The most common AI use cases

Knowing what AI is only gets you halfway. The other half is knowing what to use it for, and just as importantly, when not to use it. Exam questions and real projects both start from a business problem, not from a technology, so the skill is matching the problem to the right kind of solution. This post surveys the most common AI use cases, links each to the AWS services you are likely to meet, and then looks at the situations where AI is the wrong choice.

1. Start with the problem, not the technology

A useful habit is to describe a problem in plain words before naming any technology: “we need to sort incoming support emails by topic”, “we want to read data from scanned invoices”, “we need to predict which customers will cancel”. Once the problem is clear, you can ask which category of AI fits it. The categories below are the ones you will see most often.

2. Common use cases by category

Category	What the AI does	Example	AWS service you may meet
Computer vision	Analyses images and video	Detecting objects, faces or defects on a production line	Amazon Rekognition
Document processing	Extracts text and data from documents	Reading fields from scanned invoices and forms	Amazon Textract
Natural language processing	Understands text: sentiment, entities, topics, translation	Sorting customer reviews by sentiment	Amazon Comprehend, Amazon Translate
Speech	Converts speech to text and text to speech	Transcribing calls, reading content aloud	Amazon Transcribe, Amazon Polly
Recommendation and personalisation	Suggests items based on behaviour and preferences	“Customers also bought…”	Amazon Personalize
Prediction and anomaly detection	Predicts values or flags unusual events from historical data	Customer churn, demand forecasts, fraud flags	Amazon SageMaker (check current AWS offerings for specialised forecasting and fraud services)
Conversational interfaces	Handles dialogue with users by text or voice	A support chatbot, a voice assistant	Amazon Lex, Amazon Bedrock
Generative AI	Creates new text, images or code, and answers questions over your own data	Summarising reports, drafting emails, generating code, company Q&A	Amazon Bedrock, Amazon Q

The same use cases across industries

• Healthcare: reading medical images, extracting data from clinical documents, summarising notes.
• Finance: flagging unusual transactions, extracting data from statements, assisting customer service.
• Retail: recommendations, demand forecasting, product descriptions.
• Manufacturing: spotting defects in images, predicting equipment failures.
• Customer service: chatbots, call transcription, routing and summarising enquiries.

3. Predictive vs generative use cases

Most of the rows above are predictive or analytical: the system labels, scores, extracts or forecasts. Only the last row is generative, where the system produces new content. This matters because the two families have different strengths and risks. Predictive models are usually measured against a clear right answer, such as accuracy on labelled data. Generative models can produce fluent text that is wrong, a problem known as hallucination (Ji et al., 2023), so they need extra checks such as grounding in your own data and human review. A common mistake is reaching for a generative model when a simple classifier or a pre-built service would be cheaper and more reliable.

4. When AI is the wrong tool

AI is powerful, but it is not free, and it is not always the best answer. These are the main situations where a different approach is usually better:

• A simple rule or formula already solves it. Calculating sales tax, applying a discount code or validating a postcode needs logic, not learning.
• You need an exact, guaranteed outcome rather than a prediction. AI models produce probabilistic answers. Where the result must be right every time, such as a financial calculation or a legal threshold, deterministic code is safer. The AWS AI Practitioner exam guide includes recognising cases like this (Amazon Web Services, n.d.), so check its current wording.
• There is too little data, or the data is poor. Machine learning learns from data. Small, biased or messy datasets produce unreliable models.
• The cost outweighs the benefit. Building, running and maintaining an AI system has real costs, and Sculley et al. (2015) describe how ML systems accumulate maintenance burdens that traditional code does not. A cost-benefit comparison should come first.
• The decision is high-stakes and needs explanation or accountability. In areas such as lending, hiring or healthcare, people need to understand and challenge decisions. Risk frameworks such as the NIST AI Risk Management Framework (National Institute of Standards and Technology, 2023) stress governance and human oversight in these cases.
• You cannot define what “good” looks like. Without a measurable success criterion you cannot tell whether the AI is working.

A quick decision checklist

Ask yourself	If the answer points this way…
Can a fixed rule or formula do the job?	Yes → use ordinary code
Must the result be exact every time?	Yes → use deterministic logic, with AI at most as a helper
Do you have enough good-quality data?	No → collect data first, or use a pre-trained model
Is the benefit larger than the build and running cost?	No → do not build it
Is the decision high-stakes?	Yes → add explainability and human review

5. If AI fits: choosing how to build

When AI is the right tool, there is a common order of preference, from least to most effort. The details come in Part E, but the pattern is worth knowing now.

1. A pre-built AI service if one already does your task, such as Textract for reading documents or Transcribe for speech. This is the fastest and often the cheapest route.
2. A foundation model through Amazon Bedrock if the task involves language, summarising, answering questions or generating content, adapted with prompting or your own data.
3. A custom model on Amazon SageMaker if you have specialised data and needs that the first two cannot meet.

Exam scenarios often reward choosing the simplest option that meets the requirement, not the most sophisticated one.

6. Common confusions and exam traps

• Generative AI is not the answer to everything. Forecasting, fraud detection and recommendations are predictive tasks.
• Custom is not automatically better. If a managed service fits, building your own model usually adds cost and effort without extra value.
• “Predict” and “guarantee” are different words. If a scenario needs a guaranteed, specific outcome, AI may be the wrong tool.
• Match the service to the data type. Images point to Rekognition, documents to Textract, speech to Transcribe or Polly, and text analysis to Comprehend.
• Watch for cost and complexity clues. Phrases such as “least operational overhead” or “most cost-effective” usually favour managed services.

7. Why this matters for your AWS exam

The AI Practitioner exam asks you to identify suitable use cases for AI and to decide where it is and is not appropriate (Amazon Web Services, n.d.). Check the current exam guide for the exact task statements and weightings. The Developer exam builds on this with design choices, such as when to use a managed service, a foundation model or a custom model. Practise reading a scenario, naming the data type and the output required, and then picking the simplest tool that fits.

8. Quick check

1. Which AWS service would you consider first for extracting fields from scanned invoices?
2. Name two situations where AI is probably the wrong tool.
3. Is predicting customer churn a predictive or a generative use case?
4. Why might a generative model be risky for a task that needs exact, verifiable answers?
5. When a pre-built AI service fits the task, why is it usually preferred over a custom model?

Answers: (1) Amazon Textract. (2) Any two of: a simple rule solves the problem, an exact guaranteed outcome is required, too little or poor data, cost outweighs benefit, high-stakes decisions without oversight, no clear success measure. (3) Predictive. (4) It can produce fluent but incorrect statements (hallucinations), so the output cannot be assumed correct without checks. (5) It is faster to adopt and usually cheaper and simpler to operate, with no need to collect data and train a model yourself.

What’s next

In the next post we begin the core types of machine learning, starting with supervised learning: what it is, how it learns from labelled examples, and where you will meet it.

References

Amazon Web Services. (n.d.). AWS Certified AI Practitioner (AIF-C01) exam guide. Retrieved [Month Day, Year], from [URL of the current official exam guide]

Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), Article 248. https://doi.org/10.1145/3571730

National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). U.S. Department of Commerce. https://doi.org/10.6028/NIST.AI.100-1

Sculley, D., Holt, G., Golovin, D., Davydov, E., Phillips, T., Ebner, D., Chaudhary, V., Young, M., Crespo, J.-F., & Dennison, D. (2015). Hidden technical debt in machine learning systems. Advances in Neural Information Processing Systems, 28, 2503–2511.