AI Product Management:
Must-Know Terms (With
Real-Life Examples)
Jesal Shah
1. Machine Learning (ML)
Machine Learning is how we teach computers to learn from data like how humans
learn from experience.
Example: Netflix watches what types of shows you watch (like comedies) and uses
that information to suggest similar shows without anyone manually setting rules.
2. Supervised Learning
This is when we teach a model using labeled data, like flashcards with correct
answers.
Example: An email system is shown thousands of emails labeled as "spam" or "not
spam" so it can learn to sort new emails correctly.
3. Unsupervised Learning
Here, the model learns by itself without any labeled examples, like organizing things
based on similarities.
Example: A shopping website groups users into categories based on their buying
habits, even if no one tells the AI which group is which.
4. Reinforcement Learning
The model learns by trial and error, just like a child learning to ride a bicycle.
Example: A robot in a warehouse learns the best path to move boxes by trying
different routes and getting rewarded for being fast and accurate.
5. Natural Language Processing (NLP)
This helps computers understand and work with human language.
Example: When you ask ChatGPT a question like “Explain blockchain in simple
words,” it reads your text, understands the intent, and replies in clear language, just
like a human conversation.
6. Model Training
Model training is the process where we teach the AI to recognize patterns using past
data.
Example: We show thousands of photos of cats and dogs to a model so it learns to
recognize the difference.
7. Model Inference
Once trained, inference is when the AI makes predictions based on new data.
Example: After training, the model can look at a new photo and say whether it is a cat
or a dog.
8. Overfitting
Overfitting happens when the model learns too much from the training data and fails
on new data.
Example: A school student memorizes all past exam questions but fails the final test
because the questions are slightly different.
9. Feature Engineering
This means choosing or creating useful data points to help the AI make better
decisions.
Example: Instead of just knowing when a user last logged in, we calculate “how many
days ago” they logged in to predict if they will stop using the app.
10. Data Pipeline
It is the system that moves raw data through steps like cleaning, organizing, and
preparing for use in AI.
Example: Collecting all user activity on an app, fixing any errors, and storing it neatly
so it can be used for analysis.
11. A/B Testing
This is when you try two versions of something to see which one performs better.
Example: Half the users see Button A (blue), and the other half see Button B (green).
The company measures which button gets more clicks.
12. Precision & Recall
These are ways to measure how well a model makes decisions. Precision is how
many of its answers were right. Recall is how many correct answers it found.
Example: A fraud detection tool catches 100 frauds. If 90 are truly frauds (precision),
but there were 150 frauds in total and it caught only 100, its recall is lower.
13. ROC-AUC Score
ROC – Receiver Operating Characteristic
AUC – Area Under the Curve
Together, the ROC-AUC score helps you understand how well a classification model
separates the classes (like fraud vs not fraud).
Example: Imagine the model is trying to distinguish between bad users and good
users. ROC-AUC tells you how good it is at doing that across all possible thresholds.
A score closer to 1.0 means the model is excellent. A score of 0.5 means it is
guessing randomly.
14. Embeddings
These are smart ways of turning complex data like words into numbers that AI can
understand.
Example: When you search for “laptop” on an e-commerce site, the AI also shows
results for “notebook,” “MacBook,” or “ultrabook.” This is because all these words are
placed close together in the embedding space, meaning the AI understands they are
similar in meaning or use.
15. Model Drift
Over time, the data in the real world changes. When your model stops performing
well, that is called drift.
Example: A model trained before COVID to predict travel bookings may not work well
after travel patterns changed drastically.
16. Explainability (XAI)
It means helping humans understand why an AI made a certain decision.
Example: If a loan application is rejected, the AI explains it was due to low credit
score and high current debt.
17. RAG (Retrieval-Augmented Generation)
This is when AI searches for information first, then uses that to answer your question.
Example: You ask an AI assistant a question about the latest finance law. It first
fetches the law from the web, then answers you accurately.
18. Prompt Engineering
This is the art of writing clear and smart instructions for AI to get good answers.
Example: Instead of saying “summarize this,” you say “summarize this news article in
2 sentences for a beginner.”
19. Latency
It means how fast the AI gives you a result. Lower latency = faster system.
Example: A voice assistant that replies instantly feels smooth. If it takes 10 seconds to
reply, users will get annoyed.
20. Human-in-the-Loop (HITL)
Sometimes AI alone is not enough, humans need to review or approve decisions.
Example: An AI flags a social media post as harmful. A human moderator checks and
confirms before removing it.
