Domain 1: Business Value of AI (35-40%)
Exam Tip
The AB-731 exam is for Business Leaders, not developers. When in doubt, choose the answer that focuses on Strategy, ROI, and Governance over technical implementation or coding.
GenAI vs Traditional AI
| Traditional AI | Generative AI |
|---|---|
| Analyzes, classifies, predicts | Creates new content |
| Spam filter, fraud detection | Write emails, generate images |
Foundation vs Specialized Models
- Foundation Models: Large, generic models (like GPT-4o) pre-trained on massive data. They "know everything" generally.
- Specialized Models: Models fine-tuned for a specific task (e.g., medical diagnosis, legal coding). They are more efficient for niche tasks.
Key Terms
Key AI Terms
What affects output quality?
(Click to reveal)| Term | Remember |
|---|---|
| Prompt | Input → affects output quality |
| Token | Text unit → affects cost |
| Context window | Input + output limit |
| Temperature | 0 = predictable, 1 = creative |
Critical
Tokens are the primary cost driver. Understand that context window limits (Input + Output) directly impact both the cost and the richness of the AI's response.
Data Maturity: The Prerequisite
Before seeing ROI, an organization must have a foundation of Data Maturity.
- Quality: AI is only as good as the data it's grounded in ("Garbage In, Garbage Out").
- Accessibility: Data must be broken out of siloes so the AI can "see" it (via RAG).
- Security: Robust permissions ensured before AI deployment to prevent oversharing.
- Labels: For specialized tasks, high-quality labeled data is required for Fine-tuning.
Business Value Areas
- Productivity: Draft emails, summarize meetings, analyze data
- Decision-making: Faster insights from more data
- Automation: Handle routine tasks (FAQs, document processing)
- Customer experience: Personalized, faster responses
Value Types:
- Efficiency: Time saved, faster completion of routine tasks.
- Growth: New business models, personalized customer insights.
- Risk Mitigation: Better compliance tracking, early anomaly detection.
Trap
Questions often ask about ROI. Include all value types (Efficiency, Growth, and Risk), not just cost savings.
Model Customization Strategies
When the base model isn't enough, leaders must choose how to customize it.
| Strategy | Effort | Cost | Data Needs | Best Use Case |
|---|---|---|---|---|
| Prompt Engineering | Low | Low | None | Controlling style, formatting, and simple tasks using instructions. |
| RAG (Retrieval-Augmented Generation) | Medium | Medium | Knowledge Base | Grounding the model in your latest business data/documents. |
| Fine-tuning | High | High | Labeled Dataset | Achieving a very specific style or deep domain expertise (rarely needed). |
Strategic Rule of Thumb:
- Always start with Prompt Engineering.
- Move to RAG if the model needs to "know" specific business data.
- Only consider Fine-tuning if RAG and Prompts can't achieve the required tone or specialized format.
Model Weight Updates & Terminology
A common exam trap is confusing which customization method actually changes the model itself.
| Term | Are Weights Updated? | Core Concept |
|---|---|---|
| Pre-training | YES (Extensive) | Creating the model from scratch on massive datasets. This is where the model "learns" language. |
| Fine-tuning | YES (Targeted) | Taking a pre-trained model and continuing training on a small, specific dataset to adapt its behavior/style. |
| RAG | NO | The model remains frozen. It uses external data as "context" in the prompt to ground its answers. |
| Prompt Engineering | NO | Providing instructions and examples to guide the model's existing knowledge. |
Terminology Tip
- Grounding: Providing context via RAG so the model doesn't hallucinate. (No weight change)
- Transfer Learning: The underlying principle of Fine-tuning—building on top of a pre-trained model. (Weight change)
Adaptation vs Retrieval (Fine-tuning vs RAG)
This comparison focuses on where the knowledge comes from.
| Concept | Term | Analogy | Best for... |
|---|---|---|---|
| Adaptation | Fine-tuning | Learning a new skill or language style from a textbook. | Nuanced behavior, specialized industry jargon. |
| Retrieval | RAG | Taking an "open-book" exam with access to a library. | Facts, real-time data, and proprietary internal docs. |
Common Pitfall
Fine-tuning is NOT for real-time data. A common exam trap is asking how to provide an AI with today's stock prices. The answer is RAG/Grounding, not Fine-tuning.
Model Customization Quiz
Which method uses a Knowledge Base to ground answers without updating weights?
(Click to reveal)