Rules, Machine Learning, Generative AI, and Agents
Compare overlapping system categories and identify what each part of a real product actually does.
One product can use many parts
It is easy to imagine separate boxes for rules, machine learning, generative AI, and agents.
Real products are not that simple.
A product may use rules and machine learning together. A generative model may use deep learning. An agent may use a generative model, search, rules, and other tools.
Let us look at each idea in simple terms.
Fixed rules
A fixed rule tells a system what to do when a clear condition is met.
For example:
If an invoice is above the approval limit, send it to a manager.
A person wrote this rule. The system did not learn it from examples.
Rules are useful when the requirement is clear and stable. They may also be used inside an AI product for checks, safety limits, or routing.
Some broad meanings of AI include rule-based systems. So avoid arguments based only on the word AI. A clearer description is:
This part follows fixed rules. There is no proof of a trained model.
Machine learning
Machine learning builds behaviour from data or examples.
Think about spam. It would be hard to write a rule for every possible spam message. Instead, a model can be trained with examples of spam and normal email. It then gives a score to a new message.
People still make many choices. They choose the goal, data, labels, tests, limits, and way the model is used.
Deep learning
Deep learning is one type of machine learning. It uses models called neural networks with many layers.
Deep learning is used in many text, image, audio, and prediction systems.
Remember:
- Deep learning is part of machine learning.
- Not all machine learning is deep learning.
- Deep learning does not always create content.
Generative AI
Generative AI creates content. It can create text, images, audio, video, or code.
The word generative tells us about the output. It does not prove that the product can search the web, remember old chats, check facts, or act on its own.
Retrieval
Retrieval means finding existing material in a collection, database, or search service.
If a photo library returns an existing photo of red flowers, that is retrieval. If a model creates a new picture of red flowers, that is generation.
A product may do both. It may find documents first and then create an answer from them.
Tools
A tool is an extra ability that a system can use.
Examples include:
- Search.
- A calculator.
- A database.
- A calendar.
- An email service.
- A code runner.
A model may try to write an answer to a maths problem. A product with a calculator tool may send the problem to the calculator and use its result.
Tool access can be useful. It also creates questions:
- What can the tool read?
- What can it change?
- Does a person approve important actions?
- Is the result checked?
Using a tool does not always make a system an agent.
AI agents
An AI agent works towards a goal and can take actions over several steps. It may choose tools, check what happened, and then take the next step.
Compare these examples.
Assistant: You ask for a meeting plan. It writes an agenda and stops.
Agent: You ask it to arrange the meeting. It checks calendars, suggests a time, waits for your approval, sends invitations, checks replies, and suggests a new time if needed.
The second system does more than write text. It acts in other systems and continues towards a goal.
The term agentic AI is still used in different ways. Some people use it for any tool-using system with several steps. Others use it for a group of agents working together for a long task.
Do not trust the label alone. Ask what the system can really do.
Levels of action
Autonomy means how much a system can do without a person taking part at every step.
Here is a simple scale:
- The system suggests. A person decides and acts.
- The system suggests an action and waits for approval.
- The system acts, but a person can watch and stop it.
- The system acts within its given access without waiting for approval.
More autonomy does not remove human responsibility. People still decide to use the system, give it access, and set its limits.
Compare the systems
| Description | What the facts support |
|---|---|
| A form applies published rates | Fixed-rule calculation |
| A trained model scores new email | Machine learning |
| A many-layer neural network names objects in a photo | Machine learning and deep learning |
| A trained model creates a new image | Generative AI |
| A service finds passages in a policy collection | Retrieval |
| A service finds policy, writes a draft, and sends it for approval | A system with retrieval, generation, rules, and human review |
| A system chooses tools and takes several steps towards a goal | Possible agent; check its actions, access, and limits |
Quick check
A product finds policy passages, writes a draft answer, and requires a manager to approve it. Which answer is best?
A. It is only retrieval.
B. It is only generative AI.
C. It uses retrieval, generation, and human approval.
D. It is fully autonomous.
Check the answer
Answer: C. The parts do different jobs. Several parts do not make the product fully autonomous.
Remember
- One product can use several methods.
- Deep learning is one type of machine learning.
- Generation creates content. Retrieval finds existing material.
- An agent works towards a goal through actions and steps.
- Always check what the system can access, change, and do without approval.
Next, we will learn the difference between building a model and using it.