding.

2. Segmentation and Personalization

Segmentation

Segmentation divides a market into groups with meaningful differences in needs, behavior, value, or response. Traditional segmentation may use age, location, firm size, or industry. AI allows entrepreneurs to combine many variables, including browsing sequences, purchase frequency, average order value, service contacts, product usage, and text responses.

Common approaches include:

• Rule-based segmentation: groups are defined by explicit conditions, such as “customers with more than three purchases in six months.” It is transparent and easy to audit.
• Clustering: an unsupervised algorithm groups observations according to similarity. The entrepreneur must still decide whether the clusters are commercially interpretable and ethically acceptable.
• Propensity segmentation: customers are grouped according to predicted likelihood of an action, such as purchase, renewal, or response.
• Value-based segmentation: groups are differentiated by current or expected economic value.

A segment is useful only when it supports a different decision. A statistically distinct group that receives the same product, message, and service as every other group may add complexity without creating value.

Personalization

Personalization adapts content, timing, product presentation, price, service, or channel to an individual or segment. It may be explicit, such as addressing a customer by name, or substantive, such as recommending a compatible accessory based on a prior purchase.

Effective personalization requires a balance between relevance and intrusion. Customers may appreciate recommendations based on a current shopping session but be uncomfortable when sensitive inferences are used without explanation. Entrepreneurs should collect only data that is necessary, disclose meaningful uses, and provide practical controls.

A personalization system should be evaluated against a control group. Higher click-through rates do not necessarily mean higher profit, because personalization may increase low-margin purchases, discounts, returns, or service costs.

3. Forecasting and Demand Planning

AI forecasting uses historical observations and explanatory variables to estimate future demand, traffic, leads, or revenue. Potential predictors include seasonality, price, promotions, holidays, weather, search activity, and supply constraints.

A simple baseline is the moving average. For a window of length k, the forecast for the next period is:

Here, Y_t is observed demand in period t, F_t+1 is the forecast for the next period, and k is the number of past periods included. The formula assumes that recent observations provide useful information and that no strong trend or structural break makes the past unrepresentative. It produces demand units when each Y is measured in units.

For example, if weekly orders were 100, 120, and 110, a three-week moving-average forecast is:

More advanced models may capture nonlinear patterns and interactions, but complexity does not guarantee accuracy. Forecasts should be compared with a simple baseline and evaluated on data not used for training. A commonly used mean absolute error is:

where n is the number of evaluation periods, Y_i is actual demand, and Ŷ_i is predicted demand. MAE is expressed in the same units as demand and penalizes errors linearly. It does not indicate whether errors are systematically too high or too low, so bias should also be examined.

Forecasts support marketing by coordinating campaign timing, inventory, staffing, and cash requirements. A promotion that succeeds in generating demand but encounters stockouts may damage trust and create avoidable service costs.

4. Generative AI for Marketing Content

Generative AI can produce or transform text, images, audio, video, code, summaries, and conversational responses. Marketing uses include drafting product descriptions, adapting a campaign to multiple languages, generating subject-line alternatives, summarizing customer research, and creating first drafts for sales proposals.

A responsible content workflow includes:

1. Define the audience, objective, offer, brand constraints, and evidence requirements.
2. Provide approved source material and prohibit unsupported claims.
3. Generate several alternatives rather than accepting the first output.
4. Check factual accuracy, accessibility, cultural appropriateness, copyright, and brand tone.
5. Obtain human approval for high-impact or externally published material.
6. Monitor performance and record material model or prompt changes.

Generative models can produce fluent but false statements, often called hallucinations. They can also reproduce bias from training data, reveal confidential information through poor handling, or create content that appears personalized but is emotionally manipulative. Human review is especially important for health, finance, children’s products, employment-related services, and regulated claims.

The appropriate question is not “Can AI create this?” but “What level of automation is appropriate given the cost of error?” Low-risk headline variation may be automated with sampling and approval rules. A legally consequential product claim requires stronger evidence and human accountability.

5. Recommendation Systems

Recommendation systems estimate which products, content, or actions may be relevant to a user. Three broad approaches are:

• Content-based recommendation: recommends items similar to those a user previously engaged with, using item characteristics.
• Collaborative filtering: uses patterns of interaction among users and items. A user may receive an item because similar users liked it.
• Hybrid recommendation: combines behavioral, contextual, and item-based evidence.

A recommendation score is not automatically a measure of customer welfare. Optimizing clicks can favor sensational or low-quality content. Entrepreneurs should define the objective carefully: contribution margin, repeat purchase, completion, satisfaction, diversity, or a balanced combination.

Cold-start problems occur when a new customer has little behavioral history or a new product has few interactions. Contextual information, explicit preferences, expert rules, and carefully designed exploration can help. Recommendations should also include safeguards against discrimination, unsafe products, and excessive or inappropriate targeting.

6. AI in Sales and Lead Scoring

Lead scoring ranks prospects according to their estimated likelihood of becoming qualified opportunities or customers. Inputs may include firmographic information, engagement, product fit, source, recency, and sales activity. A score is useful only when its definition is connected to a sales action.

A predictive model may estimate:

where p_i is the estimated probability that prospect i achieves a defined outcome, Y_i = 1 indicates that the outcome occurs, and X_i represents observed features. The outcome must be specified, such as purchase within 30 days, not vaguely defined as “good lead.”

A simple expected-value rule can prioritize leads:

where EV_i is expected net value, M_i is contribution margin if the prospect converts, and C_i is the cost of sales attention or service. This simplified rule assumes that the probability estimate is reasonably calibrated and that the values are comparable. It does not include all strategic factors, such as relationship value or capacity constraints.

Worked example. Suppose Lead A has a 0.30 estimated probability of conversion, a contribution margin of $800, and a sales-contact cost of $40. Its expected value is:

Lead B has a 0.18 probability, a $1,400 margin, and a $40 contact cost:

Under this simplified rule, Lead B should be prioritized despite its lower conversion probability. This illustrates why ranking solely by probability can misallocate scarce sales time.

Models should be assessed for discrimination, calibration, stability, and fairness. Calibration asks whether prospects assigned a probability near 0.70 convert at approximately that rate over a suitable population and period. A model may rank prospects well but still produce probabilities that are not trustworthy.

7. Conversational Commerce and Customer Service

Conversational commerce uses chat, voice, messaging, or embedded assistants to support discovery, recommendation, ordering, payment, and post-purchase service. Customer-service AI may classify intent, retrieve knowledge-base content, translate messages, summarize cases, suggest replies, and route complex matters to specialists.

A robust service design separates:

• Intent recognition: what the customer is trying to do.
• Information retrieval: which approved source answers the question.
• Transaction execution: what the system is authorized to change or purchase.
• Escalation: when a human must intervene.
• Auditability: what was asked, answered, changed, and approved.

A conversational system should identify itself as an AI where appropriate, avoid pretending to have human feelings or authority, and provide a visible route to human assistance. Escalation triggers may include authentication failure, emotional distress, legal threats, high-value transactions, safety issues, repeated misunderstanding, or requests outside the system’s knowledge.

Service quality should not be reduced to containment, meaning the percentage of conversations that do not reach a human. A system can achieve high containment by frustrating customers. More balanced measures include resolution rate, time to resolution, customer effort, repeat contact, complaint rate, transfer quality, and customer satisfaction.

8. Measuring Acquisition, Retention, Lifetime Value, and Return

Acquisition

Customer acquisition cost is commonly defined as:

The numerator should specify the period and included costs, such as media, creative, agency fees, sales labor, and relevant technology. Comparing CAC across channels requires consistent attribution rules.

Conversion rate is:

The denominator must be stated. A website conversion rate based on sessions is not directly comparable with a sales conversion rate based on qualified leads.

Retention

For a cohort, a basic retention rate at time t is:

If 700 of 1,000 original customers remain active after six months, retention is 70%. Retention definitions must specify activity, cohort date, observation window, and treatment of pauses or reactivations.

Customer lifetime value

A simple contribution-based lifetime value estimate for a customer with constant periodic contribution margin and a constant retention probability is:

Here, m is contribution margin per period, r is the probability of remaining active from one period to the next, and d is the discount rate per period. The formula assumes a geometric retention process, constant margin, constant retention probability, no acquisition cost within CLV, and an indefinite horizon. If those assumptions are unrealistic, a cohort or customer-level model should forecast margins and survival by period instead.

For m = $50, r = 0.80, and d = 0.10 per month:

This is contribution value, not revenue. If acquisition cost is $60, the estimated contribution after acquisition is $73.33 under the stated assumptions.

Marketing return

Return on marketing investment can be expressed as:

Incremental means the additional contribution caused by the marketing activity relative to a credible counterfactual, usually a control group. Attributed revenue without incrementality can overstate performance because customers may have purchased anyway.

Suppose an experiment estimates $12,000 incremental contribution from a $5,000 campaign. Then:

The campaign returned $1.40 in incremental contribution for each dollar invested after recovering the investment. This does not mean that every dollar of reported revenue was caused by the campaign.

9. Integrated Applied Case: BrightBite Subscription Foods

This is a fictional illustrative case. BrightBite is a small subscription company selling customized snack boxes. It has 18,000 customers, rising digital advertising costs, frequent first-month cancellations, and a service team overwhelmed by delivery questions.

Background and problem

The founders used broad social-media targeting and sent the same email to all subscribers. They knew total sales but could not reliably compare customer value across acquisition channels. A chatbot pilot had also produced incorrect delivery promises.

Analysis

The team first defined outcomes: subscription activation, 90-day retention, contribution margin, and service resolution. It consolidated consented first-party data from orders, product preferences, delivery events, campaign exposure, and service contacts. It excluded sensitive attributes from targeting and established access controls.

Analysis showed three actionable groups: value-seeking monthly subscribers, convenience-oriented customers with low service tolerance, and high-engagement customers interested in dietary variety. These were not treated as permanent identities; segment membership could change as behavior changed.

A churn model ranked customers by predicted cancellation risk, while a separate value model estimated expected contribution. The team did not automatically discount every high-risk customer. Instead, it tested helpful interventions such as delivery-choice reminders, preference updates, and a pause option.

Approach and solution

BrightBite introduced:

• a demand forecast to plan inventory for seasonal flavors;
• recommendation rules based on allergies and declared preferences, with a safety review;
• AI-generated email drafts checked by a marketer;
• lead scoring for wholesale prospects;
• a retrieval-based service assistant limited to approved delivery policies;
• human escalation for refunds, allergy concerns, and repeated failed answers;
• randomized holdout groups to estimate incremental effects.

Outcome

After one quarter, management reported lower service handling time and improved campaign relevance. However, the most important learning came from the experiment: a generic retention discount increased short-term renewal but reduced contribution margin, while preference-update messages produced a smaller but more profitable improvement. The company therefore retained the latter intervention and limited discounts to cases where expected incremental value justified them.

Learning points

1. AI applications should be connected to explicit outcomes, not novelty.
2. Different models may be needed for propensity, value, and intervention choice.
3. A model that predicts churn does not prove that a discount will prevent it.
4. Customer benefit and contribution margin should be measured together.
5. Human escalation and policy boundaries are essential in service automation.

10. Governance, Ethics, and Regulation

AI marketing creates obligations beyond technical performance.

Privacy and consent. Organizations should establish a lawful and transparent basis for collecting and using personal data, limit collection to necessary purposes, protect data, and support access, correction, deletion, or objection rights where applicable under relevant law. Consent should not be treated as a universal solution when the processing is unexpected or excessive.

Fairness. Segmentation, credit-like lead prioritization, pricing, and service routing may disadvantage protected or vulnerable groups. Teams should test outcomes across relevant groups, investigate proxy variables, document decisions, and provide appeal or human review where decisions materially affect people.

Transparency. Customers should understand when they are interacting with AI, why important recommendations or decisions are made at a meaningful level, and how to obtain help. Transparency should be useful rather than a dense technical disclosure.

Security. Systems need authentication, authorization, encryption where appropriate, prompt-injection defenses, logging, retention controls, and vendor-risk assessment. Generative systems must not be given unrestricted access to customer records or transaction functions.

Intellectual property and truthfulness. Generated claims, images, and copy must be checked for rights, provenance, accuracy, and misleading presentation. Synthetic testimonials or fabricated customer evidence are unacceptable.

Governance. A practical governance record identifies the purpose, data sources, model owner, decision owner, performance measures, known limitations, escalation rules, review date, and incident process. Governance is not a one-time approval; models and customer behavior change over time.

11. Implementation Roadmap and Best Practices

Entrepreneurs can implement AI through a staged roadmap:

1. Clarify the decision. State who needs to decide what, when, and with what consequence.
2. Define the outcome and metric. Separate leading indicators, such as clicks, from business outcomes, such as contribution and retention.
3. Audit data. Check quality, consent, representativeness, missingness, duplication, and access rights.
4. Start with a baseline. Compare AI with a rule, human process, or simple forecast.
5. Pilot safely. Use a limited population, monitoring, approval gates, and a control group where feasible.
6. Evaluate beyond accuracy. Examine calibration, cost, latency, fairness, customer experience, robustness, and operational impact.
7. Design human responsibility. Assign owners for model outputs, exceptions, customer complaints, and shutdown decisions.
8. Scale selectively. Automate only where the expected value exceeds implementation and failure costs.
9. Monitor drift. Reassess data distributions, conversion behavior, segment meaning, and performance after market changes.

A strong entrepreneurial portfolio usually combines low-risk productivity uses, such as summarization, with carefully tested decision systems. The aim is not maximum automation. It is better customer value, sustainable economics, and accountable learning.

Contemporary Developments

Several developments are shaping the field. Multimodal models combine text, images, audio, and video, enabling richer product search and service analysis. Retrieval-augmented systems connect generative models to approved, current information rather than relying only on model memory. Agentic systems can plan and execute multiple actions, but their autonomy increases the need for permissions, transaction limits, verification, and logs. Privacy-enhancing methods, synthetic data, and on-device processing may reduce exposure of personal information. At the same time, regulatory expectations increasingly emphasize risk classification, documentation, transparency, security, and human oversight. Entrepreneurs should monitor applicable national and sector-specific requirements rather than assuming that a general AI tool is automatically lawful or suitable.

Chapter Summary

AI can support the full customer journey, from segmentation and acquisition to sales prioritization, recommendation, purchase, retention, and service. Its applications include predictive models, generative content, forecasting, conversational systems, and prescriptive recommendations. Effective use requires a clear decision, reliable and lawfully obtained data, meaningful outcome measures, and comparison with a baseline or control group.

Marketing metrics must be interpreted carefully. CAC depends on consistent cost and customer definitions; retention depends on cohort and activity definitions; CLV depends on assumptions about margin, retention, discounting, and time horizon; and ROMI should use incremental contribution rather than merely attributed revenue. AI can improve efficiency and relevance, but it can also amplify bias, privacy intrusion, error, and manipulation. Human oversight, transparency, security, fairness testing, escalation, and continuous monitoring are therefore central entrepreneurial capabilities.

Key Terms

Artificial intelligence: Systems that perform tasks associated with human intelligence, such as prediction, classification, generation, and decision support.

Customer experience: The customer’s overall perception of interactions with an organization across the relationship.

Segmentation: Division of a market or customer base into groups with meaningful differences.

Personalization: Adaptation of content, offers, products, or interactions to a customer or segment.

Propensity: Estimated likelihood that a customer or prospect will perform a defined action.

Forecasting: Estimation of future values using historical and explanatory information.

Generative AI: AI that creates new text, images, audio, video, or other outputs.

Recommendation system: A system that ranks or suggests items, content, or actions.

Lead scoring: Ranking prospects according to predicted fit, likelihood, value, or priority.

Conversational commerce: Use of conversational interfaces to support discovery, transaction, and service.

Retrieval-augmented generation: Generation grounded in information retrieved from an approved knowledge source.

Churn: The loss or cancellation of customers or subscriptions.

Retention: The proportion of a cohort that remains active over a specified period.

Customer acquisition cost: Attributable acquisition cost per newly acquired customer.

Customer lifetime value: Expected contribution or value from a customer over the relationship.

ROMI: Return on marketing investment, preferably based on incremental contribution.

Incrementality: The additional outcome caused by an intervention relative to a counterfactual.

Calibration: Agreement between predicted probabilities and observed outcome frequencies.

Model drift: Decline or change in model behavior as data, customers, or markets change.

Human-in-the-loop: A design in which people review, approve, correct, or handle exceptions to AI outputs.

Review Questions

5 Short Answer Questions

1. What is the difference between predictive and prescriptive marketing analytics?
2. Why is a commercially useful segment more than a statistically distinct cluster?
3. State two risks of using generative AI for marketing content.
4. What does calibration mean in lead scoring?
5. Why should customer-service AI provide escalation to a human?

5 Long Answer Questions

1. Explain how an entrepreneur could design an AI-enabled customer journey from acquisition through retention.
2. Compare content-based, collaborative-filtering, and hybrid recommendation systems.
3. Discuss the assumptions and limitations of the simple CLV formula presented in this chapter.
4. Explain why attributed revenue may overstate marketing effectiveness and describe how a control group helps.
5. Evaluate the governance requirements for an AI customer-service assistant.

3 Analytical / Application Questions

1. A campaign costs $8,000 and produces an estimated $20,000 in incremental contribution. Calculate ROMI and interpret the result. What additional evidence would you request before scaling the campaign?
2. Two leads have the following estimates: Lead X has conversion probability 0.25, contribution margin $900, and contact cost $50; Lead Y has probability 0.40, margin $600, and contact cost $50. Use expected value to rank them and explain why probability alone could produce a different ranking.
3. Design an experiment to test whether personalized recommendations improve profitable retention rather than merely increasing clicks. Specify the treatment, control, primary outcome, and at least two safeguards.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes prediction from causation in marketing AI?

A. A prediction always identifies the reason an outcome occurred.

B. A model can identify likely purchasers without proving that an advertisement caused their purchases.

C. Causation can be established by using a larger customer database alone.

D. Prediction and causation are identical when model accuracy is high.

Correct Answer: B. A model can identify likely purchasers without proving that an advertisement caused their purchases.

Question 2

A three-period moving-average forecast uses observed demands of 100, 120, and 110 orders. What is the forecast for the next period?

A. 100 orders

B. 105 orders

C. 110 orders

D. 120 orders

Correct Answer: C. 110 orders

Question 3

Which approach recommends items based primarily on characteristics of items a user previously engaged with?

A. Content-based recommendation

B. Randomized control allocation

C. Collaborative filtering only

D. Moving-average forecasting

Correct Answer: A. Content-based recommendation

Question 4

A lead has a 0.30 conversion probability, $800 contribution margin, and $40 contact cost. Using EV = p × M − C, what is its expected value?

A. $200

B. $240

C. $280

D. $800

Correct Answer: A. $200

Question 5

Which measure is most directly concerned with whether predicted probabilities correspond to observed frequencies?

A. Latency

B. Calibration

C. Containment

D. Click-through rate

Correct Answer: B. Calibration

Question 6

Under the chapter’s simple CLV assumptions, what is CLV when monthly contribution margin is $50, retention probability is 0.80, and monthly discount rate is 0.10?

A. $40.00

B. $80.00

C. $133.33

D. $400.00

Correct Answer: C. $133.33

Question 7

Why is a control group valuable when evaluating personalized marketing?

A. It guarantees that every customer receives the same message.

B. It estimates what would likely have happened without the intervention.

C. It eliminates the need to define a business outcome.

D. It ensures that reported revenue is always incremental.

Correct Answer: B. It estimates what would likely have happened without the intervention.

Question 8

Which is the most appropriate escalation trigger for a customer-service AI system?

A. A routine question answered from an approved knowledge base

B. A customer asking for store hours

C. A repeated failure to understand the customer’s request

D. A completed low-value transaction

Correct Answer: C. A repeated failure to understand the customer’s request

Question 9

What does ROMI in this chapter preferably use in its numerator?

A. Total website visits

B. Incremental contribution margin minus marketing investment

C. Gross impressions minus clicks

D. All reported revenue regardless of cause

Correct Answer: B. Incremental contribution margin minus marketing investment

Question 10

Which practice best supports responsible generative-AI marketing?

A. Publishing the first generated draft without review

B. Giving the model unrestricted access to customer records

C. Checking factual accuracy, rights, accessibility, and brand constraints before publication

D. Optimizing only for maximum engagement

Correct Answer: C. Checking factual accuracy, rights, accessibility, and brand constraints before publication