AB-731 generative AI business value explained
Identify the business value of generative AI solutions is worth 35–40% of AB-731, and it is the exam’s vocabulary domain. It asks you to explain how generative AI differs from other AI, what drives its cost, what makes it reliable or unreliable, and when it actually pays off. You are not asked to build anything, but you must reason like someone who signs off budgets.
Generative AI versus other AI
| Type | What it does | Example |
|---|---|---|
| Machine learning (predictive) | Learns patterns from data to predict or classify | Forecast demand, flag likely churn |
| Generative AI | Produces new content: text, images, code, summaries | Draft a proposal, summarise a meeting |
| Rules and automation | Follows fixed logic | Route an invoice above a set amount |
Choose generative AI when the task involves language or content, variation is acceptable and humans can review the output. Choose predictive ML when you need a number or a class from structured history. Choose plain automation when the logic is fixed and must be exact. The outline asks you to select a solution for a business need, so practise naming the right type before the right product.
Models, cost and value
| Concept | Leader-level meaning |
|---|---|
| Pretrained model | General model trained on broad data; fast to adopt |
| Fine-tuned model | A pretrained model further trained on your examples to behave in a specific way; costs more effort and needs good data |
| Token | A chunk of text; usage is measured and billed by tokens in prompts and responses |
| ROI | Value created (time saved, quality, revenue) against licence, usage and change costs |
Cost rises with tokens, so long inputs, long outputs and high volumes all matter. The cheapest levers are shorter prompts, sending only relevant content, and choosing a smaller model when the task is simple. Value comes from scale and automation: tasks that are frequent, language-heavy and tolerant of review.
Reliability, risk and security
- Fabrications. Models can produce confident, wrong statements. Controls: grounding, human review, clear scope.
- Bias. Outputs reflect the data. Controls: representative datasets, testing across groups.
- Reliability. Results vary between runs. Do not use generative AI where a precise, repeatable calculation is required.
- Security. Consider application security, data security and authentication: who can reach the system, what data enters it, and how it is protected.
Prompts, grounding and RAG
Prompt engineering changes output without changing the model. Useful techniques at a business level: give a role, state the task and audience, show an example, ask for a format, and break complex requests into steps. Grounding ties answers to your own data. Retrieval-augmented generation (RAG) does this by retrieving relevant content first and giving it to the model along with the question. RAG is usually the first choice for company knowledge because it stays current, keeps data under your control and costs less than fine-tuning.
Data matters as much as the model. Check type (structured, documents, images), quality (accurate, current, complete) and representativeness (does it cover the people and cases you serve?). Poor data produces confident wrong answers.
Machine learning still has a place: know the lifecycle at a high level (define the problem, prepare data, train, evaluate, deploy, monitor and retrain) and that models drift when the world changes.
Example questions
Question 1. A retailer wants to identify which customers are likely to cancel next quarter, using three years of structured account history. Which approach fits best?
- A. A generative model that writes retention emails
- B. A machine learning model trained on the account history
- C. An image generation model
- D. A fixed rule that flags every customer with no recent login
Show answer
Answer: B
Predicting a likely outcome from structured history is a machine learning task. A text generator produces content rather than reliable predictions from records. An image model is unrelated to account data. A fixed rule such as flag everyone who has not logged in cannot learn patterns and will miss most signals.
Want more questions like this? Full AB-731 practice tests →
Question 2. A firm wants an assistant to follow its house writing style and terminology consistently, and it has thousands of approved example documents. Which choice best matches this need?
- A. Make every prompt as short as possible
- B. Increase the size of the prompt input box
- C. Switch the model off and write everything by hand
- D. Consider fine-tuning a pretrained model on the approved examples
Show answer
Answer: D
Fine-tuning a pretrained model on approved examples shapes its style and behaviour consistently. Shortening the prompt does not teach a style. A bigger prompt box does not change the model. Switching off the model discards the value. A pretrained model alone can imitate a style only loosely.
Want more questions like this? Full AB-731 practice tests →
Question 3. A company wants employees to get answers from the latest version of its policies, which change monthly, with little retraining effort. What should it use?
- A. Retrieval-augmented generation over the policy library
- B. Fine-tuning the model again each month
- C. Rely only on the pretrained model’s general knowledge
- D. Tell the model never to answer policy questions
Show answer
Answer: A
RAG retrieves the current policy text at question time, so updates take effect when documents change, with no retraining. Fine-tuning would need repeating each month and does not guarantee current facts. Relying only on what the model learned in training leaves it unaware of private policies. A rule that the model must never answer makes the assistant useless.
Want more questions like this? Full AB-731 practice tests →
Next, move on to the Microsoft AI apps and services guide to see which Microsoft products implement these ideas.