ハーバードビジネススクールオンラインが教えるデータサイエンスと AI の専門家に必要な知識

# Introduction

**

You do not need to become a data scientist to benefit from data science and artificial intelligence (AI). However, you should understand what these technologies can do, where they can fail, and how to evaluate their outputs.

After working with data science and AI tools for several years, I have noticed that people often begin with the technology rather than the problem. They want to train a model, introduce a chatbot, or build an AI application before deciding what decision they are trying to improve.

For most professionals, the goal should not be to master every algorithm. It should be to develop enough data and AI literacy to ask better questions, challenge unreliable results, understand the limitations of these systems, and make informed decisions.

This article brings together insights from my interview with Iavor I. Bojinov**, Associate Professor of Business Administration at Harvard Business School, and my own experience working with data science and AI tools. Bojinov co-leads the HBS Online courses AI for Leaders and Data Science and AI for Decision Making.

# Starting With the Decision

**

Before using AI, I first define the decision that needs to be made.

"We want to use AI for customer analytics" is too broad. A clearer objective would be: "We want to identify customers who may leave so our retention team can contact them."

This connects the model to a measurable outcome and a real business action. A technically advanced system has little value if it does not improve the underlying decision.

Bojinov makes a similar point. AI should be treated as an input into decision-making rather than a substitute for judgment. Professionals need to understand its capabilities and limitations, validate its recommendations, and focus on better business outcomes instead of adopting AI for its own sake.

In his HBS Online interview, AI Adoption, Trust, and Decision-Making, Bojinov also explains that successful AI adoption depends on trust and how effectively organizations introduce and scale these systems.

Before starting an AI project, ask what problem it will solve, who will use the output, what action will follow, and how success will be measured.

# Understanding the Data First

When I begin a data science project, I do not immediately train a model. I start by exploring the data.

Exploratory data analysis (EDA) helps us understand a dataset's structure, quality, distributions, and relationships. It may uncover missing values, duplicate records, inconsistent categories, unusual observations, and potential biases.

These checks matter because accuracy alone can be misleading.

Suppose only five percent of customers leave. A model that predicts every customer will stay would be 95 percent accurate, yet it would identify none of the customers at risk.

Professionals should therefore ask whether a metric reflects the real business objective rather than relying on one impressive-looking score.

HBS Online's latest article, Data Preparation for Your AI Model: Importance and Best Practices**, also emphasizes that a model's performance depends heavily on the quality of the data behind it. The article covers cleaning, enriching, transforming, and organizing data before model development.

# Preparing the Data Carefully

**

Real-world data is rarely ready for machine learning.

Data preparation may involve handling missing values, correcting formats, removing duplicates, transforming categories, selecting useful features, and separating training data from evaluation data.

Professionals should also know where their data came from, how it was collected, what may be missing, and whether previous decisions introduced bias.

In my experience, improving the data often delivers a greater benefit than replacing a simple algorithm with a more complicated model.

A model learns from the information it receives. If that information is incomplete, outdated, inconsistent, or biased, the results will also be unreliable.

Better technology cannot fully compensate for poor data.

# Choosing the Simplest Model That Works

The best model is not always the newest, largest, or most expensive one.

For structured business data — such as customer records, transactions, sales history, or operational data — a regression model or decision tree may be more appropriate than a general-purpose large language model (LLM).

It can also be faster, cheaper, easier to explain, and simpler to monitor.

LLMs are useful when the task involves language, such as summarizing documents, extracting information, generating drafts, or analyzing customer feedback. For forecasting, classification, and tabular data, traditional machine learning may remain the better option.

Before choosing a model, I compare predictive performance, training and inference costs, speed, scalability, explainability, maintenance requirements, and the cost of incorrect predictions.

A slightly more accurate model is not always the better choice if it is much more expensive, difficult to explain, or hard to maintain.

The goal is not to use the most advanced AI. It is to choose the most reliable and cost-effective solution for the problem.

# Validating Before You Trust

A model performing well during development does not guarantee that it will work in the real world.

It should be evaluated using data that was not used to train or select it. The evaluation data should also represent the customers, markets, and conditions in which the model will operate.

Validation should continue after deployment. Customer behavior changes, markets shift, data pipelines break, and patterns that were once useful may become less reliable.

Bojinov identifies weak validation and testing as a common mistake. When organizations deploy models without testing them properly, they may become overconfident in the results.

Professionals should ask how a model was evaluated, which mistakes it makes, whether its test data is representative, and how its performance will be monitored over time.

Trust should come from evidence, not from how advanced or confident the system appears.

# Turning Customer Analytics Into Action

Customer analytics should be more than a dashboard.

A dashboard can show what happened. A predictive model can estimate what might happen next. An experiment can help determine what an organization should do about it.

Predicting that a customer may leave only becomes valuable when the company has a realistic retention strategy. It must then measure whether the intervention actually works.

Professionals should also avoid confusing correlation with causation. A feature associated with customer churn is not necessarily the reason customers are leaving.

This is where experimental thinking becomes valuable. Rather than assuming an intervention will work, organizations should test it, measure the outcome, and learn from the evidence.

The objective is not to produce more predictions. It is to make better decisions.

# Treating LLMs as Assistants

Large language models (LLMs) have made AI much more accessible. I use them regularly for coding, research, improving grammar, summarizing information, generating ideas, and learning something new.

They help me work faster and explore unfamiliar topics. However, fluent and confident language is not the same as factual accuracy.

LLMs can misunderstand context, rely on outdated information, invent sources, produce incorrect code, or overlook security vulnerabilities. HBS Online's guide, What Are Large Language Models and How Do They Work?**, provides an overview of what business professionals should understand before using them.

I never trust their output blindly. I review and reread what they produce, verify important claims, inspect generated code, test whether it works, and check for security issues.

I treat an LLM as an assistant rather than an unquestionable authority. The greater the consequences of a task, the more carefully its output should be reviewed.

# Keeping Human Judgment in the Loop

**

According to Bojinov, data literacy, experimental thinking, and the ability to critically evaluate AI outputs will become increasingly valuable as these tools become more widely used.

Professionals do not need to become AI experts, but they do need enough fluency to ask the right questions, interpret evidence, recognize weak or unreliable outputs, and understand when an answer can be trusted.

HBS Online's *Parlor Room* episode, How Mid-Career Professionals Can Lead and Grow With AI, explores how professionals can combine AI fluency with their existing domain expertise to create value in an AI-driven workplace. Another episode, Early-Career Advice for Building AI and Human Skills, discusses the importance of developing strong expertise while learning to work effectively alongside AI.

Ultimately, the professionals who stand out will be those who can combine technical understanding with domain knowledge and sound business judgment.

# Final Thoughts

Many companies are rushing to add AI everywhere because they do not want to fall behind. In doing so, they may spend heavily on APIs, infrastructure, training, and maintenance without first asking whether AI is actually the right solution.

I think the hype will eventually settle.

Companies will learn where AI adds real value, where people remain essential, and where a simpler and cheaper solution is enough.

Successful data science and AI projects do not begin with the latest model. They begin with a clear problem, reliable data, appropriate evaluation, realistic costs, and a plan for turning results into action.

The goal is not to use AI everywhere. It is to use it deliberately, validate it carefully, and apply it where it genuinely improves decisions.

Abid Ali Awan** (@1abidaliawan) is a certified data scientist professional who loves building machine learning models. Currently, he is focusing on content creation and writing technical blogs on machine learning and data science technologies. Abid holds a Master's degree in technology management and a bachelor's degree in telecommunication engineering. His vision is to build an AI product using a graph neural network for students struggling with mental illness.