1.1 Generative AI Versus Other Types of AI

1.1 Generative AI Versus Other Types of AI

Quick Answer: Discriminative and other traditional AI systems classify, predict, or detect. Generative AI creates new artifacts such as text, images, code, or audio. Pick a classifier or forecast when the business needs a label or a number. Pick generative AI when a person needs a draft to review.

The AI Transformation Leader exam (AB-731) asks you to describe how generative AI differs from other types of artificial intelligence (AI). Microsoft wrote the audience profile for business decision-makers who guide transformation with Microsoft 365 Copilot and Foundry Tools. You are not expected to write any code. You are expected to recognize the job a system is doing, choose a family of tools, and keep a human in the loop when the output can affect customers, employees, or filings.

Discriminative and traditional AI: classify, predict, detect

Most enterprise AI that shipped before large language models is discriminative. Analysts also call this traditional or classical machine learning (ML). The model learns a decision boundary. Given features from a record, an image, or a time series, it assigns a class, a score, or an alert.

Typical jobs look like this:

1. Classify — send an invoice to utilities, professional services, or needs review.
2. Predict — estimate next-quarter unit demand from history, seasonality, and promotions.
3. Detect — flag a payment that looks like fraud, or a safety incident in a plant camera feed.

The output is a label, a probability, or a numeric forecast. Teams score the system against a known answer: was the invoice coded correctly, and did demand land near the forecast? People still review exceptions. They are not editing a paragraph the model invented.

Predictive ML is the demand-forecast cousin of this family. It answers what is likely to happen. It does not write the board deck that explains the forecast to directors.

Foundry Tools is Microsoft's current name for the prebuilt API family that older documentation called Azure AI services (and, before that, Cognitive Services). Vision, speech, language, document intelligence, and search APIs often return structured results without asking you to train a foundation model. Those APIs still belong in the other-types-of-AI bucket when they classify, extract, or search rather than draft a new paragraph.

Generative AI: create text, images, code, and audio

Generative AI models learn patterns over language, pixels, or sound and then produce new content. A large language model (LLM) predicts the next token in a sequence. In business use that token stream becomes an email, a contract summary, or a first-pass board narrative. Image and audio models do the same job in other media.

What leaders actually buy is creation of artifacts:

• Draft a customer apology after a missed shipment
• Summarize a 40-page policy into talking points
• Generate a first-pass product image for a campaign review
• Suggest code or spreadsheet formulas — a valid use case even though AB-731 candidates do not write production software
• Turn a meeting transcript into notes and action lists

The output is new content, not a class label. Quality is usefulness, factual grounding, tone, and whether a named owner can safely send it. Characteristic risks are fabrications (confident wrong statements), uneven reliability, and bias. Later chapters in this guide treat those risks, token cost, and grounding in depth. This section only needs you to see that the failure mode is different from a mislabeled invoice.

Predictive ML versus generative AI

Do not treat AI as one budget line. A predictive model and a generative model solve different verbs.

Microsoft's product map as of the July 22, 2026 skills list puts everyday creation work in Microsoft 365 Copilot and Microsoft Copilot experiences. It puts model catalog, agents, evaluation, and many prebuilt APIs in Microsoft Foundry and Foundry Tools. Older pages still say Azure AI Foundry or Azure AI services; teach the current names. You do not need SKU-level product detail to answer this bullet. You need the job-to-be-done split. Selecting a solution is the next section. Copilot and Foundry chapters come later.

When a leader should pick each

Use discriminative or predictive AI when:

• The business already knows the allowed answers, such as a closed set of invoice codes.
• Success is a number you can score, such as forecast error or detection rate.
• You need the same decision every Tuesday, not a unique paragraph.
• Legal or finance requires a deterministic rule plus a score, not a story.

Use generative AI when:

• A skilled employee currently writes, rewrites, or synthesizes language or media.
• The artifact will be reviewed by a person who owns the send button.
• Variety is the point. Every customer email is different. A single label is not enough.

Use both, in sequence, when:

• A classifier decides what kind of work this is, then a generative model drafts the response.
• A forecast produces the number, then a generative model drafts the narrative a chief financial officer (CFO) still edits.

Scenario: invoices versus customer email

An operations vice president asks for AI on accounts payable. If the pain is mis-coded invoices, that is classification. A document-understanding or language classifier reads the PDF and proposes a general-ledger code. Generative AI is the wrong first tool. You do not need a poem about the invoice. You need the right bucket.

If the pain is slow, uneven replies when a vendor disputes a deduction, that is generation. Copilot in Outlook can draft a reply from the thread and an attached statement. A person in payables still sends it. Security and legal care that the draft did not invent a settlement the company never approved.

Scenario: demand forecast versus board narrative

A CFO already has a demand-planning model. Replacing it with an LLM is a category error. The forecast is a predictive ML job. Asking Copilot or a Foundry-hosted model to write the board narrative from the approved forecast, last quarter's commentary, and risk notes is a generative job. The board still expects the CFO's judgment. The model only accelerates the first draft.

Human resources (HR) and legal see the same split. Screening résumés into stages is classification. Drafting an offer letter in the company's voice is generation, with counsel review. Security sees it in detection: a model that scores a suspicious sign-in is not the same system as a model that drafts the incident report.

How this shows up for AB-731 candidates

Microsoft's skill bullet is: describe the differences between generative AI and other types of AI. Expect stems that name an outcome — route, score, draft, explain — rather than a brand. Match the system to the output type. If the stem asks for a number or a class, do not reach for generative AI. If it asks for language or media a human will edit, do not reach for a classifier alone.

Keep Microsoft's responsible AI expectations in view even at this foundation layer. Generative systems can invent plausible text. Discriminative systems can still be biased or brittle, but they do not typically fabricate a clause that was never in the contract. That difference drives later choices about grounding, human review, and when not to automate send.

Official Microsoft pages to reopen

• Study guide for Exam AB-731 — skills measured as of July 22, 2026, including this foundational bullet
• What is Microsoft Foundry? — current platform name for cataloguing and deploying models and agents
• What are Foundry Tools? — prebuilt vision, speech, language, search, and related APIs
• What is Microsoft Copilot? — LLM-powered productivity experiences grounded in web and/or work data