BL603 Transcript Notes May 10-May 11

1. Course Philosophy: Decision-Centric Thinking

 Primary Goal: Help organizations (for-profit or NGO) make better
decisions.
 Key Applications: Visual analytics, product design, strategic
improvements – all centered around enhancing decision-making.

2. Socio-Technical Systems vs. IT

 IT: Hardware, software, infrastructure (e.g., Microsoft Office, networks).
 IS (Information Systems): IT + people + processes + organizational
management.
 Socio-technical lens: Integrates management, anthropology, and
data science with tech.

3. Organizations as Systems of Processes

 Post-WWII capitalism shaped the view of organizations as mechanistic
systems.
 Business processes are visualized as diagrams (blocky charts) showing
flows of work.
 Employees = “human resources” = part of larger productivity
machinery.
 Continuous improvement is a default assumption (driven by
capitalism/competition).

4. Hierarchy of Knowledge (DIKIW)

Level Meaning
Data Raw, objective facts
Informati
Processed data with context and structure
on
Knowled Patterns and understanding derived from
ge information
Insight Interconnections that spark actionable ideas
Wisdom Experience-based judgment (tacit)
 Subjectivity Alert: While data is neutral, information includes
interpretation bias.
 Feedback Loop: We analyze data → produce information → uncover
gaps → seek new/better data.

5. Types of Knowledge
Type Characteristics Examples
Explici
Codified, clear rules Laws, policies
t
Implici Intuitive but learned Class patterns, best practices
t
Experiential, hard to Cultural norms, leadership
Tacit
articulate intuition

6. Decoding Analytics Buzzwords

Major Terms:
 Data Analytics: Broad term for analyzing raw data to extract insights
(business or otherwise).
 Business Analytics (BA): Uses data analysis to inform business
decisions.
 Business Intelligence (BI): Mostly retrospective, focused on what
happened.
 Data Science: More technical, often experimental.
 Artificial Intelligence (AI): Aim to replicate intelligent behavior.
 Machine Learning (ML): Subfield of AI – trains models to learn
patterns.
Analytics by Time Horizon:
Past Present Future
BI, Real-time BA, Predictive,
Descriptive analytics ML/AI

7. Four Types of Analytics

Type Purpose Example
Descriptiv
What happened? Monthly sales report
e
Diagnosti Why did it
Root cause of sales drop
c happen?
What might
Predictive Demand forecast
happen?
Prescriptiv What should we Automated price
e do? adjustments

8. Building Models for Better Decisions

 What is a model? A simplified representation of a system (e.g.,
revenue projection, supply chain).
 7 Modeling Steps:
1. Define the problem
2. Gather data
3. Build model (Excel, Python, etc.)
4. Test model
5. Refine model
6. Interpret results
7. Implement insights
 Warning: Garbage In = Garbage Out. Data quality drives model
reliability.
9. Excel Tools Used
Goal Seek
 Use when you have one desired output and one input to manipulate
(e.g., break-even point).
Data Tables
 One-way: Vary one input across a range (e.g., sales).
 Two-way: Vary two inputs (e.g., store size and rent) to analyze impact
on an output like profit.
Scenario Manager
 Create labeled cases like Best, Worst, Most Likely.
 Compares results of different input combinations.

10. Williams Coffee Case Applications

 Key Questions:
o What’s the break-even sales level?
o How to achieve $300K annual profit?
o How do store size and rent affect profit?
o How do best/worst-case scenarios change decision?
 Tools Applied:
o Goal Seek for break-even and profit targets.
o One/Two-Way Data Tables to test sales vs. expense/profit.
o Scenario Manager for sensitivity planning.

11. Data Quality Dimensions

Dimension Meaning
Reflects real-world
Accuracy
conditions
Completene All necessary data is
ss included
Consistency Uniform across sources
Uniqueness No duplicates
Recency Is it up to date?
Relevance Fit for purpose
Reliability Can it be trusted?
Presentation Understandable layout
Can users access it
Accessibility
fairly/safely?
 Data Audit: Review datasets for issues.
 Data Cleaning: Fix inconsistencies, fill gaps.

12. Big Data Concepts

 Big Data ≠ Just Large Data. It's data requiring specialized tools.
 5 V’s of Big Data:
1. Volume – Scale
 2. Velocity – Speed of generation
3. Variety – Data types (text, image, etc.)
4. Veracity – Trustworthiness
5. Value – Usefulness for decisions
 Tool Example: Hadoop (open source) – breaks large tasks into parallel
computing jobs.

13. Open Source vs. Proprietary Software

 Proprietary: Owned/licensed (e.g., Microsoft, Google).
 Open Source: Community-developed, transparent code (e.g., Linux,
VLC).
Pros:
 Free, modifiable, promotes innovation.
Cons:
 Risk of abandonment, unclear support, licensing complications.
 Real-World Issue: Debate over open-source AI (e.g., Meta’s LLaMA vs.
OpenAI’s GPT).
Goal Seek Refresher
 Use case: Adjust one input to achieve one desired output.
 Requires a formula in the "Set Cell".
 Examples: Break-even analysis, setting target profit.
 Common mistake: Using incorrect assumptions (e.g., not amortizing
franchise fee correctly).
 Garbage In = Garbage Out: Outputs depend on quality and logic of
input formulas.

One-Way Data Tables

 Used for: Varying one input across multiple possible values to observe
effects on multiple outputs.
 Example Variable: Selling price per coffee (e.g., $0.90 to $1.50).
 Outputs observed: Demand, Revenue, and Profit.
 Connected to Excel formulas (not static values) for dynamic
calculation.
 Used Excel features like:
o Autofill series
o Conditional Formatting for visual analysis
 Important logic:
o Profit = Revenue – Expenses
o Demand drops as price increases unless artificially fixed
(simulating non-price-sensitive behavior).

Two-Way Data Tables

 Used for: Testing the interaction between two input variables (e.g.,
unit cost and selling price) on one output (e.g., profit).
 Example setup:
o Rows = Unit Cost (e.g., $0.20 to $0.40)
 o Columns = Selling Price (e.g., $0.90 to $1.50)
o Output = Profit
 Limitation: Excel only supports two-way tables due to its 2D
spreadsheet structure.
 Visualization via Conditional Formatting:
o Most profitable outcomes in bottom right (low cost, high price).
o If human behavior is fixed (no price sensitivity), max profit
always increases with price.

Williams Coffee Pub Case Application

 Tasks:
o One-way table: Vary annual sales from $600,000 to $1.7M to
observe profit and expenses.
o Two-way table: Vary store size (sq. ft.) and rent to observe
profit.
 Lessons:
o Profit is sensitive to store size and rent cost.
o Must incorporate assumptions (e.g., bigger stores = more sales).
 Example improvement: Tie sales to square footage using formula like
Sales = base + rate × sq.ft.

Conceptual & Strategic Takeaways

 Build models that reflect realistic human/business behavior.
 Always question:
o Are my formulas grounded in logic or assumptions?
o What is the actual business decision being supported?

Residency DAY 2 - May 11, 2025

Part 1: Excel Analytics Recap & Completion
1. Review of Tools Learned So Far
Input
Tool Outputs Use Case
s
Goal Seek 1 1 Break-even or target value solve
One-Way Data
1 Multiple Sales impact on multiple outputs
Table
Two-Way Data Dual input impact (e.g., rent and size
2 1
Table on profit)
Scenario Many (or Best/Worst Case Sensitivity
Many
Manager 1+) Analysis
2. Scenario Manager Deep Dive
 Used for complex forecasting (5–10+ year financial projections).
  Can model combinations like:
o Sales growth
o Drive-thru revenue
o Operating expenses
 Enables structured planning for uncertain outcomes by allowing
predefined sets of assumptions.
3. Williams Coffee Case Application
 Inputs: Growth %, Drive-thru $, Expenses %
 Scenarios Created: Worst, Likely, Best
 Outputs Analyzed: After-tax profits, cumulative profit, and Net
Present Value (NPV)
 Conclusion: The business is viable but not highly profitable; risk of
low ROI in worst case; further refinement or alternative investment
options advised.

Part 2: Transition to AI Foundations

1. Purpose of Introducing AI
 To "cheat" Excel models using AI tools like CoPilot, ChatGPT, Gemini.
 Highlight where AI adds value or falls short in business analysis.
2. Definition Hierarchy
Concept Description
AI (Artificial The umbrella term: “machines that mimic human
Intelligence) intelligence.”
Machine Learning Subset of AI focused on learning from data.
Representation
Converts real-world concepts into numerical features.
Learning
Multi-layer neural networks for complex pattern
Deep Learning
recognition.
AI that creates (e.g., ChatGPT, DALL·E, Stable
Generative AI
Diffusion).
3. Historical Framing
 1956 Dartmouth Conference coined "AI"
 Early tools like Enigma led to AI's military and strategic importance
 Geoffrey Hinton (UofT) pioneered Deep Learning

Neuroscience & Statistical Analogy

 Neurons & Networks: Deep learning models mimic brain-like
pathways using layers and weights
 Example: Classifying cats/dogs/snakes based on features like “number
of legs” and “face shape”
 Prediction & Regression: Introduced with housing price model—uses
historical data to forecast unknown outcomes

Key Takeaways
1. Excel Tools teach the logic of modeling & sensitivity.
 2. Scenario Manager is powerful for multi-year business forecasting.
3. AI models can support, but not replace, human intuition—useful in
augmenting decision-making but limited by:
o Data quality
o Feature selection
o Model overfitting or hallucination
4. “All models are wrong, some are useful.” Always question
assumptions, test multiple scenarios.

Action Items for Students

 Download the Williams Coffee “Sunday” Excel Sheet from Avenue
(pre-built tables).
 Create and run 3 scenarios (worst, likely, best) with inputs: sales
growth, drive-thru revenue, expense %.
 Use Scenario Manager’s summary function to compare NPVs and
profits across scenarios.
 Begin reflecting on AI integration in business decision-making.
Final Project Overview
 The group final project will start around Module 2 and focus on
strategy and transformation.
 It will integrate topics like:
o Human-centered design
o Change management
o Analytics (including financial or numerical calculations)
 You'll pitch a transformative project (as if to a boss or investor).
 Flexibility in format/content—depends on your idea.
 Deeper work begins during Residency 2 (e.g., stakeholder analysis).

Machine Learning (ML) vs. Deep Learning (DL)

 ML: Requires humans to specify features (e.g., count wheels to
recognize cars).
 DL: Learns features on its own from raw data; more powerful but
less explainable.
o Often selects non-intuitive features (e.g., lighting edges).
o Hard to interpret—introduces risk due to lack of transparency.

Supervised vs. Unsupervised Learning

 Supervised: Labeled data (e.g., “cat” vs. “dog”), allows classification.
 Unsupervised: Unlabeled data; performs clustering to find hidden
patterns.
o Advantage: Less human bias, may discover new groupings.
o Needs more data typically.

Narrow AI vs. General AI (AGI)

 Narrow AI (e.g., chess bots, self-driving cars): Performs specific
tasks.
  General AI: Matches human-level intelligence across tasks.
o Raises ethical/philosophical concerns (alignment, autonomy,
existential risk).
 Professor identifies as AGI agnostic—skeptical of claims of its near-
term arrival.

Gartner Hype Cycle & AI Winters

 Generative AI (GenAI) is near the "Trough of Disillusionment".
o Overhyped tools may lead to another AI Winter (like in the
'70s/'90s).
 Many GenAI apps are useless or redundant, but some are showing
real value in niche use cases.

Prompt Engineering
 Historically important but becoming less relevant as models auto-
refine prompts.
 Good practices:
o Be specific, use examples, positive phrasing, and clear
objectives.
o Use frameworks like CRISPY (Capacity, Role, Insight, Statement,
Personality, Your style).

Practical Demos (Case Analysis)

 Tried using CoPilot vs. ChatGPT to solve Excel-based business cases.
o CoPilot failed; ChatGPT generated useful—but not always correct
—answers.
o Emphasized the importance of human verification and risk
management.
o Caution: Outputs may "seem" right but require understanding
and cross-checking.

Transformer Models
 Power modern LLMs like GPT and Gemini.
 Use autoregression: predict next word based on previous words.
o Not rule-based (“if X, then Y”)—instead probabilistic predictions
based on large datasets.
o Computationally expensive and potentially environmentally
impactful.

Visual AI & Diffusion Models

 Diffusion models (e.g., DALL·E, Midjourney) generate images by
reversing noise added to real images.
 Key principle: train by turning images to noise → learn reverse process.
 Enables creation of new images, hybrids (e.g., dog-cat), and
hallucinations when data is lacking.
Ethical Considerations (Teaser for Later)
 Risks in GenAI: hallucinations, privacy, lack of explainability, model
bias, jailbreaking.
 Will be covered deeper during the ethics module.