Fundamentals of AI and ML · Topic 4 of 8
When to Use AI and ML
Decide when AI and ML add value and when simple rules are better, and match business problems to the right ML technique and common real world use cases.
- 4 min
- Use AI and ML for large scale, pattern based, hard to define tasks, not for exact rule based calculations.
- AI is not suited when data is missing, results must be fully guaranteed, or the cost exceeds the benefit.
- A category output means classification, a number output means regression, and unlabeled grouping means clustering.
What it is
AI is a tool, not a goal. The exam expects you to judge whether a business problem is a good fit for AI and ML, and which kind of technique suits it.
AI works well when the task depends on patterns in data that are hard to write as fixed rules. It works poorly when a clear rule or a simple calculation already solves the problem.
How it works
When AI and ML add value
AI is a good fit in these cases.
- Scale: The task is too large for people to do by hand, such as reviewing millions of transactions.
- Complexity: The rules are too many or too fuzzy to write down, such as recognising a face or a tone of voice.
- Prediction: You need to estimate a future value or outcome from past data.
- Personalisation: Each user needs a different answer, such as product suggestions.
- Language and media: You must process text, speech, images or video.
- Automation of judgement: The decision is repeated often and follows learnable patterns.
When AI and ML are not the right tool
Avoid AI in these cases.
- A deterministic rule gives the exact answer every time, such as calculating tax from a fixed table.
- There is not enough data, or the data is poor or unrepresentative.
- The cost of building and running the model is higher than the value it brings.
- The decision needs a guaranteed, fully explainable result that a probabilistic model cannot give.
- A wrong answer would be costly or dangerous and no human can review it.
A prediction from a model is a probability, not a certainty. If a use case cannot accept occasional errors, plain software or a human check is a better choice.
Matching the problem to the technique
| Business question | Technique | Example |
|---|---|---|
| Which category does this belong to | Classification | Spam filter, fraud flag, sentiment label |
| What number will this be | Regression | House price, delivery time |
| Which items are alike | Clustering | Customer segments |
| What is unusual here | Anomaly detection | Odd network traffic |
| What will come next in time | Forecasting on time series | Demand planning |
| What should we suggest | Recommendation | Products or films |
| What best strategy to follow | Reinforcement learning | Robot movement |
| Create new content | Generative AI | Drafts, summaries, images |
Common real world applications
| Area | Use case |
|---|---|
| Retail | Product recommendations, demand forecasting |
| Finance | Fraud detection, document processing |
| Customer service | Chatbots, call transcription, sentiment analysis |
| Healthcare | Reading medical documents and images, with expert review |
| Manufacturing | Predicting equipment faults, visual quality checks |
| Marketing | Customer segmentation, content drafting |
| Software | Code suggestions and explanations |
Questions to ask first
Before choosing AI, check four things: is there a clear business goal, is the data available and good, can the result be measured, and is a human review needed. A clear success measure, such as fewer support calls or higher conversion, shows if the project works.
What to remember for the exam
- Use AI and ML for large scale, pattern based, hard to define tasks, not for exact rule based calculations.
- AI is not suited when data is missing, results must be fully guaranteed, or the cost exceeds the benefit.
- A category output means classification, a number output means regression, and unlabeled grouping means clustering.
- Fraud detection is usually anomaly detection or classification.
- Recommendation systems suggest items based on user behaviour.
- Forecasting predicts future values from time series data.
- Document, speech, image and language tasks often have a ready made AWS AI service, so building from scratch is rarely the first choice.
- Define a measurable business outcome before starting.
If a question describes a simple fixed rule, such as a discount for orders over a set amount, the answer is not machine learning.
Common mistakes
Choosing AI because it sounds modern. A rule or a database query is cheaper, faster and fully predictable when it fits the problem.
Ignoring the data. A model cannot learn what the data does not show. No data usually means no ML.
Expecting perfect accuracy. Models make mistakes, so the design should handle errors, for example with a human in the loop.
Using generative AI for every problem. Classic ML is often a better fit for predicting numbers or categories, and it costs less to run.
Skipping the business measure. Without a clear goal, nobody can tell if the model is worth keeping.
Check yourself
Name a case where AI is a poor fit.
Calculating a total from a fixed price list is a poor fit, because a simple rule gives the exact answer every time.
Which technique suits grouping similar customers?
Clustering, which is unsupervised learning.
Why is a human review sometimes needed with AI output?
Models give probable answers, not guaranteed ones, so a person can catch costly errors.
Topic 4 of 8 in Fundamentals of AI and ML
Test yourself on Fundamentals of AI and ML
The full AWS AIF-C01 question bank has 849 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by AWS. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.