Generative AI Solutions

One filter decides every "is this generative AI?" question on the exam: is the output freshly composed content (a drafted email, a generated image, a conversational answer) or a fixed prediction like a number, a class, or a bounding box? That single test is what the Identify features of generative AI solutions objective rewards. The cheat-sheet above frames generative AI as one of the AI-900 workload categories; this section pins down the mechanics by name, so a stem can quote any one of them and you still place it in the larger picture.

Generative AI answers "compose something new," not "what is this?"

The other AI-900 workloads return a fixed result: a classification model answers "which category is this?", a regression model answers "what number?", and a computer-vision model returns a label or a bounding box. Generative AI instead produces new, original content (natural-language text, code, images, or audio) in response to an input. Microsoft Learn frames it directly: generative AI powers applications that can create content, answer questions, and assist with tasks^[1]. Typical generative tasks are summarizing a report, drafting or rewriting an email, answering a question conversationally, or generating an image from a description. The single mental filter for the exam: if the output is brand-new content, it is generative AI; if the output is a number, a class, or a box, it is predictive AI.

LLMs and the transformer architecture

Text-based generative AI is driven by large language models (LLMs) and their smaller relations, small language models (SLMs). Microsoft Learn describes them as models that encapsulate the linguistic and semantic relationships between the words and phrases in a vocabulary^[2], so the model can reason over natural-language input and generate relevant responses. LLMs are built on the transformer architecture, which Microsoft describes as consisting of two "blocks": an encoder block that creates the embeddings by applying a technique called attention^[2] (embeddings being the numeric vectors that capture each token's meaning) and a decoder layer that uses the embeddings calculated by the encoder to determine the next most probable token in a sequence^[2]. The crucial conceptual point: an LLM does not look up an answer. Microsoft says LLMs are trained to generate completions based on prompts^[2] and likens them to a super-powerful example of the predictive text feature on many cellphones^[2], it predicts the most probable next token, over and over. That probabilistic mechanism explains both the model's fluency and why its output is not guaranteed to be factually correct.

Tokens are the atomic unit, not whole words

Although humans think in words, LLMs operate on tokens. Microsoft Learn notes that while we tend to think of language in terms of words, LLMs break down their vocabulary into tokens^[2], and that tokens include words, but also sub-words (like the "un" in "unbelievable" and "unlikely"), punctuation, and other commonly used sequences of characters^[2]. The first step in training is to break the training text into these distinct tokens and assign a unique integer identifier to each one^[2], building a vocabulary of hundreds of thousands of tokens^[2]. Because the model consumes the prompt and produces its completion as sequences of tokens, token count, not word count, is the atomic unit the exam expects you to recognize. (Turning those tokens into the meaning-capturing embeddings introduced above, via attention, is worth knowing by name but is a level of depth AI-900 only touches lightly.)

Prompts in, completions out

The input/output contract is the most exam-portable fact here. Microsoft Learn states plainly: a prompt is simply the input you give to an LLM to get a response^[3], and the model responds to a prompt with a completion^[3]. There are two kinds of prompt: system prompts that set the behavior and tone of the model, and any constraints it should adhere to^[3], and user prompts that elicit a response to a specific question or instruction^[3]. The system prompt is normally set by the application; the user prompt is what a person types. Keep "prompt" (input) and "completion" (output) crisply apart, the exam tests that pairing directly. Pulling the pieces together, the figure below traces the path from prompt to completion: the prompt is split into tokens, the encoder turns them into embeddings using attention, and the decoder reads those embeddings to predict the next most probable token, looping one token at a time until the completion is assembled.

Knowing what an LLM is gets you halfway; the AI-900 blueprint also expects you to recognize how generative AI is delivered (copilots), how it is made trustworthy on your data (grounding via RAG), and how its specific risks are managed (responsible generative AI). These three ideas are the back half of "features of generative AI solutions."

Copilots wrap the prompt-and-completion loop into everyday apps

A copilot is an AI-powered assistant that embeds the prompt-to-completion loop inside the tools people already use. Microsoft 365 Copilot, for example, brings generative AI into Word, Excel, Outlook, and Teams so users can draft, summarize, and create content without leaving the app. To keep a conversation coherent, copilots and chat assistants often keep track of the conversation history and include summarized versions of it in subsequent prompts^[3], giving the model ongoing context to build on across turns. For the exam, recognize a copilot as the user-facing packaging of an LLM, not a different kind of model.

Prompt quality is a lever you control

Because a completion is generated from the prompt, the prompt materially changes the result. Microsoft's guidance is concrete: be clear and specific^[3] (explicit instructions beat vague language), add context^[3] (topic, audience, format), use examples^[3] of the style you want, and ask for structure^[3] such as bullet points or tables. The exam-portable point: better results often come from a better prompt, not a bigger model.

Grounding with retrieval augmented generation (RAG)

An LLM only knows what was in its training data, so it cannot answer about your private documents or recent facts, and, generating probabilistically, it can produce confident but wrong answers (commonly called hallucinations). Grounding fixes the knowledge gap. Microsoft Learn describes retrieval augmented generation (RAG) as an approach that involves retrieving information, like documents or emails, and using it to augment the prompt with relevant data^[3], so that the response generated by the model is then grounded in the information that was provided^[3]. The canonical example: an expenses assistant first retrieves the relevant section of the company expenses policy, adds it to the prompt alongside the user's question, and the model answers from that supplied context rather than guessing. Recognize RAG as the answer whenever a stem says the assistant must reflect internal, organization-specific, or up-to-date information.

Responsible generative AI: harms and layered mitigation

Because generative AI creates content, it carries heightened responsible-AI concerns. Microsoft frames the work as four stages applied after the initial planning of a solution. Microsoft Learn organizes it as planning the solution, then map potential harms, measure potential harms, mitigate potential harms, and manage (operate) a responsible solution^[4], i.e. map, measure, mitigate, and manage, the four stages the exam shorthand names. Mitigation itself is applied in layers, not one place: the model^[5] itself, a safety system with content filters, the system message and grounding data that steer and constrain the model, and the user experience (the application UI and documentation). The takeaway for AI-900: generative AI does not replace responsible-AI thinking, it intensifies it, and the mitigation is a stack rather than a single switch.

Generative-AI questions on AI-900 follow a handful of repeatable shapes. Naming the pattern lets you answer on recognition under time pressure instead of recall. Each pattern below pairs the stem signal with the right answer and the distractor that is built to look right.

Pattern 1: "Is this generative AI?", decided by the output type

The stem describes a task and asks which AI workload it is. The deciding clue is the kind of output:

• Summarize a document, draft/rewrite text, generate code, answer questions conversationally, create an image → generative AI (the output is new content).
• Predict a number (price, demand) → regression; assign a category (spam/not-spam, sentiment) → classification; detect objects / read a form → computer vision or document intelligence. These are not generative AI because the output is a fixed prediction, not composed content.

The classic distractor labels a summarization or chat task as "classification" or "NLP analysis." It is wrong because nothing is being categorized, something new is being written.

Pattern 2: "Which component / term?", the vocabulary chain

The stem quotes one mechanism and asks you to name it, or offers four terms and asks which fits. Anchor on the chain tokens → embeddings → transformer (encoder/decoder + attention) → prompt → completion:

• "The unit an LLM breaks text into" → token (not "word", that is the built-in trap, since humans think in words).
• "The architecture LLMs are built on" → transformer.
• "The input you give the model" → prompt; "the model's response" → completion. Swapping these two is the most common distractor.
• "Sets the model's behavior and constraints" → system prompt, versus the user prompt that asks the specific question.

Pattern 3: "The assistant must use our private/recent data", grounding with RAG

When a stem says responses must reflect internal documents, company policy, or current facts the model was never trained on, the answer is grounding / retrieval augmented generation (RAG), retrieve the relevant data and add it to the prompt. The distractor proposes "retrain the model" or "fine-tune on the documents," which AI-900 treats as heavier and unnecessary for simply supplying current context; RAG is the lightweight, exam-correct move. A second distractor says "just ask the base model," which is wrong because the model has no knowledge of private or recent data.

Pattern 4: "Why is the answer wrong / made up?", the probabilistic-generation limitation

The stem describes the model returning a fluent but factually incorrect or invented answer and asks for the cause or the fix. The cause is that the model generates the next token probabilistically rather than retrieving a verified fact, a hallucination. The fix is grounding (RAG) to supply real data and/or responsible-AI mitigations (content filters, a clear system message, human review). A distractor that says "the model is broken / needs more training data" misses the point: fluent-but-wrong output is an inherent property of generative models, mitigated by grounding and guardrails, not a defect cured by more of the same training.

Pattern 5: "What's a feature/benefit vs a limitation/risk of generative AI?", the balanced pair

Some stems ask you to pick a true capability (creates original text/code/images, summarizes, answers conversationally, powers copilots in everyday apps) or a true limitation/risk (can produce confident but inaccurate output, lacks knowledge of private or recent data, can generate offensive or biased content, is non-deterministic). The trap pairs a real capability with an overstated claim like "always factually accurate" or "gives the same answer every time", both false, because generation is probabilistic. When a stem frames generative AI as deterministic, auditable, or guaranteed-correct, that option is wrong; those are properties of a deterministic program or a conventional predictive model, not an LLM. (Naming the specific Azure services that host generative AI, Azure OpenAI, Azure AI Foundry, the model catalog, is the separate Azure GenAI services subtopic; this section is about the concepts, not the SKUs.)