Art of AI · The systems guide
AI systems,
made understandable.
From recognising a sound to creating an image or coordinating a task: discover how AI systems work, where they help, and what to examine before you rely on them.
By Penny Wraiter · Updated 10 October 2026 · AI-assisted editorial guide
01 / Get your bearings
What is an AI system?
An AI system takes inputs and infers outputs that serve an objective: a prediction, a recommendation, a piece of content or a decision. Its inputs might be words, images, measurements or signals from the world. Systems differ in how independently they operate and whether they adapt after deployment. This broad view follows the OECD’s explanation of an AI system.
A chatbot is one familiar interface. AI also appears inside search, cameras, translation software, design tools and industrial equipment. Start by asking what a system actually does, what information it uses and what happens when it gets something wrong.
Model
A model maps inputs to outputs. In machine learning, training adjusts it using examples; inference uses it to produce a result. Some AI also uses explicitly represented knowledge and rules.
System
The model operates within software, data flows and controls. A document assistant might combine a model, a search index, access permissions and a user interface.
Product or service
This is what you install, subscribe to or integrate. Compare its actual features, settings and terms: the same model can sit behind very different user experiences.
02 / Find the right kind of help
What can AI systems do?
These are useful ways to group tasks, rather than rigid product categories. A single application may combine several of them.
Identify and classify
Turn speech into text, sort incoming requests or flag a possible defect in an image.
Ask: which important cases does it miss, and how often does it raise a false alarm?
Predict and detect patterns
Estimate demand, identify unusual sensor readings or support planning with historical data.
Ask: does the test data resemble the situation where this will be used?
Create and transform
Draft text or code, propose an illustration, transform an image or help prepare audio and video.
Ask: is the result accurate, useful and suitable for the intended audience?
Find and explain information
Search a collection of documents and use retrieved passages to help answer a question.
Ask: do the linked sources actually support the answer, and are they current?
Work across media
Multimodal systems handle more than one kind of input or output, such as a photo and a written question.
Ask: which formats are supported, and what detail is lost during processing?
Coordinate tasks
Tool-connected assistants can carry out steps such as retrieving a file, calculating a result or preparing an update.
Ask: what can it change, what needs approval, and can you inspect or undo the action?
For the underlying learning approaches, see Google’s introduction to machine learning. The examples and evaluation questions here are AOI’s practical reading of these capabilities.
03 / Look inside the workflow
An answer is only one part of the system.
A useful way to inspect an AI application is to follow the information through it. This simplified workflow also shows where people can set limits and check results.
- STEP 01InputA question, file, image or sensor reading. Define the task and the data it may use.
- STEP 02ProcessingA model works with instructions and context; some systems also retrieve information or call tools.
- STEP 03OutputA score, draft, image, recommendation or proposed action reaches a person or another program.
- STEP 04ReviewCheck the result, approve consequential actions and record failures so the workflow can improve.
This is an explanatory map, not a claim that every product includes a review step. Whether that step exists is something to verify.
Training, context and memory
Training changes a machine-learning model. Supplying a document or continuing a conversation can change the context available for an answer without retraining that model. A product’s saved “memory” is another mechanism again. Check its settings before assuming information will be remembered, forgotten or excluded from future training.
Retrieval and connected tools
Retrieval-augmented generation, or RAG, brings selected external information into a generation workflow. It can make an answer more relevant to your documents, but the system still has to retrieve the right passage and use it correctly. Google Cloud explains the retrieval process. Tools add capabilities; they also add permissions and possible side effects to examine.
04 / Decide where it runs
Local, cloud or a combination?
The deployment choice affects convenience, control and ongoing work. Compare the whole application, including storage, extensions and external connections.
| Approach | What it means | What to examine |
|---|---|---|
| Local | The model runs on your own computer or infrastructure. Depending on the application, inference may work offline. | Hardware capacity, installation, updates, backups and any network traffic. Local execution alone does not prove that every component keeps data on the device. |
| Cloud | A provider operates the model remotely, through an app, website or software interface. | Data handling, retention controls, usage charges, service availability and whether results and settings can be exported. |
| Hybrid | Different parts of the workflow run in different places: for example, local preparation followed by remote processing. | Exactly which data crosses the boundary, which service receives it and who maintains each part. |
Open weights and open source also need a closer look. Downloadable model weights do not, by themselves, tell you which code, training materials or usage permissions are available. Read the specific licence and documentation before choosing a deployment.
05 / Choose by evidence
Start with a task you can judge.
A spectacular demonstration answers someone else’s question. Write down your own: “Can this help turn our approved notes into a useful first draft?” is easier to evaluate than “Which AI is best?”
For a first comparison, use a small, non-sensitive set of realistic examples. Include awkward inputs and one case where the correct response is to ask for more information. Keep the inputs and acceptance criteria consistent across systems.
Count the time spent correcting the result as well as the time spent generating it. Sometimes a search, spreadsheet or ordinary rule is the clearer solution.
- Define success. Specify the output, audience and acceptable error level.
- Check data requirements. Identify what must be uploaded, stored or connected.
- Try representative tasks. Record the product, version where available, date, settings and results.
- Measure the whole workflow. Include review effort, recurring cost and failed attempts.
- Check control and exit. Confirm permissions, export options and how to stop or reverse actions.
06 / Keep judgment in the loop
Capability does not remove uncertainty.
Generative systems can produce confident but incorrect statements, fabricated references or misleading details. Other risks include privacy exposure, harmful bias and information-security failures. NIST’s Generative AI Profile describes these risks and approaches to evaluating and managing them.
Our practical rule is to match review to consequence. An unexpected texture in an exploratory image is different from an invented figure in a published report. Check important claims against their sources; test code before using it; review generated media for errors and suitability.
For systems that act, define the allowed workspace and the point at which a person must approve a change. A helpful conversational tone is not evidence that an action was completed. Look for the saved file, confirmed update or other observable result.
AI is most useful when you can tell what it contributed, what remains uncertain and where your own decision begins.
07 / Put the ideas to work
Choose your next guide.
Move from the general picture to a particular tool, creative workflow or business problem.
40 AI App Builders and Development Tools
Choose by use case: web products, mobile apps, business tools, websites and cloud workspaces. A substantial reality check with official sources, practical evaluation questions and linked art-library detours.
Read the 2026 guide →Deep Dream Generator
Understand the platform, distinguish it from the original DeepDream technique, and plan a first image experiment.
Read the platform guide →Repeatable AI images
Develop a clear brief, keep useful settings and compare results in a way you can learn from.
Build an image workflow →Build your working toolkit
Continue with tools, tutorials and resources for using AI in everyday creative and technical work.
Explore the tools and learning hub →Bring a business problem
Explore how Art of AI and PRODAT can help scope a project around its workflow, people and intended outcome.
Explore business applications →Also explore the AI-o-philic toolkit and the question of control, or use the Art Style Database to develop a visual vocabulary for creative work.
A few common questions
Is every AI system generative?
No. Many systems classify information, estimate values or recommend an option. Generative AI is the part of the field concerned with producing content, including text, images and sound.
Does an AI assistant automatically know current information?
No. Check whether the product can retrieve current information, whether that capability was actually used and what sources it returned. A recent-looking answer or a citation is not enough by itself.
What makes an AI system an “agent”?
The term is used in different ways. For evaluation, look past the label: can it choose and execute steps, use tools, inspect results and continue towards a goal? Then check the permissions, stopping conditions and approval controls around that behaviour.
Do I need to learn programming first?
Many applications can be used through a normal interface. Programming becomes useful when you need custom integrations or repeated automated tasks. Begin with a small task you understand well enough to check.
Sources and further reading
Primary references checked on 4 October 2026. This overview explains general concepts; product capabilities and terms should be checked in each provider’s current documentation. The selection questions and workflow examples are AOI editorial guidance.
- OECD: Explanatory memorandum on the updated definition of an AI system — scope, outputs, models and autonomy.
- Google for Developers: What is machine learning? — learning approaches and applications.
- Google Cloud: Retrieval-augmented generation — retrieval and generation in one workflow.
- NIST: AI Risk Management Framework and Generative AI Profile (PDF) — evaluating risks in context.