#15. What Can AI Do? A Functional Classification of AI

In previous articles, we looked at AI by the type of data it handles, such as text AI, image AI, voice AI, and video AI.

This article uses a different standard.

Even when two AI systems use the same image data, they may perform very different tasks. One may detect objects in a photo, another may predict how those objects will move, and another may generate an entirely new image.

The data may be the same, but the function is different.

To understand AI more clearly, we need to separate two questions.

1. What type of data does the AI use?
2. What does the AI do with that data?

There is no single universal way to classify every AI function. In this article, we will organize AI according to the process through which it receives information, produces a result, and connects that result to an action.

Recognition and Classification → Prediction → Recommendation and Selection → Generation and Transformation → Planning and Optimization → Execution and Control

These six functions are not completely separate technologies. They are categories that help us understand the role an AI system performs.

In real AI services, several functions are often connected or used at the same time.

How Are AI Functions Connected?

The process through which AI analyzes a situation and turns that analysis into action can be understood as follows.

1. Recognition and Classification: Identify what is currently present.
2. Prediction: Calculate what is likely to happen next.
3. Recommendation and Selection: Choose a suitable option from several candidates.
4. Generation and Transformation: Create a new output or change an existing one.
5. Planning and Optimization: Determine the steps and methods needed to achieve a goal.
6. Execution and Control: Carry out the decision through a system or machine.

Not every AI system performs all six steps.

A spam filter may stop after recognition and classification. An image generation service mainly focuses on generation and transformation.

A self-driving vehicle or robot, however, may use several functions in sequence because it must understand its surroundings and act within them.

1. AI That Recognizes and Classifies

Recognition and classification AI identifies useful features within input data and determines what the data represents.

This includes detecting a person in a photo, deciding whether an email is spam, or identifying which words were spoken in an audio recording.

Common examples include:

1. Detecting people or objects in an image
2. Verifying a face or fingerprint
3. Converting speech into text
4. Separating defective products from normal products
5. Classifying emails as spam or legitimate messages
6. Detecting unusual signs in medical images

Recognition and classification are closely related, but they do not perform exactly the same role.

Recognition is mainly about locating or identifying a target within data. Classification is about assigning that target to one of several predefined categories.

For example, when a factory camera detects a scratch on the surface of a product, it is performing recognition. When the system labels that product as normal or defective, it is performing classification.

This does not necessarily mean that AI understands the object in the same way a person does.

It calculates which object or category the input most closely resembles based on patterns found in its training data.

2. AI That Predicts Future Outcomes

Prediction AI uses past and current data to calculate which outcome is most likely to occur next.

It is used to estimate events that have not yet happened, such as future sales, demand, equipment failures, or traffic conditions.

Common examples include:

1. Predicting next month’s product sales
2. Calculating the likelihood of equipment failure
3. Predicting whether a customer will cancel a service
4. Estimating the risk that a borrower will not repay a loan
5. Predicting traffic conditions and arrival times
6. Calculating the likelihood of developing a disease

Prediction AI does not know the future with certainty.

It identifies relationships that repeatedly appeared in existing data and calculates what is likely to happen under similar conditions.

For that reason, predictions may become inaccurate when a new situation differs significantly from the past or when important information is missing from the data.

Recognition and classification determine what the current input is.

Prediction calculates what is likely to happen after that.

3. AI That Recommends and Selects

Recommendation and selection AI evaluates several candidates and presents the option it considers most suitable for a user or situation.

Product recommendations in online stores, video recommendations on streaming platforms, and personalized playlists in music apps are familiar examples.

Common examples include:

1. Recommending videos a user may enjoy
2. Displaying products a customer is likely to purchase
3. Suggesting music or movies that match a user’s preferences
4. Ranking news articles a user may find interesting
5. Recommending suitable job postings to applicants
6. Displaying search results that are most relevant to a question

Recommendation systems often use prediction as part of their process.

An AI system may analyze a user’s previous behavior and the choices made by similar users. It then predicts how likely the user is to click, watch, or purchase each item.

The service uses those predictions to rank the available options or select which ones to display.

The difference between prediction and recommendation lies mainly in their purpose.

Prediction calculates the likelihood of different outcomes. Recommendation and selection use those calculations to rank candidates or choose an appropriate option.

For example, calculating how likely a user is to watch a particular movie is prediction. Using that result to decide which movies appear first on the home screen is recommendation.

A recommended result is not always the option that is best for the user.

Depending on the service’s objective, the system may prioritize content that keeps users engaged longer, produces more advertising revenue, or is more likely to lead to a purchase.

4. AI That Generates and Transforms

Generation and transformation AI creates new outputs from an input or changes existing content into another form.

This has become one of the most familiar AI functions as generative AI has become widely used.

Common examples include:

1. Writing an answer to a question
2. Creating a draft article or report
3. Summarizing a long document
4. Translating text into another language
5. Generating an image from a written description
6. Creating speech or music
7. Writing computer code
8. Editing an existing image or video

Generation and transformation are often used together, but there is a difference between them.

Generation creates a new output based on the input. Transformation changes the language, format, length, or expression of content that already exists.

For example, writing a new article from a topic is generation. Translating an existing Korean article into English is transformation.

Turning a long meeting transcript into a short summary is also closer to transformation.

Generative AI produces results by using patterns and relationships learned from training data.

However, a natural-looking result is not always an accurate one. AI may produce incorrect explanations or describe information that does not exist as though it were real.

Important outputs therefore need to be checked separately.

5. AI That Plans and Optimizes

Planning and optimization AI examines a goal and its conditions, then determines the steps and methods needed to achieve that goal.

It compares several possible options and searches for a more suitable method based on factors such as time, cost, distance, safety, or efficiency.

Common examples include:

1. Finding the fastest travel route
2. Determining the delivery order for logistics vehicles
3. Adjusting a factory production schedule
4. Assigning tasks to employees or equipment
5. Optimizing electricity usage
6. Determining a sequence of actions in a game
7. Organizing a multi-step business process

Prediction asks, “What outcome is most likely to occur?”

Planning asks, “What needs to be done, and in what order, to achieve the goal?”

Optimization becomes important when several possible methods are available. It searches for the option that best satisfies a defined criterion or set of criteria.

For example, estimating when a delivery vehicle will arrive at each destination is prediction. Deciding the order in which the destinations should be visited is planning.

Finding the route that minimizes travel distance and time is optimization.

Agentic AI and Planning

AI systems that can receive a goal, divide it into smaller tasks, determine the order of those tasks, select tools, and continue working through the process have recently gained significant attention.

These systems are commonly described as agentic AI or AI agents.

A conventional generative AI system primarily focuses on producing an answer or output. Agentic AI organizes the work required to achieve a broader objective.

For example, suppose a user asks an AI system to prepare materials for a meeting next week. The AI may perform the following steps.

1. Determine which materials are needed.
2. Search for relevant information.
3. Organize the key points.
4. Create a document or presentation.
5. Review the result and revise missing sections.

This process does not rely on generation alone.

Planning is used to divide the goal into smaller tasks. Selection is used to choose suitable information and tools. Generation is used to create the document.

When the AI enters information into an external service or saves a file, execution is also involved.

Agentic AI is therefore better understood not as one completely new function, but as a way of connecting several AI functions around planning so that the system can carry out an entire goal-oriented process.

6. Execution and Control AI Turns Decisions into Actions

Execution and control AI connects an analysis or plan to an actual action performed by a system or machine.

Instead of ending with a result displayed on a screen, it directly changes something in a digital or physical environment.

Common examples include:

1. Controlling the direction and speed of a self-driving vehicle
2. Moving a robotic arm to pick up an object
3. Controlling the altitude and direction of a drone
4. Automatically adjusting a building’s temperature and lighting
5. Blocking a transaction after detecting suspicious activity
6. Sorting or sending emails according to defined conditions
7. Entering data into a business system and saving documents

The difference between planning and execution-and-control is whether the AI causes a real change in the environment.

Planning AI decides what should be done and in what order. Execution and control AI changes the state of a system or moves a machine according to that plan.

For example, determining the movement path and action sequence of a robot is planning. Rotating the motors, moving the robotic arm, and picking up the object are execution and control.

The same distinction applies in software.

Writing a draft email is generation. Deciding which email should be answered first involves planning or selection.

Sending the email through an authorized account is execution.

Errors in execution and control can have a greater impact than errors in some other AI functions.

An incorrect sentence and an incorrect movement by a vehicle, medical device, or industrial robot do not create the same level of risk.

Execution and control AI therefore requires more than accurate results. It also needs permission management, limits on what the system can do, emergency stop mechanisms, and human review.

Why Do These Functions Overlap?

The six functions are not completely independent technologies.

Their boundaries may appear to overlap because an AI system often uses several functions in sequence to produce one result.

A product recommendation service, for example, may recognize a user’s behavior, predict the likelihood of purchasing each product, and then rank the products for recommendation.

1. Identify the user’s search and purchase history.
2. Predict the likelihood that the user will select each product.
3. Rank the products according to the service’s criteria.
4. Display the highest-ranked products.

A self-driving vehicle combines even more functions.

1. Recognition and Classification: Detect roads, traffic lights, vehicles, and pedestrians.
2. Prediction: Estimate how nearby vehicles and pedestrians will move.
3. Planning and Optimization: Determine a safe and efficient driving route.
4. Execution and Control: Accelerate, brake, and steer the vehicle.

An AI agent also connects several functions.

1. Understand the user’s request and the current situation.
2. Estimate the tasks and results that may be needed.
3. Select appropriate information and tools.
4. Generate documents or responses.
5. Plan the order and method of the work.
6. Operate programs or services within its authorized permissions.

For this reason, describing an AI system with only one functional label may not fully explain how it works.

Calling a system “generative AI” may simply mean that generation is the most visible part of the service.

Recognition, selection, prediction, and planning may also be operating behind the scenes.

Data Type and Function Are Different Classification Standards

One of the most common sources of confusion is mixing data-based categories with function-based categories.

Image AI refers to AI that handles image data. Prediction AI refers to AI that calculates possible future outcomes.

These terms are based on different classification standards.

The same AI system can therefore be described according to both the data it uses and the functions it performs.

Text, images, speech, video, and sensor data describe the forms of data that enter an AI system.

Recognition and classification, prediction, recommendation and selection, generation and transformation, planning and optimization, and execution and control describe the roles the AI performs with that data.

What Should We Check When Evaluating an AI Service?

The label “AI” alone does not explain what a service actually does.

When examining an AI service, it is useful to ask the following questions.

1. What type of data does it receive?
2. What does it identify within that data?
3. Does it classify the current situation or predict a future outcome?
4. Does it rank several options or select one?
5. Does it create a new output or transform existing content?
6. Does it determine the steps needed to achieve a goal?
7. Does it stop after presenting a result?
8. Does it directly operate a system or machine?

These questions help us understand not only the role of the AI, but also what may happen when it makes an error.

An AI system that generates sentences and an AI system that controls a vehicle cannot be governed by the same safety standards.

An AI system that recommends content and one that automatically blocks payments also have different levels of responsibility and impact.

To understand an AI system properly, we need to look beyond its name and examine what data it receives, what function it performs, and how far its results are allowed to go.

DANA NOTES in One Sentence

AI functions can be divided into recognition and classification, prediction, recommendation and selection, generation and transformation, planning and optimization, and execution and control. In practice, most AI services work by connecting several of these functions together.