Digital Transformation & AI

Session 13: Prediction Machines, ML and AI Canvas

Sridhar Srinivasan
Recap of topics introduced

• Data characteristics

• AI vs ML vs DL: what is, when to use

• Programmed approach vs ML approach

• ML project process model

• The Tesla advantage

Types of ML methods
• Supervised Learning
• Unsupervised Learning
• Reinforcement Learning
• Transfer Learning
• ..
Supervised Learning
Rooms Age … Price Product Price … Buy / Not
Skew buy

Grad Work Ex … Campus Movie # of songs … Hit / Flop

Job Pay Length

• Supervised learning models a target column y as a function of a set of predictor columns X1, X2 etc.
• So, y = f(X)
• If y is continuous numeric, it is a regression scenario
• If y is categorical, it is a classification scenario
Unsupervised Learning
• There are many scenarios under this method.
• Clustering scenario involves grouping similar samples based on
distances. This is popularly used in recommenders, market
segmentation, document clustering, anomaly detection etc.

User Song 1 Song 2 Song 3 Song 4 Song 5

1 0 0 3 5 0 Data showing no
2 12 1 1 9 7 of times users
played songs
3 8 0 0 6 6
4 0 1 4 6 0
5 0 0 2 3 0
• Imagine you’re playing a video game where you control a character (the Agent). The game world is
the Environment.
• State (( S_t )): This is the current situation or position of your character in the game. For example, your
character might be standing at the start of a level.
• Action (( A_t )): This is the move you decide to make, like jumping over an obstacle or picking up a power-up.
• Environment Response: After you take an action, the game updates. This gives you a new situation or state ((
S_{t+1} )) and a reward (( R_{t+1} )). For example, if you successfully jump over an obstacle, you move to the
next part of the level and might get points as a reward.
• Reward (( R_{t+1} )): This is the feedback you get from the game. If your action was good, you get a positive
reward (like points). If it was bad, you might get a negative reward (like losing a life).
• Learning: Your character (the agent) uses this feedback to learn. Over time, it figures out which actions lead to
the best rewards and starts making better decisions.
Applications of Reinforcement Learning

Optimize
Learn to play
Navigation, treatment Portfolio
games e.g. Go,
control of robots strategies, Management
Chess
perform surgeries

Optimize
Recommendation
manufacturing Traffic control
systems
process
Transfer Learning
• Involves taking a model that has already been trained on one task and
applying it to a different but related task

• Pre-trained Model: A model is trained on a large dataset for a specific

task.
• For example, a model might be trained to recognize objects in images.

• Transfer Knowledge: The knowledge gained from this task is then used
to help solve a new, but related task.
• For instance, using the object recognition model to identify specific types of
animals in images.
What is the core
competency of Computers?
 data
processing

cryptography …

performing
arithmetic
calculations networking
simulations
& comm.

graphics running
and image software
processing applications
What is the core
competency of AI?
 next word
generation

… image
recognition

prediction
content
recommendation forecasting

autonomous sentiment
cars classification
Uses of AI’s ability to predict

USE PAST SALES OF HOUSES, USE PAST PATIENTS’ HISTORY USE EXAMPLES OF CAT USE CUSTOMER DATA TO
TO PREDICT A PRICE OF A NEW TO PREDICT REMISSION OF IMAGES TO RECOGNIZE IF A PREDICT / CLUSTER SIMILAR
ONE LISTED CURRENT PATIENT NEW IMAGE IS A CAT CUSTOMERS
When cost of an input becomes cheaper

That input is used more in existing uses

More ways are discovered to use that input

The demand for complementary items

increases
AI Canvas
Action: What are you trying to do? Based on the decision
Input: What data do you need to run the predictive algo?
Prediction: What is the prediction to be made?
Judgment: How do you value different outcomes and errors?
Outcome: What are success metrics?
Feedback: How can you use the outcomes to improve the algo?
Training: What data do you need to train the predictive algorithm?
Class discussion :
Home alarms

• Reports say that 97% of the alarms

from home security are false alarms.
• How can AI help improve the
response to these?
• AI can make prediction cheaper, faster
and hence scale decision making also
• What does it take to include such a
machine in your decision process?
How can AI improve the way security firms respond
to home security alarms?

• Apply AI canvas:
• Prediction: What is the prediction to be made?
• Judgment: How do you value different outcomes and errors?
• Outcome: What are success metrics?
• Feedback: How can you use the outcomes to improve the algo?
• Input: What data do you need to run the predictive algo?
• Action: What are you trying to do?
• Training: What data do you need to train the predictive algorithm?
AI canvas: How to improve responses to home
security alarms using AI?

• Prediction: predict if alarm was caused by unknown intruder

• Judgment: Cost of responding to a false alarm to cost of not responding to a true
alarm
• Action: Dispatch a security response to alarm or not
• Outcome: Observe if action taken to alarm was correct
• Input: Data from cameras, sensors, contextual data at that time e.g. street level
• Training: Historical sensor data matching with outcome data e.g. history of false
alarms or history of true alarms
• Feedback: Site data matched with data from outcomes, improve model
prediction
Class discussion: Healthcare
Emergency Response
• Context: Emergency rooms (ERs) in India face significant challenges due
to high patient volumes and limited resources. Efficient triage is crucial
for prioritizing patients and optimizing resource allocation.

• India has one of the highest burdens of emergency cases globally, with
emergency departments (EDs) in major hospitals like AIIMS, New Delhi,
receiving approximately 150,000 patients annually. Overcrowding in
Indian ERs is a critical issue, leading to increased wait times and strain
on healthcare resources.
• Acute conditions like coronary syndrome, stroke, respiratory diseases,
and injuries are leading causes of death and disability in India,
highlighting the need for efficient emergency care.
• The implementation of structured triage systems, such as the AIIMS
Triage Protocol (ATP), has shown improvements in managing patient flow
and reducing waiting times.
How can AI improve healthcare emergency
response?
• AI canvas:
• Prediction: What is the prediction to be made?
• Judgment: How do you value different outcomes and errors?
• Outcome: What are success metrics?
• Feedback: How can you use the outcomes to improve the algo?
• Input: What data do you need to run the predictive algo?
• Action: What are you trying to do?
• Training: What data do you need to train the predictive algorithm?
Healthcare Emergency Response
• Prediction: Predict the severity of incoming emergency cases.
• Judgment: Weigh the cost of misclassifying a severe case as non-severe
against the cost of over-allocating resources to non-severe cases.
• Action: Allocate medical resources and prioritize patients based on
predicted severity.
• Outcome: Observe if the resource allocation and prioritization were
appropriate.
• Input: Data from patient symptoms, medical history, vital signs, and
contextual data (e.g., time of day, location).
• Training: Historical patient data, including symptoms, outcomes, and
resource usage.
• Feedback: Match patient outcomes with initial predictions to refine the
model.
Class discussion: Fraud
Detection in Financial
Transactions
• Context: Predicting fraudulent activities in real-time to prevent
unauthorized transactions and reduce financial losses.

• Merchant losses due to fraud are projected to reach $38 billion

in 2023. AI-based fraud detection systems can catch up to 94%
of fraudulent transactions in real-time
• Banks implementing real-time AI fraud solutions have seen an
80-90% reduction in false positives. Financial institutions can
save between $3 million and $15 million annually by preventing
fraud through AI. Approximately 63.6% of financial institutions
using automated fraud prevention methods successfully
prevent fraud before it occurs

• These statistics highlight the critical need for advanced AI

solutions in fraud detection and the substantial benefits they
can bring to financial institutions.
Fraud Detection in Financial Transactions
• AI canvas:
• Prediction: what is the decision to be made?
• Judgment: How do you value different outcomes and errors?
• Outcome: What are success metrics?
• Feedback: How can you use the outcomes to improve the algo?
• Input: What data do you need to run the predictive algo?
• Action: What are you trying to do?
• Training: What data do you need to train the predictive algorithm?
Fraud Detection in Financial Transactions
• Prediction: Predict the likelihood of a transaction being fraudulent
• Judgment: Balance the cost of false positives (legitimate transactions flagged as
fraud) against the cost of false negatives (fraudulent transactions not detected)
• Action: Approve or flag transactions for further review based on the prediction
• Outcome: Monitor the accuracy of fraud detection and the impact on customer
experience
• Input: Transaction data, user behavior patterns, historical fraud data, and
contextual data (e.g., location, time)
• Training: Historical transaction data labeled as fraudulent or legitimate
• Feedback: Update the model based on the outcomes of flagged transactions and
customer feedback
Digital Transformation & AI
Session 14: AI Human Collaborative Intelligence
Sridhar Srinivasan
Topics

How is AI being deployed in organizations

How can employees adapt to AI and redefine

their roles
What is the partnership model between AI and
humans
How is AI being deployed in
organizations
How is AI being
deployed in
organizations

• Cognitive process automation (CPA)

• Cognitive insights
• Cognitive engagement
CPA
• Use of AI technologies to automate complex business
processes that typically require human intelligence
• This includes tasks such as decision-making, problem-
solving, and learning from data.

• Examples:
• AI-powered chatbots, intelligent document
processing, and automated decision-making systems
The evolution of process automation

Autonomous
systems
Hyper
automation •end to end
systems
Cognitive •end to end automation
process
Process automation
Automation
Robotic •Model enabled
Process execution of tasks
requiring
Automation intelligence
•software to
Manual execute rule-
processes based tasks
Class discussion
In organizations, employees often create tickets to resolve various
issues. There are some frequently raised type of tickets and some
rare complex ones.

1. Traditionally, the ticket is received by a helpdesk team who

study the ticket and assign to the right team / specialist.
2. The recipient then proceeds to engage and resolve the ticket.

How can RPA and CPA be used to add value in this scenario?
Cognitive insights
• Detecting patterns in data and interpreting their meaning using
statistically-based machine learning algorithms
• Example:
• anomaly detection, forecasting
Class discussion

• The Singapore Changi Airport Group

wanted to use AI to predict the arrival
time of long-haul flights.

• Apply the six stages of ML projects

process to this.
Cognitive engagement
• Engaging employees and/or customers using natural language
processing chatbots, intelligent agents and machine learning
Cognitive engagement
• After you successfully delivered the long-haul flights arrival time
project, the management of CAG was impressed and want your help
with another initiative.

• They want to use a chatbot to help passengers interact, get the

information needed regarding various aspects like arrival, departure,
seats availability, fares, food etc.

• What are the benefits?

• How will you approach?
• What are the challenges you foresee?
 Predictive maintenance

Shopping suggestions

Class discussion
Fraud detection
We saw three types of AI applications
being used in the industry
Flight ticket booking assistant
- Cognitive Process Automation
- Cognitive Insights
- Cognitive engagement Personalized learning

Identify which type is relevant in these.

Supply chain optimization

Virtual health assistant

How can employees adapt to
AI and redefine their roles
The Five Step model
• These are ways that people can ‘step’ in response to AI adoption

• Step-up
• Step-in
• Step-forward
• Step-narrowly
• Step-aside
Step-up
• At org level or unit level Executives who will make the decisions on
when, where and how to invest in AI capabilities across the
portfolio of the organisation’s business needs
Step-in
• Symbiotic Relationship Between Experts and AI Systems
• Suggests the best sources of data to use to train the machine
• Knows exception scenarios beyond available data that models
can learn from
• Recognizes new patterns in data that warrant re-training of model
Step-forward
• Analytics specialists, data science professionals who develop, test and
deploy the new models
• Employees and vendor associates
Step-narrowly
• Those involved in tasks where ROI of automation is not high

• This could be due to small size of teams or nature of work involved

• E.g. Private Wealth Manager, Forensic accountant

Step-aside
• Roles which will make way for AI to do what it is best at and
redefining their roles around creativity or empathy
• Their earlier role was customer facing, they will now help explain the AI
outputs to customers, motivate them to use / lower resistance;
• Their roles will focus more on enablement of usage
Class discussion
• Comment on the frequency distribution of these roles

• Step-in: Domain experts collaborate with AI, identifying data sources and
validating outputs
• Step-forward: IT and data science professionals develop and implement
AI solutions
• Step-narrowly: Specialists continue their highly specialized roles with
minimal changes
• Step-up: Executives and leaders prioritize and select AI applications
• Step-aside: Customer-facing roles transition to explain AI outputs and
encourage usage
What is the partnership model
between AI and humans
The Case for Human-Machine Collaboration
• Profitability requires growth and productivity
• Growth requires adaptation to change
• Productivity demands efficiency and standardization

• Human Strengths:
• Flexible, adaptable, and dynamic.
• Capable of understanding complex, unquantifiable contexts
• Generate ideas without needing prior data

• Machine Strengths:
• Highly efficient and consistent
• Perform at high productivity levels
• Limited in handling uncertainty and rapid changes
The missing middle

How can humans and machine

100%
collaborate to produce superior Automation
Human
results?
AI – Human partnership models
• AI extended by human
• Human extended by AI
Two modes of human thinking
• System 1 mode: Intuitive, effortless, quick, implicit, emotional
• System 2 mode: Logical, deliberate, takes time

When under uncertainty, people fall back on System 1 mode

Thinking, Fast and Slow, Daniel Kahneman

The upside of System 1
• Context Understanding: Grasping physical, social, and situational contexts.
• Intent and Needs Recognition: Understanding the intent and needs of others.
• Commonsense Reasoning: General-purpose reasoning without domain-specific
training.
• Associative Reasoning: Using prior knowledge to generate suggestions even with
limited or no specific data.
• Question Identification: Identifying important questions and strategic issues.
• Hypothetical Inquiry: Asking and exploring ‘what if’ questions and counterfactual
scenarios.
• Multiple Perspectives: Viewing situations from various angles.
Source:
AI: Augmentation, more so than automation
Steven M. MILLER
AI extended by human

Prepare datasets to train models

Train Train on how to interact with humans

Ability to develop and share ‘Chain of

Explain thought’ to improve trust in predictions
“Why did you do / say / predict that?”

Help AI continue to perform properly, safe,

Sustain responsibly
3H - Helpful, Honest and Harmless
Human extended by AI

Amplify
• Scale human ability to process big data, multi criteria scenarios, optimize
applying conflicting priorities
• Generate options quickly complying with given expectations
• Help handle “What if” queries

Interact • Use natural language abilities to transcribe meetings, search videos, interact at
scale with employees conveying messages and taking feedbacks
• Superior version of IVR first level menu

Embody • With their sensors, motors, actuators and materials suitable for conditions they
can work alongside humans in factories, mines, disaster scenes and labs
• Hyundai “Cobot”
The traditional approach
• Minds - people think up or invent a better way of doing something
• Products - to monetize the idea
• Core - build a company around making that product or service
Three trends shaping modern economy
Traditional Emerging

Mind Machine

Product Platform

Core Crowd

From economies of scale to economies of scalable learning

 Flexibility

5 areas of Speed
improvements
for
Scale
organizations Decision-making
Personalization
Digital Transformation & AI
Session 15: Stitch Fix case study
Sridhar Srinivasan
Topics
• Business topics in case
• ML Topics in case
• Case analysis
Business Topics
Business concepts
• Funding series
• Pitch to investors – equity for funds
• Inventory turnover
• Buy on consignment
• Strategy
• Business plan
• Key metrics for ecommerce firms
• WOM vs ad spend
• Blue ocean strategy basics
Funding series
• Seed Funding
• Series A, B, C funding
• IPO
Pitch to investors
• Funds in exchange for equity stake
• Funds as debt
• Other creative forms

What do investors look for?

Inventory turnover
• No of turns = COGS / avg inventory
• Shows the ability to leverage working capital efficiently
Buy on consignment
• An arrangement where one person or entity of gives another
custody over materials or goods to sell while still legally owning
them
• When a sale is made, the supplier is paid
What is a business model?
Customer value proposition
• Statement of the value that the company intends to provide

Profit formula
• The business model generates value for your company through factors such as revenue model,
cost structure, margins, and inventory turnover.

Key resources and processes

• Has resources and processes required to deliver value proposition to targeted customers.
• Resources: people, technology, products, facilities, equipment, and brand
• Processes: training, manufacturing, service

“All it really meant was how you planned to make money”

What is a Strategy?
Strategy is a plan of action designed to
achieve a long-term or overall aim
 Mission and Target Value
Vision segments proposition

Competitive Revenue
Elements of positioning Streams
Channels

business
strategy Cost structure
Resource
Allocation
Organizational
Structure

Alliances Key activities

Key metrics for ecommerce firms
1.Customer Acquisition Cost (CAC): The cost associated with acquiring a new customer.
2.Customer Lifetime Value (CLV): The total revenue expected from a customer over their
entire relationship with the business.
3.Cart Abandonment Rate: The percentage of shopping carts that are abandoned before the
purchase is completed.
4.Average Order Value (AOV): The average amount spent each time a customer places an
order.
5.Return Rate: The percentage of products that are returned by customers
WOM impact
Consider this scenario
• Customer Acquisition Cost = 100 Rs
• New customers added = 1000
• Referral i.e. WOM customers = 200
• So, savings = 200 * 100 Rs
Blue Ocean Strategy
• Value innovation – differentiation at low cost
• Creates uncontested space, makes competition irrelevant
• Red approach takes industry structure as given and attempts competitive
advantage within that structure
• More value @ higher cost or comparable value @ lower cost
Blue vs Red
 Strategy Canvas
6
5
4
3 value(stitch fix)
2 value(amazon)
1
0
1 2 3 4 5 6 7

Competitive Factors Value (SF) Value(Amazon)

1. Personalization High 5 Medium 3
2. Customer Engagement Medium 3 High 5
3. Variety of Products Medium 3 High 5
4. Styling Expertise High 5 Medium 3
5. Pricing Medium 3 Medium 3
6. Convenience High 5 High 5
7. Data-Driven Recommendations High 5 High 5
ML Topics
ML topics
• Data platform
• Use of mobile app for data collection and engagement
• Use of social media – Pinterest Stich Fix inspirations boards to share
preferences with stylists
• Quality of data
• Rank score of inventory
• Customer States
• State transition matrix - deterministic
• Markov chain models - probabilistic
• Genetic algorithms – iteratively evolve till best results reached
Data sources to build Customer State
1. Customer Registration Step: Collects basic demographic information and style preferences.
2. Customer Social Media Handles: Analyzes social media profiles for style insights.
3. Fix Box Feedback: Gathers detailed feedback on each item sent in the Fix box.
4. Use of iOS App: Tracks user interactions and preferences through the app.
5. Finery: Integrates wardrobe data from the Finery app.
6. Style Profile Surveys: Detailed questionnaires about style preferences, sizes, and lifestyle.
7. Browsing Behavior: Tracks how customers interact with the website and app, including items viewed and time spent.
8. Customer Reviews: Collects and analyzes reviews and ratings of items.
9. Social Media Engagement: Monitors likes, shares, and comments on Stitch Fix’s social media posts.
10.Email Interactions: Tracks responses to marketing emails and newsletters.
11.Referral Programs: Gathers data from referral activities and new customer sign-ups.
12.Stylist Notes: Uses insights from personal stylists’ notes and interactions with customers.
13.Event-Based Data: Collects data related to specific events or promotions, such as holiday shopping behavior.
14.Third-Party Data: Integrates data from third-party sources for additional insights.
Data platform
• provides the infrastructure and tools needed to derive valuable
insights from data, supporting data-driven decision-making

• Key components typically include:

• Data Collection and Ingestion
• Data Storage
• Data Processing
• Data Analysis and Visualization
• Data Governance
Sources of data, quality of data
• Registration stage
• Social media handles
• Customer feedback on fixes
• Pinterest Stitch Fix inspirations
• Style Shuffle iOS app
Algorithms
• Creates client “states”
• Tracks all interactions and creates a state
• Detects changes and triggers
• Uses state transition matrix and Markov chain models to predicts demand
• Rank-order inventory for each customer
• Recommendation systems
• Warehouse assignment
• Intelligence logistics
Case analysis
Case discussion topics
• Why did investors show reluctance to fund?
• Lake learnt from three industries to create SF – retail, personal stylists
and subscription boxes. What did she take / leave from each?
• For a subscriber, how does SF add value compared to stores options?
• How does this compare with Amazon’s offering? Should SF be
concerned?
• Finery adds an additional barrier, value and makes it harder for
imitation. How can SF use Finery to add value to customers?
• How do you SF continue to grow in rev and profit?
Why did investors show reluctance to fund?
• Lake saw value in the data she could collect by owning
• Investors wanted to avoid the cost of buying
• Lake could envision her business model, but it was not so obvious
to others
Lake learnt from three industries to create SF – retail, personal
stylists and subscription boxes. What did she take / leave from
each?

Industry Borrowed Left out

Retail Try before you buy Physical stores

Convenience of online stores, Free Items selected by buyer

delivery & easy returns
Personal Stylists Personalized styling Premium
Subscription boxes Surprise element Outside in instead of inside out
From a buyer perspective, what are the unique ways SF offered a leap in
value beyond other retailers (brick & mortar, online)? How do they
compare with Amazon offerings, esp. Personal Shopper?

In each utility lever,

use a S to show what red
ocean is offering and a W to
show what is a blocker

www.blueoceanstrategy.com
Buyer Utility Levers
1.Customer Productivity: Enhancing the efficiency and effectiveness with which customers
can achieve their goals, often by simplifying processes or providing tools that save time
and effort.
2.Simplicity: Making products or services easy to understand and use, removing
unnecessary complexity to improve the customer experience.
3.Convenience: Offering products or services in a way that fits seamlessly into the
customer’s lifestyle, making them easily accessible and reducing effort required from the
customer.
4.Risk: Minimizing the perceived or actual risks associated with using a product or service,
such as financial, physical, or emotional risks, to build customer trust and confidence.
5.Fun and Image: Creating products or services that provide enjoyment and enhance the
customer’s self-image, making the experience more enjoyable and appealing.
6.Environmental Friendliness: Designing products or services that are sustainable and
have a minimal negative impact on the environment, appealing to eco-conscious
customers.
 Utility Map – In-store retailers
Visit Store Select Try Finalize Buy Return
Customer Productivity

Simplicity

Convenience

Risk reduction

Fun & image Utility Map – In-store retailers

Environment friendliness Visit Store Select Try Finalize Buy Return
Customer Productivity

Simplicity

Convenience

Risk reduction

Fun & image

Environment friendliness
Utility Map – Stitch Fix
Visit Store Select Try Finalize Buy Return
Customer Productivity

Simplicity

Convenience

Risk reduction

Fun & image

Environment friendliness
 How does this compare with Amazon’s
offering? Should SF be concerned?
Identify 1-2 points where SF can score over Amazon and continue to
grow?
What are the enablers of these?

Utility Map – Stitch Fix Utility Map – Amazon Personal Shopper

Visit Store Select Try Finalize Buy Return Visit Store Select Try Finalize Buy Return
Customer Productivity Customer Productivity

Simplicity + Simplicity

Convenience Convenience

Risk reduction + Risk reduction

Fun & image Fun & image

Environment friendliness Environment friendliness

Finery adds an additional barrier, value and makes
it harder for imitation. How can SF use Finery to
add value to customers?
• Finery gives SF access to info on the 70% clothing customers are
buying elsewhere
• Can launch more services learning what they are missing
currently
• Adds entry barrier in a legal way – IP purchase
How do you SF continue to grow in rev and
profit?
• Costs
• Customer acquisition
• Inventory turnover
• Logistics
• Risks
• Expansion
• Inducting new associates
• Pricing
• Manage price rise vs churn
• Rediscover
References
• https://www.youtube.com/watch?v=BbxHxNb5TOQ
• What is Stitch Fix?

• https://www.cnbc.com/video/2018/10/09/finery-ceo-on-disrupting-
the-fashion-industry.html
• What is Finery?

• https://www.youtube.com/watch?v=OsTnOerTE9U
• https://www.youtube.com/watch?v=i3leT-hWa7E
• Amazon Personal Shopper is shutting down
Digital Transformation & AI
Session 16: "Storrowed" case study
Sridhar Srinivasan
Topics
• Recap of ML methods
• Supervised Learning
• Unsupervised Learning
• Reinforcement Learning
• Transfer Learning
• Recap of CRISP-DM process model
• Image basics
• Convolutional Neural Networks
• Image segmentation
• “Storrowed” case study
Recap
Supervised Learning
Work Ex CGPA Degree Campus Work Ex CGPA Degree Got
Job Pay Placed?

Regression - Continuous numeric label Classification - Categorical label

Common algorithms to solve regression Common algorithms to solve classification

• Linear Regression • Logistic Regression
• K Nearest Neighbors Regressor • K Nearest Neighbors Classifier
• Decision Tree Regressor • Decision Tree Classifier
• Deep Neural Network Regressor • Deep Neural Network Classifier
• … • …
Unsupervised Learning - Clustering
Customer ID Film 1 Film 2 Film 3 … …

Voter ID Topic 1 Topic 2 Topic 3 … …

Common algorithms to solve create

clusters
• K-Means Clustering
• Hierarchical Clustering
• …
Reinforcement Learning

• Applications
• Navigation, control of robots
• Learn to play games e.g. Go,
Chess
• Optimize treatment strategies,
perform surgeries
• Portfolio Management
• Involves taking a model that has already been trained on one
task and applying it to a different but related task

• Pre-trained Model: A model is trained on a large dataset for

a specific task.
• For example, a model might be trained to recognize
objects in images.

• Transfer Knowledge: The knowledge gained from this task is

then used to help solve a new, but related task.
• For instance, using the object recognition model to
identify specific types of animals in images.
 Typical ML Project – the process model

Cross Industry Standard

Process for Data Mining

(CRISP-DM) is a standard

process used for data mining

8
Image basics

Upload image and inspect pixel values https://pixspy.com/

How digital images are stored, pixel values, RGB https://www.analyticsvidhya.com/blog/2021/03/gra
yscale-and-rgb-format-for-storing-images/

9
CNN basics
• Convolutional Neural Networks are used to handle images and
videos

https://developers.google.com/machine- Short video on CNN

learning/practica/image-classification
https://cloud.google.com/vision/docs/drag-and- Drag and drop an image and see object
drop recognition
https://www.dtksoft.com/image-recognition- Drag and drop any image of vehicle and license
demo plate and can be recognized
https://deeplizard.com/resource/pavq7noze2 CNN demonstration tool
https://poloclub.github.io/cnn-explainer/ Online tool demonstrating CNN
https://saturncloud.io/blog/a-comprehensive- Blog post describing CNN steps
guide-to-convolutional-neural-networks-the-eli5-
way/ 10
Image segmentation

Upload an image and hover mouse to see image https://segment-anything.com/demo#

segmentation

11
Case Study : “Storrowed”

‘Storrowing:' A Boston tradition officials wish would https://www.nbcboston.com/news/local/storrowin

end g-a-boston-tradition-officials-wish-would-
end/3128351/

WBC Question Everything – how to avoid storrowing https://www.youtube.com/watch?v=Oh7YDHY_Gvc

12
What business
are we in?

13
Class exercise
• Research the topic to arrive at a problem statement to take
forward

14
Business Understanding

15
Business Understanding

What is the Why is it worth Who is the Who are

problem? solving? sponsor? stakeholders?

What has been What is the current

Why are these Suggest success
done so far to value of this
steps falling short? metric(s)?
prevent? success metric?

What are the

causes?

16
 •Context
•Background information or scenario setting.
•Role Play by Agent
•Specify the role or persona the AI should adopt.
•Task Definition
•Clearly state the specific task or question.
•Specificity
•Include precise details or requirements for the response.
•Constraints
•Define any limits (e.g., word count, tone, style).
•Examples
•Provide examples of desired outcomes or responses.
•Input and Output Format
•Specify how the input should be handled and how the output should be

ChatGPT Prompts structured.

•Clarifications and Edge Cases
•Mention any ambiguities or exceptions to address.
- best practices •Tone and Style
•Indicate the desired tone (formal, casual) and style (narrative, bullet
points).
•Iteration and Refinement
•Suggest steps for refining or iterating on the response.
•Error Handling
•Include instructions for managing potential mistakes or
misunderstandings.
•Multi-Step Guidance
•Break down the task into smaller steps if needed.
•Bias Mitigation
•Instruct the AI to avoid specific biases or perspectives.
•Reinforcement
•Reinforce key points or priorities in the task.
•Ethical Considerations
•Address any ethical concerns or guidelines. 17
Prompt structure
Clearly state what you want the AI to do (e.g.,
Clear Task Definition:
generate, summarize, analyze).

Provide necessary background information or

Context:
context to guide the AI's response.

Be specific about any requirements, constraints, or

Specific Instructions:
desired outcomes (e.g., word limit, tone).

Include examples or templates to illustrate the

Examples (if needed):
desired response style or format.

Conclude with a direct question or command to

End with a Question or Command:
prompt the AI to generate the response.

18
Chain of Thought Prompting
• This is a technique used in prompt engineering to guide the AI
through a logical reasoning process step by step
• This method helps the AI break down complex tasks or problems
into smaller, more manageable steps, ultimately leading to a more
coherent and accurate response

19
Few-Shot Learning
• Provide the model with a few input-output pairs as examples
within the prompt, and then ask the model to generate the output
for a new input based on those examples

Example 1:
Text: "I loved the movie, it was amazing!"
Sentiment: Positive

Example 2:
Text: "The food was terrible, I would not recommend it."
Sentiment: Negative

Now, classify the sentiment of this text: "The service was good, but the food was just okay."
20
Multi-Step Prompts
• This involves breaking down a complex task into a series of
smaller, sequential prompts
• This approach guides the AI through a series of steps, each
building on the previous one, to achieve a more comprehensive
and accurate outcome
Task: Summarize a research paper.
Step 1: "First, list the key points discussed in the introduction of the research
paper."
•AI Response: Lists key points from the introduction.
Step 2: "Now, summarize the methodology used in the study."
•AI Response: Provides a summary of the methodology.
Step 3: "Finally, what are the main findings and conclusions drawn in the paper?"
•AI Response: Summarizes the findings and conclusions.
21
Which approach?
• There are basically different approaches possible
• List some possible approaches
• Consider an approach and indicators that are conducive to using
that

22
Class exercise
• Research the topic to arrive at solution ideas

23
Data Understanding

24
Data Understanding

NYC open data – info on low bridges, past incidents https://data.cityofnewyork.us/Transportation/NYC-

of bridge strikes Bridge-Strike-Data/jdn9-td9w/about_data

https://data.cityofnewyork.us/Transportation/Citywid
e-Low-Bridges/rn6h-i66u/about_data

25
Model building

26
Model building
• This approach depends on the datasets available and the
approach adopted

• For e.g. a traditional ML approach using datasets so a set of

predictors can predict that a truck is expected to get storrowed

• For e.g. a image recognition approach can be used to send alert to

the driver on mobile

27
Model evaluation,
deployment

28
What follows?

29
Digital Transformation & AI
Session 17: Generative AI and Future of Work
Sridhar Srinivasan
Topics
• Recap: ML Methods
• Recap: Prompting Best practices
• Problem Formulation
• Article: AI Prompt Engineering Isn’t the Future
• Article: To work well with Gen AI you need to learn how to talk to it
• Class Exercise: Future of Work tasks
Recap
ML Methods and DL techniques
ML Method Characteristics
Involves modelling a target variable using a set of predictors. The
Supervised Learning target can be continuous (Regression) or categorical
(Classification)
Includes techniques like clustering, which involves identifying
Unsupervised Learning
and grouping similar instances
Deep Learning A subset of ML that utilizes neural network architectures
Involves retraining a pre-trained model for a related but different
Transfer Learning
task
Focuses on learning to perform tasks in an environment to
Reinforcement Learning
maximize rewards through actions
A deep learning technique used to recognize patterns in images
Convolutional Neural Networks
and videos
 •Context
•Background information or scenario setting.
•Role Play by Agent
•Specify the role or persona the AI should adopt.
•Task Definition
•Clearly state the specific task or question.
•Specificity
•Include precise details or requirements for the response.
•Constraints
•Define any limits (e.g., word count, tone, style).
•Examples
•Provide examples of desired outcomes or responses.
•Input and Output Format
•Specify how the input should be handled and how the output should be
ChatGPT Prompts structured.
•Clarifications and Edge Cases
•Mention any ambiguities or exceptions to address.
- best practices •Tone and Style
•Indicate the desired tone (formal, casual) and style (narrative, bullet
points).
•Iteration and Refinement
•Suggest steps for refining or iterating on the response.
•Error Handling
•Include instructions for managing potential mistakes or
misunderstandings.
•Multi-Step Guidance
•Break down the task into smaller steps if needed.
•Bias Mitigation
•Instruct the AI to avoid specific biases or perspectives.
•Reinforcement
•Reinforce key points or priorities in the task.
•Ethical Considerations
•Address any ethical concerns or guidelines.
5
Prompt structure
Clearly state what you want the AI to do (e.g.,
Clear Task Definition:
generate, summarize, analyze).

Provide necessary background information or

Context:
context to guide the AI's response.

Be specific about any requirements, constraints, or

Specific Instructions:
desired outcomes (e.g., word limit, tone).

Include examples or templates to illustrate the

Examples (if needed):
desired response style or format.

Conclude with a direct question or command to

End with a Question or Command:
prompt the AI to generate the response.

6
Chain of Thought Prompting
• This is a technique used in prompt engineering to guide the AI
through a logical reasoning process step by step
• This method helps the AI break down complex tasks or problems
into smaller, more manageable steps, ultimately leading to a more
coherent and accurate response
Solve this problem and show the steps you used to arrive at the final answers. Also, verify the
answers.
2x + 3y = 13
x–y=1

a driver was going in the wrong direction in a one way street but the policeman did not fine
him. why? give the answer and show how you arrived at that.
7
Few-Shot Learning
• Provide the model with a few input-output pairs as examples
within the prompt, and then ask the model to generate the output
for a new input based on those examples

Example 1:
Text: "I loved the movie, it was amazing!"
Sentiment: Positive

Example 2:
Text: "The food was terrible, I would not recommend it."
Sentiment: Negative

Now, classify the sentiment of this text: "The service was good, but the food was just okay."
8
Multi-Step Prompts
• This involves breaking down a complex task into a series of
smaller, sequential prompts
• This approach guides the AI through a series of steps, each
building on the previous one, to achieve a more comprehensive
and accurate outcome
Task: Summarize a research paper.
Step 1: "First, list the key points discussed in the introduction of the research
paper."
•AI Response: Lists key points from the introduction.
Step 2: "Now, summarize the methodology used in the study."
•AI Response: Provides a summary of the methodology.
Step 3: "Finally, what are the main findings and conclusions drawn in the paper?"
•AI Response: Summarizes the findings and conclusions.
9
Topics
• Recap: ML Methods
• Recap: Prompting Best practices
• Problem Formulation
• Article: AI Prompt Engineering Isn’t the Future
• Article: To work well with Gen AI you need to learn how to talk to it
 Article: AI Prompt
Engineering Isn’t the Future

11
Prompt Engineering
• Prompt engineering involves crafting and refining the input
prompts given to AI models, particularly language models like
GPT, to elicit the desired response or behaviour

• Prompt engineering may cease to be a hot skill as Gen AI evolves

• Problem formulation can assist in improving the effectiveness of

Prompt Engineering.

12
Problem formulation
• Problem formulation is the process of defining and structuring a
problem in a way that makes it solvable using AI or other
computational methods

• Requires grasp of the problem domain and understand related

issues

13
 Diagnosing problems is still a
weakness in many organizations

One of the reasons it is underdeveloped is the

overemphasis on problem solving

14
Four components of
problem formulation

• Problem diagnosis
• Problem Decomposition
• Problem Reframing
• Problem Constraint design

15
Problem diagnosis
• Identifying the core problem (for AI to solve)
• What is the main objective to be accomplished?
• Examples of approaches –
• E.g. Five Whys to formulate problems

A superior problem formulation improves the quality and chances

of the solution
16
Problem decomposition
• Breaking down complex problems into manageable sub-problems

• Especially when involving multifaceted problems

• For e.g. develop a framework for cybersecurity of an organization
• If assigned as a prompt directly versus as subproblems

• Examples of decomposition: Functional decomposition or WBS

• Decomposition can
• help visualize complex problems
• simplify the identification of individual components and their interconnections

17
Problem reframing
• Involves changing the perspective and enabling alt interpretations
• This can also help broaden the scope of potential solutions
• Can also help workaround constraints
• See from multiple perspectives
• Explore analogies
• Use abstraction
• Proactively question the problem objectives
• Try to identify any missing components in the problem definition
(do not take a given problem statement as final)

18
Try Problem reframing in these cases
• Shortage of parking spaces at office building
• At airport arrival area, bags takes too long to arrive
• Office elevator is very slow
• There are frequent load shedding power cuts

19
20
Effective reframing
• Involve a team, induct them on the approach
• Bring outsiders into the discussion
• Write down alternate definitions to the problem
• Ask what is missing – people react to what they see, what is
missing has to be explicitly looked out for
• Ask for multiple categorization of the problem
• Analyze positive exceptions
• Question the objectives, surface latent aspects

21
Problem constraint design
• Delineate boundaries of problem by defining input, process and
output restrictions of the solution search
• Use constraints to direct solution efforts
• For tasks focused on getting things done efficiently, use clear rules and
goals
• For tasks that need creative thinking, play around with the rules to find
unique solutions

22
Article: To work well with
GenAI you need to learn
how to talk to it

23
How to speak to AI?
Computers need more explicitly stated context

Seek AI to generate 3 types of responses and select or synthesize

Use “few shot learning” when examples can help convey the task better

Enable humans to prompt better using template prompts instead of from

scratch
This Photo by Unknown Author is licensed under CC BY

Breaking down complex tasks into sub tasks

If one approach did not work, try using another way

Include data, facts and also context and structure

24
Retrieval
Augmented
Generation
(RAG)

25
Problem formulation using GPT
1. Research and Information Gathering: Use GPT to explore background information, trends, and
industry insights related to the problem
2. Brainstorming and Ideation: Generate potential problems and diverse perspectives using GPT to
identify key areas of focus
3. Problem Statement Drafting: Employ the "who, when, what, how, why" framework with GPT to draft a
clear and precise problem statement
4. Decomposition and Analysis: Break down the problem into smaller parts and explore root causes with
GPT's assistance
5. Reframing and Alternative Approaches: Use GPT to suggest alternative problem framings and explore
hypothetical scenarios
6. Constraint Identification: Identify and prioritize constraints with GPT to understand their impact on
problem-solving
7. Feedback and Iteration: Refine the problem formulation iteratively by incorporating GPT's feedback
and new insights
8. Finalization: Synthesize information into a coherent problem formulation and prepare presentations
with GPT's help

26
Problem statement
components
• Description of the current state
• Impacts & importance of these – why solve this?
• Stakeholders
• Metrics / KPI for status quo and future state
• Components of the problem, their relationships
• Constraints
• Priorities / Urgency / Time aspects
• Assumptions

27
Class Exercise using Gen AI
• Step 1: Develop a problem statement for one of these
• Step 2: Develop a 2-3 solutions using the problem statement
developed

A. Employers want reduce or remove WFH while employees prefer the flexibility

B. Govts are concerned about privacy & ownership of data shared by users when accessing GenAI

28
Detailed slides for
reference

29
Detailed version of how to use GPT for
problem formulation
• To develop a problem formulation using generative AI like GPT, a student can follow these steps to leverage the AI's capabilities effectively:
• 1. Research and Information Gathering
• - Initial Exploration: Use GPT to gather background information on the topic or industry related to the problem. Ask the AI for summaries of
relevant articles, case studies, or reports.
• - Trend Analysis: Request insights on current trends, challenges, and opportunities in the field to understand the broader context of the
problem.
• 2. Brainstorming and Ideation
• - Prompt Generation: Use GPT to generate a list of potential problems or challenges within a given context. This can help in identifying areas
that might need attention.
• - Diverse Perspectives: Ask the AI to provide different perspectives on the problem, considering various stakeholders and their potential
concerns.
• 3. Problem Statement Drafting
• - Structured Framework: Use the "who, when, what, how, why" framework to draft a comprehensive problem statement. GPT can help refine
each aspect by providing suggestions and examples.
• - Clarity and Precision: Request the AI to rephrase or simplify complex ideas to ensure the problem statement is clear and concise.
• 4. Decomposition and Analysis
• - Break Down Components: Ask GPT to help decompose the problem into smaller, manageable parts. This can include identifying sub-
problems or related issues.
• - Root Cause Analysis: Use the AI to explore potential root causes of the problem, asking for explanations or examples of similar situations.

30
..contd.
• 5. Reframing and Alternative Approaches
• - Reframe the Problem: Request the AI to suggest alternative ways to view or frame the problem, which can lead to innovative solutions.
• - Scenario Exploration: Use GPT to explore hypothetical scenarios or "what-if" questions to understand the implications of different problem
formulations.
• 6. Constraint Identification
• - List Constraints: Ask GPT to identify potential constraints or limitations that might affect the problem-solving process, such as budget, time,
or resources.
• - Prioritization: Request advice on which constraints are most critical and how they might influence the approach to solving the problem.
• 7. Feedback and Iteration
• - Review and Feedback: Use GPT to review the drafted problem formulation and provide feedback or suggestions for improvement.
• - Iterative Refinement: Continuously refine the problem statement by incorporating feedback and exploring new insights provided by the AI.
• 8. Finalization
• - Synthesize Information: Ask GPT to help synthesize all gathered information into a coherent and compelling problem formulation.
• - Presentation Preparation: Use the AI to draft a presentation or report that clearly communicates the problem formulation to stakeholders.

31
Additional reading material
• https://hbr.org/2024/01/to-solve-a-tough-problem-reframe-it

• https://www.ted.com/talks/mihaly_csikszentmihalyi_flow_the_se
cret_to_happiness?geo=hi&subtitle=en

32
Digital Transformation & AI
Session 18: Case Study – Lemonade Insurance
Sridhar Srinivasan
Topics
• Recap
• Case Study: Lemonade Insurance
What is a business
model?
What is a business model?
Customer value proposition
• The model helps customers perform a specific “job” that alternative offerings don’t
address.

Profit formula
• The model generates value for your company through factors such as revenue model, cost
structure, margins, and inventory turnover.

Key resources and processes

• Has resources and processes required to deliver value proposition to targeted customers.
• Resources: people, technology, products, facilities, equipment, and brand
• Processes: training, manufacturing, service

“All it really meant was how you planned to make money”

Class What are the
elements of
discussion business strategy?
 Mission and Target Value
Vision segments proposition

Competitive Revenue
Elements of positioning Streams
Channels

business
strategy Cost structure
Resource
Allocation
Organizational
Structure

Alliances Key activities

 Class discussion: Innovation domains

What are the various areas in which organizations can pursue AI-driven
innovations?
Innovation Domains
• Products: Developing new or improved products or services to meet evolving customer needs or
preferences.
• Services: Innovating in service delivery to provide customers with better experiences or more efficient
solutions.
• Processes: Streamlining internal processes and operations to improve efficiency and productivity.
• Business models: Exploring new ways of creating and capturing value, such as shifting from a
product-based to a service-based model.
• Customer experiences: Creating unique and personalized experiences for customers that
differentiate the organization from its competitors.
• Innovation in distribution channels: This involves exploring new ways to deliver products or services
to customers, such as through e-commerce platforms or subscription models.
• Innovation in marketing: This involves using new technologies and techniques to promote products or
services and reach new customers, such as through social media advertising or influencer marketing.
• Innovation in supply chain management: This involves optimizing the supply chain to reduce costs
and improve efficiency, such as through the use of automation, data analytics, and real-time tracking.
• Innovation in organizational culture: This involves fostering a culture of innovation within the
organization, such as by encouraging risk-taking, experimentation, and collaboration among
employees.
• Innovation in sustainability: This involves developing environmentally sustainable products or
processes, reducing waste, and minimizing the organization's carbon footprint.
Lemonade Insurance
• https://www.youtube.com/watch?v=tbOfboCmv6I
• Lemonade Insurance - How It Works
• https://www.youtube.com/watch?v=flSLI2JmWVE
• Lemonade insurance app
• https://www.youtube.com/watch?v=wrRLbbWLSJ4
• How Lemonade is using AI to Disrupt Insurance
Lemonade product portfolio
Class • What are the broader impacts of Lemonade on
discussion the industry?
Impact of AI on Lemonade’s Vision / Mission

• Enabled new business opportunities

• Created a more personalized and efficient approach to insurance
• Made insurance more affordable and accessible to a wider range of customers

Broader takeaways

• AI made the company's vision of creating a fairer and more transparent insurance
industry possible
• AI can be a powerful tool for disrupting traditional industries, enabling new business
models

This Photo by Unknown Author is licensed under CC BY-SA-NC

Class discussion

• How did Lemonade change the customer journey?

Customer journey transformed
1.Fully Digital Experience
2.AI-Powered Chatbots
3.Instant Quotes and Claims
4.Transparent Pricing Model
5.Social Impact Focus
6.Mobile Accessibility
7.Data-Driven Insights, personalized
Class discussion

• How is their strategy and business model

different from traditional companies?
How is Lemonade different?
• Built grounds up to modern realities
• Strategy
• Make it a lovable brand
• Social impact
• Capital intensive and longer gestation period

• Business Model
• Rely on AI
• Behavioral economics
• Social good oriented
Invoke virtuous cycle

• https://www.lemonade.com/blog/save-money-every-month/

An estimated 5% of policyholders who have had their claims paid, returned their money after
their stolen item was found
Class discussion

• How are their internal processes different?

Lemonade’s processes
• Focus on customer experience
• Digital enabled
• Reduce customer effort
• Reduce costs using trust and tech
• Fewer siloes
• Cross functional collaboration
Class discussion

• What helped Lemonade offer better

pricing?
Impact of AI on customer segmentation
• AI is helping the company to identify and target high-risk customers
• Traditionally, insurance companies used: manual underwriting,
statistical methods, behavioral economics to identify and target
high-risk customers
• AI can be used to identify patterns in data that would be difficult for
humans to see. For e.g., identify customers who are more likely to
file claims for water damage, targets these customers with
messages about how to prevent water damage, such as installing a
water leak detector
• AI can segment customers based on their needs and interests,
allows Lemonade to customize offerings. For e.g., Lemonade offers
different types of renters insurance policies, depending on the size
of the apartment and the number of roommates.
 Customers switching to Lemonade from market leaders cos
https://www.lemonade.com/blog/lemonade-exposed-eight-
months/

As more data is available, precise segments are identified and targeted with apt pricing
https://www.lemonade.com/blog/precision-underwriting/

https://www.lemonade.com/blog/lemonade-sets-new-world-
Class discussion

• What helped Lemonade handle claims settlement differently?

What helped Lemonade handle claims
settlement differently?
• Transparency
• Availability
• Agile
• Tech enabled submission
• Trust based
• Peer reviewed, community vote on disputes
• Risks covered thru superior clustering
• Reduced cost of processing
Class discussion – Agility

• Lemonade's agility allows it to quickly respond to market changes and customer

needs. Responded to new needs triggered by COVID:
• In 2020, Lemonade launched a new product called "Pet Insurance for
Pandemic Pets." This product was designed to meet the needs of pet owners
who were adopting new pets during the pandemic.
• In 2021, Lemonade launched a new feature called "Lemonade Giveback."
This feature allows customers to donate a portion of their premiums to a
charity of their choice.
• In 2022, Lemonade launched a new product called "Lemonade Car
Insurance." This product is designed to make it easier and more affordable for
people to get car insurance.
Discussion
1. What are the enablers of Lemonade’s agility in their key activities?
2. How does Lemonade use its agility to quickly respond to market changes and customer needs?
Agility enablers
Flat organizational Focus on Integrated
structure innovation processes

Cross functional
Tech enabled Data integration
collaboration

Data-driven Empowering
decision making culture
 While AI is a powerful tool that
can help businesses to create
Class new value and improve their
performance.

discussion What should businesses watch

out for when using AI to define
their business strategy?
What should businesses watch out for when using AI to
define their business strategy?

Businesses need to carefully consider how they can use AI to their

advantage.

Businesses need to tailor their strategy to their specific

circumstances.

As AI continues to develop, businesses will need to continue to

rethink their strategy in order to stay ahead of the curve.