Getting Started with AI

Last reviewed: August 2026

Decide First: Is AI the Right Fit for This Problem?

Before choosing an interface or a model, confirm whether this is even an AI problem. Generative AI is strong at producing plausible text, code, and images, but it is not a source of verified facts, up-to-date internal data, or deterministic calculation. Use these four questions to filter.

Is this a problem you can solve precisely with rules and computation, or one that tolerates uncertainty like generation, summarization, or classification?
Do you need up-to-date information or internal, company-specific grounding?
Can a human review wrong answers or wrong actions, or can you constrain them with permissions?
Can you define success criteria, representative test inputs, and a cost ceiling?

Nature of the problem	Recommended approach
Solved by deterministic rules/computation	Conventional software (no AI needed)
Structured-data prediction/classification	Traditional ML (ML Platforms)
Generation, summarization, coding assistance	Generative AI (via the three ways below)
Requires up-to-date/internal grounding	Generative AI + RAG or tool use
High-impact judgment or state-changing actions	Human approval and least privilege required (AI Security)

If the problem ends with rules and statistics, stop here and use conventional software. If AI is a fit, decide which interface to use below.

Three Ways to Use AI

The first thing to decide in AI adoption is not “which technology” but “through which interface will we use AI.” There are three broad ways, and one organization often uses all three in parallel.

Way to use	What it is	Primary users	Representative examples
① Conversational AI apps & platforms	Subscribe to finished chatbots/copilots and use them right away. Build internal bots with no-code builders	Non-technical staff, business planners	ChatGPT, Gemini app, Microsoft 365 Copilot, Copilot Studio
② AI coding tools	Developers delegate code writing/editing to AI in the IDE or terminal	Software engineers	GitHub Copilot, Claude Code, Codex, Kiro, Grok Build
③ API & SDK	Apps and agents call models programmatically and embed them in products	Development teams, system integrators	Amazon Bedrock API, Azure Foundry SDK, api.openai.com

This document covers the concepts and starting points of ① and ②, and how to use ③ (building directly with APIs) below.

Why Use AI on the Cloud

Pre-training a frontier foundation model from scratch requires tens of millions of dollars in massive GPU clusters, trillions of training tokens, and months of compute time. Most organizations do not do this themselves and instead use ready-made AI services provided by cloud vendors.

By analogy, it is like drawing electricity from a power plant or the power grid rather than generating your own. We focus on “what to do with electricity,” not “how to generate electricity.”

Three Key Concepts

1. Foundation Model (FM)

A general-purpose AI model pre-trained on massive data. The most common type is an LLM (Large Language Model). GPT (OpenAI), Claude (Anthropic), Gemini (Google), and Nova (Amazon) are examples.

Used via API calls — no need to train yourself.
Think of it as “a smart assistant that answers questions.”

2. Prompt

The input message sent to the model. Written in natural language, such as “Tell me the capital of Korea.” Output quality varies greatly depending on how you ask.

3. Token

The unit a model uses to process text. Roughly one word ≈ 1–2 tokens. Most APIs bill by input/output token count.

Background: The AI Technology Lineage

“AI” is not a single technology. It is a collective term for decades of techniques, and each generation coexists with rather than replaces the previous one. The three ways above mainly deal with the latest generations — generative and agentic AI — but traditional ML and deep learning are still in use.

Generation	Core Tech	What It Does	Cloud Services
Traditional ML	Regression, classification, clustering, trees	Predict/classify structured data (churn, anomaly detection, recommendations)	SageMaker AI, Azure ML, Vertex AI, OCI Data Science
Deep Learning	CNN, RNN, Transformer	Process unstructured data (image recognition, speech, translation)	GPU instances + ML platforms
Generative AI	Foundation models (LLM, multimodal)	Generate text/image/code/speech	Bedrock, Microsoft Foundry, Gemini
Agentic AI	LLM + tool use + autonomous execution	Given a goal, plans, executes, and verifies on its own	AgentCore, Foundry Agents, Gemini Agent Platform

Method: How Do You Extend AI

Even with the same foundation model, results differ depending on the method by which you adapt it to your problem. The methods below are not exclusive to one way of using AI — they run through all three ways. For example, prompt improvement appears as “custom instructions” in ① conversational apps, connecting your data appears as “file upload” in ① or knowledge grounding in a no-code builder, and as a RAG pipeline in ③ the API.

Cost and complexity increase as you move down, and the stages are most distinct when building directly with ③ (API/SDK).

graph TD
    A[1. FM API Call] --> B[2. Prompt Engineering]
    B --> C[3. RAG - Connect Your Data]
    C --> D[4. Fine-tuning]
    D --> E[5. Train From Scratch]

Stage 1: API Call

The simplest starting point. Send a question to one of Amazon Bedrock, Microsoft Foundry, Gemini Enterprise, or OCI Enterprise AI and receive an answer. Ready to use for chatbots, document summarization, translation, and similar tasks. See AI Platform and Model Comparison for vendor comparisons, and When This Helps for a broader list of use cases.

Stage 2: Prompt Engineering

Same API, dramatically different results depending on how you ask. Giving a clear role and instruction improves quality without writing any code. For detailed design techniques, see Prompt Engineering.

Stage 3: RAG (Retrieval-Augmented Generation)

Foundation models are rich in general knowledge but do not know your company’s data. RAG retrieves your documents, includes the relevant parts in the prompt, and then sends it to the model (analogy: “an open-book exam”). Used for internal document-based chatbots, product FAQ, and legal/medical document lookup.

RAG works with Vector Stores. See Advanced RAG Patterns for implementation details.

Stage 4: Fine-tuning

Fine-tune the model with your data. It can be optimized for a specific task but cost and time increase significantly.

Use cases:

Understanding industry-specific terminology
Reflecting a company’s unique tone/style
Enforcing a specific output format

Stage 5: Train From Scratch

Unnecessary for most organizations. This is what companies like Google, OpenAI, and Anthropic do.

For model operations, evaluation, and cost tracking see LLMOps. For AI security and guardrails see AI Security.

When to Use Which Method

The table below takes a beginner’s learning-order view (simple → advanced). For a practical routing view that maps requirements directly to approaches, technologies, and docs, see AI System Lifecycle — Technical Task Selection Guide.

Situation	Recommended
Want a quick prototype	Stage 1 (API call)
Using API but quality is lacking	Stage 2 (Prompt engineering)
Want answers grounded in company documents	Stage 3 (RAG)
Specialized domain the general model doesn’t know well	Stage 4 (Fine-tuning)
Entirely new problem with no existing model	Stage 5 (Train from scratch)

Why Not Build Your Own AI Model

Factor	Train from Scratch	Cloud API
Initial cost	Tens of millions in GPU infrastructure	Cents to dollars per API call
Time	Months to years	Minutes (API integration)
Data	Trillions of training tokens needed	Model already trained
Talent	Many ML engineers/researchers	Achievable with general developers
Updates	Retraining required	Vendor auto-updates
Effectiveness	State-of-the-art possible	Sufficient for most tasks

Common Mistakes

Skipping prompt engineering and jumping to fine-tuning — Prompt improvements alone often suffice, yet teams invest the cost/time of fine-tuning first.
Designing without considering token costs — Sending long system prompts on every request or including unnecessarily large documents in context causes cost explosion.
Relying solely on model internal knowledge without RAG — FMs don’t know post-training information or internal data, causing hallucinations.

Checklist

Following the staged approach: API call → prompt engineering → RAG → fine-tuning
Monitoring token usage and costs with budget caps set
RAG pipeline configured for cases requiring answers based on internal data