BUAN 3301 — AI in Business

Week 11: AI in Customer Service

Dr. Ashim Bose — Section 004
March 31 + April 2 (both sessions) | Comprehensive Visual Edition
Quiz Prep Notes

1. Profit Center vs. Cost Center

Dr. Bose spent significant class time on this. In any business, departments are classified as
either generating revenue or purely consuming costs.

Profit Center Cost Center

Definition: A department or function where Definition: A department or function that only
revenue coming in offsets the costs. The incurs costs with no direct revenue. It’s necessary
department contributes directly to the company’s for the business but doesn’t generate money.
top line.

Examples: Sales (you sell products/services and Examples: Customer Service, HR,
collect payment), Product (you build and sell a Finance/Accounting, Engineering/IT. All
product that generates revenue) necessary, but none directly generate revenue.

Treatment: Well-funded, glamorous positions. Treatment: HR, Finance, and Engineering still get
Employees get bonuses, awards, and strong respect because they require highly specialized
incentives for performance. skills. Customer Service gets deprioritized
because the skill sets are seen as more general
and entry-level.

Why CS gets neglected: While every company pays lip service to customer service, it’s often
not truly prioritized. Sales, marketing, product, and engineering are the “glamorous” positions.
Customer service is frequently staffed with entry-level workers in low-cost locations (India,
Philippines) at $3–10/hour. This underfunding creates a cycle: poor service → unhappy
customers → more complaints → more burnout.

Feedback bias: Customers are psychologically far more likely to provide feedback when they’re
upset than when satisfied. Very few people will proactively praise a company for good service.
This means the data flowing into CS teams is overwhelmingly negative, which contributes to the
burnout problem and also skews any analysis of customer sentiment.

CS burnout: Approximately 90% of customer service interactions involve unhappy customers.

Combined with low investment, lack of recognition, and repetitive work, this leads to high
turnover rates in CS teams.
2. The Customer Service Pyramid (Levels 0–3)
Dr. Bose walked through how a 100-person customer service team is structured. The support
organization follows a pyramid model where the base is the largest and cheapest layer, and
each successive level is smaller, more expensive, and more specialized.

Level Headcoun Specific Responsibilities Key Details from Class

t
Level 0AI / AI (no Deflect issues from the service CHEAPEST level. This is where AI
Self-Service humans) desk using self-healing and targets FIRST because the savings
self-service. potential is highest.
Handles incidents marked with a Uses chatbots, automated FAQ systems,
star (★) on slide 7: ticket creation, self-service portals.
general troubleshooting, FAQ Goal: handle as much volume here as
support, and knowledge base possible so humans don’t have to.
solutions. The ★ star items on the slide indicate
These are issues that can be fully tasks that COULD be automated at L0.
resolved without any human
intervention.

Level ~50 Ticket creation and ticket handling Entry-level workers, often fresh out of
1Entry people(50 General troubleshooting college or high school.
Agents % of team) FAQ support Largest layer = most people, lowest cost
Installation support per person.
Configuration control Typically staffed in low-cost locations
Solutions from knowledge base (Manila, Hyderabad) at $3–10/hr.
Escalation to Level 2 when unable First human contact when L0 can’t
to resolve resolve the issue.
Dr. Bose described visiting a Manila call
center at 2 AM — the city was buzzing
 because agents work US nighttime
hours.

Level ~25 Escalation from Level 1 More experienced, higher cost per
2Senior people(25 Updating ticket status person.
Agents % of team) Detailed issue tracking Handles problems that L1 couldn’t solve
Check integration issues — typically more complex technical
Hardware and software issues.
interoperability These agents understand how systems
Simulation interact with each other (interoperability).
Escalation to Level 3 You want to protect this layer by
resolving as much as possible at L0/L1.

Level ~15 Escalation from Level 2 The “rock stars” — Dr. Bose’s words.
3Expert people(15 Performance tuning Walk on water, can solve any problem.
Specialists % of team) Remote troubleshooting MOST EXPENSIVE per person. You
Test lab simulation never want them taking routine calls.
Environmental simulation Close to engineering — can diagnose
Bug management product-level bugs.
Development escalations They also maintain the knowledge base
that L0/L1/L2 rely on.
Change management
Responsible for training lower levels, so
Training (for lower levels)
their expertise propagates down.
Updating knowledge base
This is where development escalations
go when a real code fix is needed.

Managemen ~10 Oversee the entire support Top of the pyramid.

t people(10 operation Responsible for scheduling, staffing
% of team) Set strategy, manage budgets decisions, and AI implementation
Report to executive leadership strategy.
The ones who receive the “reduce by
20” mandate from above.

The star (★) system from the slide: Items marked with a star at Level 1 are incidents that
COULD be resolved at Level 0 if AI/self-service is implemented. This is the key insight — ticket
creation, general troubleshooting, FAQ support, and knowledge base lookups are all
automatable. That’s why AI targets L0 first: it absorbs the starred tasks that currently require L1
humans.

The economic logic: If you have 50 people at L1 earning $3–10/hr, and AI can handle 60–70%
of their workload, the savings are massive. As you move up the pyramid, AI becomes less
useful because L2/L3 problems require human judgment, complex diagnosis, and product-level
knowledge. You typically stop automating before L2/L3 because humans in the loop are
essential for nuanced problems.

3. Routing & Escalation Flow

When a customer contacts support, the issue enters at Level 0 (AI/self-service). The AI attempts
to resolve it automatically. If it can’t, it routes to Level 1 (entry agents). If L1 can’t resolve, they
escalate to L2, then L3. At each step, the unresolved issue moves to a more expensive, more
specialized person.

Why this matters: Each escalation costs more money and takes more time. A call that gets
resolved at L0 costs nearly nothing. The same call at L3 might cost 10x more because you’re
paying a highly skilled expert to handle it. The entire system is designed to filter: resolve as
much as possible at the cheapest level, and only pass up what truly requires human expertise.

Real-world context: When you chat with a company and first get a bot, then get transferred to
a person, then get transferred to a “specialist” — that’s this escalation path in action. The “for
training purposes” call recordings feed back into the AI to improve L0 resolution over time.
4. Customer Service Pain Points
Pain Point Details How AI Solves It
Long wait times Even now, calling a 1-800 number virtually AI handles inquiries instantly.
guarantees a wait. Dr. Bose: “The question is No hold time. 24/7 availability.
how long? 3 minutes? 5 minutes?” Newer Reduces queue length for
approach: companies ask for your number and remaining human agents.
call you back.

Repetitive queries Agents answer the same questions hundreds AI chatbots handle repetitive
of times per day — “Where is my package?” queries (60-70% of total
“How do I reset my password?” No human volume). Agents freed up for
wants to do this. It’s the #1 cause of burnout. complex, meaningful work.

Inconsistent You email and get one answer, then chat and AI gives identical,
responses get a different one. Even two human agents policy-compliant answers every
interpret rules differently. Dr. Bose: “As a time, across every channel.
customer, you don’t want a dice roll. You want Eliminates human
the best outcome.” interpretation variance.

24/7 coverage Students take classes during the day — the AI doesn’t sleep, doesn’t take
only time to resolve issues is at night, when breaks, doesn’t have timezone
agents are off. Call centers in Manila work US constraints.
nighttime hours, but coverage gaps remain.

Scaling costs Every 10% increase in call volume requires AI scales to 10x volume with
more hires, more training, more office space. minimal cost increase. The 10x
Cost grows linearly with volume. scalability factor from the ROI
framework.

The 7-to-1 Rule

It takes approximately 7–8 positive experiences to overcome a single negative one. Dr. Bose
emphasized this applies everywhere — customer service, workplace relationships, classroom
interactions. “If you lose control once, it takes 7–8 times to recover.” This is why the early IVR
implementations (which were terrible) are still haunting companies — customers remember the
bad experience, and it takes many good AI interactions to rebuild trust.

Workplace implication: “This may not be a big deal for you as students, but when you’re in the
workplace dealing with customers and colleagues, you definitely don’t want to get into situations
where you need that many recoveries.”
5. Traditional vs. AI-Enabled Journey

Step Traditional (Manual) AI-Enabled

Routing Human operators manually classify each AI auto-classifies inquiries across all
inquiry and route it to available agents or channels (chat, email, phone, social media)
departments. Slow, error-prone, and routes instantly to the right person or
inconsistent. bot.

Triage Agents manually assess urgency and AI triage with sentiment analysis and
priority based on customer interaction and priority scoring. Frustrated customers get
predefined protocols. Subjective judgment. flagged for immediate human attention.

Resolutio ALL issues, including common FAQ Simple issues (60-70% of volume)
n questions, require a human agent. Even auto-resolved by AI. Complex issues routed
password resets go through a person. to humans with AI providing real-time
suggestions.

Knowledg Agents rely on static knowledge bases, AI provides real-time suggestions, relevant
e personal experience, and often lack knowledge articles, customer history, and
real-time context. They manually search for sentiment analysis during the interaction.
answers.

Follow-up Agents manually send follow-up emails, Automated follow-up communications.

surveys, and reminders after each System learns from each interaction to
interaction. Inconsistent and improve future responses.
time-consuming.

Analytics Periodic, manual review of feedback and Continuous AI-driven analytics. Real-time
performance data. Slow optimization cycle dashboards, trend detection, and automatic
— changes take weeks or months. optimization.
Key point from April 2: AI can be applied across the ENTIRE customer service process —
there is no step where AI doesn’t have an opportunity to add value. This is why the AI customer
service market is $15B and growing at double-digit rates (24% CAGR through 2030).

The AI-Enabled Process Flow (Slide 11)

Inquiries arrive through multiple channels (chat, email, phone, social media). AI performs initial
triage, instantly resolving straightforward questions while routing complex issues to the right
human expert. Even when humans handle cases, AI provides real-time suggestions, relevant
knowledge articles, and sentiment analysis. After resolution, the system learns from each
interaction, continuously improving.
6. The Six Customer Service KPIs
These were the class exercise topics. Students presented on each one. Know all six in depth:
what it measures, the formula, and specifically how AI can improve it.

6a. Average Handle Time (AHT)

Definition: The total time spent on a single customer interaction, including talk time, hold time,
and after-call work (filling out the ticketing system, documenting what happened, logging the
resolution).

Formula: AHT = (Talk time + Hold time + After-call work) / Total number of calls

Why it matters: Lower AHT means agents can handle more calls per shift, reducing the
number of agents needed and lowering cost per contact.

How AI improves it: (1) AI listens to the call in real-time and queries the knowledge base
instantly — no more putting customers on hold while agents manually search. (2) AI surfaces
relevant information before the agent even asks. (3) AI automates after-call work by
summarizing the conversation and auto-filling the ticket, so agents can move to the next call
immediately.

From student presentation (Raj Kedia): Customer service reps currently have to manually
search a knowledge bank, which takes time and requires putting the customer on hold. AI with
access to the same knowledge bank can retrieve information instantly and present it to the
agent during the call.

6b. First Contact Resolution (FCR)

Definition: The percentage of customer issues that are fully resolved on the first interaction,
without the customer needing to call back or follow up. Higher is better.

Formula: FCR = Issues resolved on first contact / Total contacts

Why it matters: Every follow-up contact costs money and erodes satisfaction. If a customer has
to call three times to fix one problem, that’s 3x the cost and a frustrated customer.

How AI improves it: (1) AI scans the customer’s account and recent activity BEFORE the call
connects — it already knows what you’re likely calling about. (2) During the call, AI surfaces the
exact policy or resolution relevant to the issue. (3) Intelligent routing sends the issue to the
agent most likely to resolve it. (4) Self-service deflection handles simple issues entirely without
human involvement.

Proactive customer service: Dr. Bose highlighted this as exceptional: “Wouldn’t it be nice if
they already knew what you were calling about? And had a solution? What if they proactively
called YOU? That’s proactive customer service, where they predict a problem and solve it
before you even know there’s a problem.”
From student presentation: Connect Plus (telecom) used AI to pre-scan accounts before calls
connect, checking for open tickets, overcharges, and missed payments. FCR went from an
industry average to 81% within six months. Average handling time dropped to 5.4 minutes.

6c. Cost per Contact (CpC)

Definition: How much money a company spends to handle one customer support interaction —
one call, one email, or one chat.

Formula: CpC = Total support costs / Number of customer interactions

Why it matters: This is the fundamental efficiency metric. If you can lower CpC while
maintaining quality, you’re doing more with less.

How AI improves it: (1) AI chatbots handle simple questions at near-zero marginal cost. (2)
Faster routing means less wasted agent time. (3) Agents solve problems more quickly with AI
assistance. (4) The combination means fewer total human hours per interaction.

6d. Customer Satisfaction (CSAT)

Definition: A direct measurement of how satisfied customers are, typically on a 1–5 or 1–10
scale, collected via post-interaction surveys.

Formula: CSAT = Sum of satisfaction scores / Total number of responses

Limitation: Traditional surveys suffer from low response rates and sampling bias. They capture
only the very happy or very unhappy — not the middle.

The silent majority problem: Most customers don’t respond to surveys. They’re out there
using the product, but they’re not talking to you. So how do you measure their satisfaction?

How AI addresses the silent majority: (1) Predictive CSAT scoring: AI analyzes chat/call
transcripts to score 100% of interactions, not just the surveyed ones. (2) Indirect signal analysis:
AI looks at retention patterns, repeat usage, purchase frequency, and engagement metrics. (3)
Extrapolation: from the 10 calls you DO receive feedback on, AI can model the likely sentiment
of the 90 you don’t.

From student presentation (Kylie): Pre-match predictive scoring turns the silent majority into
permanent data. AI scores every interaction, achieving 100% coverage instead of the small
percentage who fill out surveys.

6e. Net Promoter Score (NPS)

Definition: Measures the likelihood that a customer will recommend the company to friends and
family. Scale from -100 to +100.

Formula: NPS = % Promoters (score 9-10) minus % Detractors (score 0-6)

Important nuance: NPS is NOT a perfect predictor of company growth. It works best for
tracking loyalty and sentiment before and after a specific change.
How AI improves it: (1) Reduces customer effort (less frustration = more promoters). (2)
Predicts detractors before they leave — AI identifies at-risk customers from behavior patterns.
(3) Auto-analyzes all NPS comments to find the top 5 reasons for promoters and top 5 for
detractors. (4) Coaches agents in real-time to improve interactions. (5) Automates “closed-loop”
process: reaches back out to detractors to understand why and try to recover them.

From student presentation (Nathan): Starting NPS of +32 (pretty bad). After AI
implementation: Day 1 — AI handles 60% of inquiries. Day 7 — AI starts predicting detractors.
Day 14 — auto-analyzes all NPS comments. Day 30 — agent performance improves 20% via
real-time coaching. Day 45 — closed-loop automation. Result: NPS jumped from +32 to +54.

6f. Agent Productivity (AP)

Definition: A measure of efficiency and effectiveness for individual agents, balancing quality
(are they solving problems well?) with quantity (how many are they handling?).

Formulas: Productivity = Total output (interactions) / Total input (hours worked). AP Score =
(Resolved tickets × Quality score) / (Active handling hours × Work order).

Core components: Quantity/volume (how many calls), Efficiency (how fast), Effectiveness (did
they resolve it), Quality (adherence to standards and client requirements).

How AI improves it: (1) AI co-pilot gives real-time assistance during calls. (2) Automated
post-call work (summaries, ticket filling). (3) Low-value queries deflected to AI, so agents handle
only meaningful work. (4) Automated QA and coaching — AI reviews calls and provides
feedback to agents.

From student presentation (Daniela): Without AI, a typical interaction: 3 minutes searching
customer account + 7 minutes troubleshooting + 5 minutes wrapping up = 15 minutes. With AI
assistance, each of those steps is significantly reduced.
7. Sentiment Analysis for Agent Evaluation

Use 1 — Real-time customer monitoring: AI monitors the customer’s emotional state

throughout the call. If the customer started neutral but becomes angry or frustrated, the AI flags
this for immediate escalation to a human expert. This prevents bad situations from getting
worse.

Use 2 — Agent evaluation and cross-checking: After each call, agents enter data into the
system about what happened and how it was resolved. But agents might misinterpret — or
misrepresent — what happened. Companies now collect separate customer feedback (“Rate
this person 1–5”) and compare it against what the agent reported. If there are big gaps between
what the agent says happened and what the customer experienced, that’s a red flag.

Dr. Bose (April 2): “The customer service person might say ‘Yeah, I helped resolve this
customer’s issue.’ They can put whatever they want in their report. But the company wants to
know from YOU as the customer what your sentiment was, to cross-check what they put into the
system. If there are a lot of variations between what’s in the system and what the customer
feels, then where’s the problem? It’s the agents.”

8. Knowledge Management for AI

What forms of knowledge exist in a company:

•​ Customer data: Account information, purchase history, interaction logs, support tickets.
Structured data that lives in databases.
•​ Product/service documentation: Descriptions, specifications, pricing, feature lists. The
official record of what the company sells.
•​ User manuals: Installation guides, operating instructions. Dr. Bose: “These are quite
precious because they contain detailed how-to information.”
•​ FAQs: Frequently asked questions with standard answers. A critical knowledge source
for both humans and AI.
•​ Call transcripts: Recordings and transcripts of actual customer interactions. Extremely
valuable because they show the AI what to do AND what not to do.

Why folders don’t work: Dr. Bose asked the class: “If everybody has documents in their own
folder, does that work for a company with thousands of employees?” No. Nobody has access to
everyone else’s files.

Three types of data (from earlier lectures): Structured (tables, databases), Unstructured
(documents, text, transcripts), Semi-structured (JSON, XML, logs). All three need to be
organized and accessible for AI.

Storage systems: Structured data → Traditional databases (Oracle, Microsoft SQL Server,
PostgreSQL, Google BigQuery). Unstructured/both → Modern platforms like Databricks and
Snowflake that handle structured AND unstructured data. Cloud storage for document
management. Dr. Bose: “Data is the new oil. These data companies have very big valuations.”

Prerequisite for AI: All of this knowledge must be organized and accessible to the AI system. If
your data is scattered across folders, siloed in different departments, or locked in legacy
systems, the AI cannot access it. This is why data quality and integration is barrier #1.
9. RAG Architecture in Customer Service

What RAG is: Retrieval Augmented Generation. A process where a foundation model (like
Claude or GPT) is connected to company-specific data so it can give accurate, factual answers
instead of making things up.

The Hyundai example (April 2 lecture): A customer visits the Hyundai website and clicks the
AI chat widget. They ask to schedule a service appointment. The chat connects to a foundation
model running on the back end. But can you just use a raw foundation model? No — it doesn’t
know Hyundai’s dealer availability, service schedules, or customer accounts.

How RAG solves this: The model sends a query to the RAG system. RAG retrieves relevant
data from Hyundai’s databases — dealer schedules, customer account info, service history —
and feeds it back to the model. The model then generates a grounded, accurate response:
“Your nearest dealer has an opening on Thursday at 2 PM for your 50,000-mile service.”

Without RAG = hallucinations: The model makes up answers that sound confident but are
completely wrong. With RAG = factual, company-specific responses grounded in real data.

Connection to the class RAG exercise: Dr. Bose referenced the RAG exercise from earlier in
the semester — same concept, now applied to customer service.

10. How AI Actually Works in CS (Slide 14)

Step 1 — LLMs process input: Large Language Models trained on massive text datasets
understand context and generate human-like responses. Think of them as prediction engines
that learned patterns from billions of conversations. Examples: Claude (Anthropic), GPT
(OpenAI), Gemini (Google).

Step 2 — RAG ensures accuracy: Retrieval Augmented Generation connects the LLM to your
specific knowledge base — product docs, policies, FAQs — pulling verified information to
ground responses in facts rather than hallucinations.

Step 3 — System integration delivers context: Integration with CRM systems and databases
provides customer history, order details, and account status. This enables personalized,
informed responses instead of generic answers.

Step 4 — Processing matches use case: Real-time processing for live chat and voice (needs
to be instant). Batch processing for email analysis and trend reports (can take longer). The
architecture adapts to latency requirements and cost constraints.

11. Core AI Technologies (Slide 13)

Technology What It Does Key Details from Class
Chatbots / Text-based interactions handling Most popular AI use case in CS. This is what
Conversational FAQs, order status, and guided companies implement first. Replaces L0/L1
AI troubleshooting through natural humans for routine queries. Biggest area of
dialogue. investment.

IVR (Interactive Phone-based automation with Early implementations were terrible (“press 1
Voice speech recognition and natural for this, press 2 for that”). These bad
Response) language understanding for memories persist. Takes 7-8 positive
hands-free service. experiences to overcome. Most current IVR is
still pre-LLM old tech.

Sentiment Real-time emotion detection Dual purpose: (1) Flag unhappy customers for
Analysis identifying frustrated customers escalation, (2) Evaluate agent performance by
who need immediate human comparing agent reports vs. actual customer
attention. sentiment.

Knowledge Answer engines and retrieval Requires organized data infrastructure

Management systems finding the right (Oracle, SQL Server, Databricks, Snowflake).
information from vast databases “Data is the new oil.” Without this, AI has
instantly. nothing to retrieve.

Predictive Anticipating customer needs and Also used for workforce scheduling (call
Analytics potential issues before they volume trends, special events, staffing
escalate into service calls. optimization). Dr. Bose walked through the
Manila scheduling example.

12. Predictive Analytics for Workforce Scheduling

If you manage a call center of 100 people, you need to build a weekly schedule. Dr. Bose: “You
can’t just randomly throw names into a spreadsheet.”
What the AI needs to analyze:

•​ Call volume trends: What time of day has the most traffic? Typically ramps up in the
morning, peaks around lunchtime/early afternoon, drops off at night. The AI maps this
curve to determine how many agents are needed at each hour.
•​ Request complexity: Some requests take 5 minutes, others take 30. The AI factors in
the mix of simple vs. complex expected at different times.
•​ Special events: Product recalls, software patches, outages, or promotions cause
sudden call spikes. If there’s a recall, “call volume goes off the charts because
everybody’s trying to get their fix.” The AI needs access to event calendars.
•​ Agent preferences: Especially for overseas teams in Manila working nighttime hours.
Agents want to know their schedule so they can plan family time. The AI generates
schedules that balance coverage needs with agent availability.

The output: an optimized schedule that says “I need X people for the morning shift, Y for the
afternoon peak, Z for the night shift.” This is predictive analytics applied to operations.

13. ROI Framework (Slide 16)

Cost Savings Revenue Impact Experience Metrics

30% labor cost reduction 15% conversion lift 40% faster resolution
50% training time decrease 25% upsell increase 12-point NPS improvement
10x scalability factor 20% retention boost 90% CSAT scores

Break-even: Typically 6–12 months for companies implementing AI thoughtfully. Start with
high-volume, low-complexity use cases where automation rates exceed 70%. As the system
learns, expand to more complex scenarios.

How to calculate ROI: Don’t just measure cost savings. Also measure revenue impact from
improved customer experience and agent capacity to handle higher-value interactions
(upselling, retention).
Cost reality check (April 2): Many call centers pay agents $3–10/hour in low-cost locations.
The cost of implementing AI (development, data prep, API fees, maintenance) must be LOWER
than the cost of the human resources it replaces. That’s the fundamental business case math.

Week 8 business case connection: Customer service operation: 50,000 calls/month at $8/call
= $4.8M/year. AI chatbot handles 40% of calls. Initial investment: $450K. Ongoing: $220K/year.
Year 1 net benefit: $740K. Year 3 net benefit: $1.9M. Total 3-year net profit: $2.99M.
14. Four Critical Barriers to AI Implementation

Barrier What Goes Wrong How to Address It

1. Data Quality Fragmented systems create data silos. Data governance programs. API
Outdated information produces wrong integration between systems. Careful
answers. Privacy regulations (GDPR, handling of sensitive information. Invest
CCPA) add complexity. Different in modern data platforms that centralize
departments have different databases structured and unstructured data.
that don’t talk to each other.

2. AI can confidently state completely RAG architecture (ground responses in

Hallucinations wrong information — a brand and legal real data). Human review loops for
nightmare. See the Air Canada case: the high-stakes interactions. Confidence
chatbot promised a refund that didn’t thresholds (if AI isn’t sure, escalate to
exist, and the company was legally human). Extensive testing before launch.
forced to honor it.

3. Cost Mgmt Per-query API costs add up fast at scale. Model costs carefully and optimize
Infrastructure requires ongoing continuously. Start with high-ROI use
maintenance. Hidden costs include cases (L0 automation). Monitor token
integration, training, and content usage and cost per interaction. Compare
creation. Many underestimate the total against human agent cost per interaction.
cost of ownership.

4. Org Agents fear job loss. Managers doubt AI Incentivize AI adoption (leaderboards,
Resistance capability. Executives expect magic. The rewards). Involve teams early. Set
whole team (100 people) might leave if realistic expectations. Redeploy
they hear 20 are being cut. displaced workers into AI oversight, data
curation, and knowledge management
roles.

15. Air Canada Hallucination Case

Dr. Bose described this real legal case in the April 2 lecture as a cautionary tale:

What happened: A customer contacted Air Canada’s AI chatbot about a bereavement fare. The
chatbot told the customer they could book at full price now and get a retroactive refund within 90
days by submitting a bereavement form. It gave specific details about the refund amount and
process. This was entirely fabricated — a hallucination. Air Canada’s actual policy did not allow
retroactive bereavement fare refunds.
The fallout: Finance caught the discrepancy and told the customer no refund would be issued.
The customer went to small claims court in Canada, presenting screenshots of the chatbot’s
conversation as evidence that the company had made a promise.

The ruling: The court ruled in the customer’s favor. Air Canada argued the chatbot was a
separate entity and not responsible for accurate information. The court rejected this: the
company is responsible for ALL information on its website, whether provided by a static page or
a chatbot.

Takeaway: If you implement AI for a company, YOU are responsible for everything it says. AI
hallucinations aren’t just embarrassing — they create real legal liability. This is exactly why RAG
(grounding in real data), confidence thresholds (if unsure, escalate), and human review loops
(for high-stakes promises) are not optional — they’re legal necessities.

16. The “Reduce Your Team by 20” Scenario

A real-world scenario discussed across both lectures. Your boss comes to you and says: “In six
months, I want you to do more with AI and reduce your 100-person team to 80.” This is
happening at every company right now.

Step 1 — Assess the work: Look at what each level is actually doing. What percentage of L1
calls are simple enough for AI? Review call transcripts, ticket data, and resolution patterns.

Step 2 — Gather training data: Call transcripts (the most valuable — they show what to do
AND what not to do). Product documentation. FAQs. Training materials that agents received
when they joined. All of this becomes training data for the AI.

Step 3 — Build and deploy models: Start at Level 0. Build AI that can handle the starred (★)
items from the service levels slide. Test extensively before launch.

Step 4 — Handle the organizational challenge: This is the hardest part. If all 100 people hear
through the grapevine that AI is replacing 20 jobs, you risk losing ALL 100 — then you have no
one.

Solution: Incentivize the team to learn and use AI. Create an AI leaderboard (“Who are the top
10 people using AI this week?”). Build a culture that embraces AI. People who adopt AI become
more valuable. People who resist become more vulnerable to cuts.

The balanced story: Dr. Bose (April 2): “It’s not just a bad news story about people getting laid
off. It’s a balance story. Some people get laid off, but other people get hired to implement AI,
manage the data, curate knowledge, and oversee the systems.”

CEO dashboards: This is happening right now. CEOs look at AI dashboards showing which
employees are using AI, how many tokens they consume, and how productively they use it.
Employees not using enough AI are being challenged: “Hey, you’ve got to step up your AI game,
or else.”
17. Case Studies
Zendesk
Company: Global CRM platform serving 170,000+ customers.

Challenge: Managing increasing ticket volume while maintaining quality.

AI implementation: Answer Bot for automated responses to common questions. Content Cues
for recognizing patterns in customer feedback. Intelligent ticket routing to send issues to the
right agent.

Results: 15% reduction in resolution times. 20% increase in customer satisfaction. Successfully
scaled across multiple languages and time zones.

Bank of America — Erica

Company: Major financial institution serving millions of customers.

Challenge: Providing personalized financial guidance at scale.

AI implementation: Conversational banking interface (you can talk to Erica like a person).
Proactive financial notifications (alerts about unusual spending, bill reminders). Seamless
mobile integration.

Results: 19.5+ million active users. 105+ million client requests processed. Significant call
center volume reduction. Enhanced financial literacy among customers (Erica teaches people
about their finances).

Key insight: Proactive AI doesn’t just reduce inbound calls — it builds customer loyalty.
Customers appreciate being warned about problems before they happen.

18. AI Use Cases in Customer Service (Slide 21)

Use Case AI Types Used Real Example
Proactive Support Predictive Analytics, NLP Zendesk Answer Bot identifies potential issues
from behavior patterns and proactively reaches
out with solutions before customers report
problems.

Health Monitoring Predictive Analytics, Gainsight Customer Health Score analyzes

Clustering Algorithms product usage, support interactions, and
engagement metrics to predict churn risk.

Feedback Analysis NLP, Sentiment Analysis, Adobe Experience Cloud auto-categorizes

Topic Modeling thousands of feedback points into actionable
themes and sentiment scores to prioritize
product improvements.

Smart Routing Classification Algorithms, Salesforce Service Cloud routes issues to the
Decision Trees most qualified agent based on issue complexity,
 agent expertise, and predicted resolution
success rate.

Virtual CSMs Conversational AI, NLP Intercom Resolution Bot handles repetitive
customer success tasks like feature explanations
and best practice recommendations, freeing
human CSMs for complex work.

Renewal/Upsell Predictive Analytics, HubSpot uses AI to identify optimal timing for

Propensity Modeling renewal conversations and the most relevant
product expansions based on usage patterns.

Voice of Customer NLP, Sentiment Analysis, Qualtrics XM Platform integrates feedback

Entity Recognition across channels (surveys, social media, support
tickets) to create comprehensive voice of
customer insights.

19. Building Customer Loyalty (Slide 8)

Definition: Customer loyalty is the ongoing emotional connection between a customer and a
brand, characterized by a willingness to repeatedly purchase from that company over its
competitors. Loyal customers do more than just buy products — they act as brand advocates,
promoting the company through positive word-of-mouth and social media.

Nine strategies to increase loyalty: (1) Personalization — tailor services to preferences. (2)
Consistency — maintain reliable quality. (3) Rewards — incentivize repeat purchases. (4)
Engagement — maintain regular communication. (5) Feedback — listen and respond. (6)
Support — address issues fast with proactive assistance. (7) Optimization — simplify the buying
process. (8) Exclusivity — early access to unique products. (9) Transparency — be open about
policies and pricing.
20. Aztea Class Exercise (Slide 22)
Scenario: A CRM solution provider called Aztea has undergone a rewrite of their software to
make it cloud native. This has been disruptive to customers: their business processes and
integrations have to be changed, and there are more bugs being reported. Customer
satisfaction and loyalty are going down. Many customers are shopping to replace Aztea with a
competitor.

Team structure: Implementation team (responsible for new software deployments) and
customer support team (responsible for solving issues reported through website or phone).

Exercise: Recommend how Aztea’s customer service should change to improve satisfaction.
Dr. Bose noted this was based on a real-world experience he personally managed.

Recommendation details:

•​ AI Chatbot (L0/L1): Automate triage for known bugs, migration FAQs, and routine
requests. Resolves issues at the cheapest level, reduces wait times, and provides 24/7
coverage — directly addressing the support surge from the rewrite.
•​ RAG Knowledge Base: Centralize product docs, known-issue lists, migration guides,
and FAQs into a structured knowledge system. Connect via RAG to prevent
hallucinations and ensure the AI gives accurate, Aztea-specific answers.
•​ Sentiment Analysis: Monitor all channels in real-time to detect frustration early. Flag
at-risk customers for immediate human escalation. Cross-check agent reports against
actual customer sentiment to ensure quality.
•​ Omnichannel Consistency: Ensure identical answers across website, phone, and chat.
Inconsistent responses erode trust — especially during a disruptive migration.
•​ Predictive Analytics & Proactive Outreach: Forecast call spikes around patch
releases to optimize staffing. Proactively contact customers affected by known bugs
BEFORE they call in — exceptional service that rebuilds loyalty.
•​ Change Management: Offer incentives for support staff to train the AI system. Redeploy
displaced agents into AI oversight, data curation, and knowledge management roles.
Overcome resistance through retraining, not just layoffs.

21. Dr. Bose’s Personal Turnaround Story

This story directly parallels the Aztea scenario and comes from his real 35-year career:
He was brought in to manage a company that had gone through the same type of disruptive
software change. The support team of ~500 people across L0–L3 had become completely
demoralized. When he listened to call recordings, he heard agents cussing at customers,
complaining, and being hostile. The team had absorbed the constant negativity and stopped
caring.

The turnaround process:

•​ Got the entire team together on a weekend and made a statement: “We’re going to fix
this, and we’re going to work extra hard to do it.”
•​ Locked themselves in a room and brainstormed: identified every problem, the root cause
of each, and potential solutions.
•​ Created a 90-day transition process. Assigned dedicated customer service managers to
specific customer sets — these managers would know their customers well and
advocate for them with internal teams.
•​ The engineering team was based in Israel (timezone challenges). Set up structured
communication processes to bridge the gap. Created a prioritization system for the
biggest cross-customer issues.
•​ Slowly but steadily moved customers toward recovery. Communicated commitments
(“we’ll solve X by this date”) and then stuck to those commitments.

The payoff: The company hadn’t held a customer conference in years because leadership
feared what would happen if all the angry customers got together in one room. Eventually they
scheduled one. Dr. Bose went on stage not knowing what to expect — “I ducked under my
podium thinking I was going to get a rotten tomato thrown at me.” Instead, someone in the front
row stood up and started clapping. Then another. Then the entire room gave a standing ovation.
They had seen the difference.

His 35-year career takeaway: “Customer service is a really tough position. You’re basically
always dealing with complaints and negativity. But if you can turn things around, it’s among the
most meaningful and rewarding experiences in a career.” He still gets choked up telling this
story.

22. Pre-LLM vs. Post-LLM Customer Service

The timeline: ChatGPT launched in late November 2022. Everything before that is pre-LLM.
Everything after is post-LLM.

Pre-LLM (before 2022): Expert systems, basic machine learning, rule-based chatbots, rigid IVR
phone trees (“press 1 for billing, press 2 for support”). These were limited, frustrating, and gave
AI in customer service a bad reputation.

Post-LLM (2022+): Large language models that can understand context, handle ambiguity,
generate natural responses, and learn from interactions. A fundamentally different capability.
The critical gap: Even in 2026, most customer service AI that people interact with is STILL old
pre-LLM technology that hasn’t been upgraded. Companies are working on upgrades but many
haven’t launched yet. So the negative perception of “AI customer service” is largely based on
legacy limitations, not the capabilities of modern AI. As more companies upgrade their stacks,
the experience will improve dramatically.
HIGHEST-PROBABILITY QUIZ TOPICS
# Topic to Know
1 Profit center vs. cost center (definitions, examples, why CS gets neglected)

2 The pyramid / Levels 0–3 (headcount, cost per level, starred items, where AI targets first, the
economic logic)

3 Six KPIs: AHT, FCR, CpC, CSAT, NPS, AP (definitions, formulas, how AI improves each, student
presentation details)

4 Silent majority concept (low survey response + sampling bias, how AI predicts from transcripts and
indirect signals)

5 Proactive customer service (predict + solve before customer contacts, the “what if they called you
first” concept)

6 The 7-to-1 rule (7–8 positive experiences to overcome one negative, applies to CS and workplace)

7 “Reduce by 20” scenario (data gathering, model building, org challenge, incentivize adoption,
balanced story)

8 ROI numbers: 30% labor reduction, 50% training decrease, 6–12 month break-even, $4.8M →
$2.99M net profit

9 RAG architecture (foundation model + company data via retrieval = accurate answers, Hyundai
example)

1 Knowledge management: 5 types of knowledge, 3 data types, storage systems (Oracle,

0 Snowflake, Databricks)

1 Sentiment analysis dual purpose: (1) flag frustrated customers, (2) cross-check agent reports vs.
1 reality

1 Predictive analytics for scheduling (call volume curves, special events, Manila timezone example)
2

1 Air Canada hallucination case (AI promised fake refund, court ruled company liable, legal
3 implications)

1 Pre-LLM vs. post-LLM (most current chatbots still use old tech, negative perception from legacy
4 systems)

1 Case studies: Zendesk (15% resolution, 20% CSAT) and BofA Erica (19.5M users, 105M requests,
5 proactive)

1 Aztea scenario: the 6 recommended AI interventions and Dr. Bose’s personal parallel story
6

1 Cost reality: AI cost must be lower than the $3–10/hr agents it replaces, fundamental business
7 case math

1 Customer loyalty: definition, 9 strategies, connection to CS quality

8

1 AI use cases table: 7 use cases with specific AI types and real company examples
9