AI-901 practice questions and answers

Correct answer: A — Training learns patterns from data, while inference applies the trained model to new data

Training is where the model's parameters are fitted to data, and inference is where those fixed parameters produce predictions for new input. The two are distinct operations, and inference does not need labels because the label is what it predicts.

Correct answer: A — Object detection

Object detection returns both the class and the bounding box coordinates for each object found. Image classification assigns a single label to the whole image, and OCR extracts text characters rather than locating general objects.

Correct answer: A — The chunk size and overlap used during ingestion

Broken chunks are an ingestion problem, so tuning chunk size and adding overlap so ideas are not split fixes the root cause. Temperature, region, and replica count do not repair damaged retrieval content.

Correct answer: A — Optical character recognition

OCR converts text in images into machine-readable characters, including handwriting in modern models. Speech recognition works on audio, sentiment analysis judges tone in text that has already been extracted, and entity linking connects mentions to knowledge base entries.

Correct answer: A — Microsoft Entra ID with a managed identity and role-based access control

A managed identity removes secrets from the application entirely and RBAC scopes what the identity may call. Hardcoded keys and repository-stored credentials leak, and anonymous access removes authorisation altogether.

Correct answer: A — Recall

Recall measures the proportion of actual positives that were caught, which is what matters when missing a fraud case is worse than investigating a false alarm. Precision measures how many flagged cases were real, and mean absolute error and R-squared are regression metrics.

Correct answer: A — Image captioning

Captioning produces a natural language description of an image. Face detection locates faces, spatial analysis interprets movement in video, and background removal isolates the foreground.

Correct answer: A — Regression

Predicting a continuous numeric value such as a price is regression. Classification predicts a discrete category, clustering groups unlabelled records, and anomaly detection flags unusual observations rather than estimating a value.

Correct answer: A — In the tool definitions and the identity's role assignments, so the capability does not exist

A prompt is guidance, not a security boundary, so a destructive capability the agent must never have should simply not be exposed as a tool or permitted by its role. Temperature and token limits are generation settings with no authorisation effect.

Correct answer: A, B — Temperature · Top P

Temperature and top P both shape the sampling distribution over next tokens and therefore how varied the output is. Max tokens caps length rather than creativity, and region and endpoint name are deployment details.

Correct answer: A, B — It predicts the next token based on the tokens it has already seen · It is pre-trained on very large volumes of text

Language models are next-token predictors trained on huge corpora, which is what gives them broad capability. They do not store documents verbatim for lookup, they can produce confident but wrong output, and they can perform many tasks with no task-specific labelled data.

Correct answer: A — Clustering

Grouping unlabelled records by similarity is clustering, an unsupervised technique. Multiclass classification requires known labels, regression predicts numbers, and object detection applies to images.

Correct answer: A — The model catalog and deployments in the Foundry portal

Microsoft Foundry provides a model catalog for discovery and comparison and a deployments experience for standing up an endpoint. Monitor workbooks visualise telemetry, Resource Graph queries inventory, and Policy enforces governance rules.

Correct answer: A — Inclusiveness

Inclusiveness is the principle about empowering everyone and engaging people across abilities and backgrounds. Reliability concerns consistent behaviour, privacy concerns data protection, and accountability concerns human answerability.

Correct answer: A — Sentiment analysis

Sentiment analysis scores text for emotional polarity. Key phrase extraction returns the main talking points, language detection identifies which language the text is in, and translation converts it to another language.

Correct answer: A — Azure AI Search

Azure AI Search provides vector, keyword, and hybrid retrieval and integrates directly with Foundry as a grounding data source. Service Bus and Event Grid are messaging services, and Batch runs compute jobs.

Correct answer: A — Flagging a credit card transaction that does not fit the account's normal pattern

Anomaly detection identifies observations that deviate from an established normal pattern, which is exactly what unusual transaction flagging does. Translation, image generation, and summarisation are language and generative workloads.

Correct answer: A, B — Ground responses in retrieved source documents · Instruct the model to answer only from the supplied context and to say when it does not know

Grounding and an explicit instruction to abstain when the context is silent are the two standard mitigations. Higher temperature increases invention, removing the system message removes the guardrail instructions, and concurrency is a scaling concern.

Correct answer: A — Azure AI Document Intelligence

Document Intelligence provides prebuilt and custom models that return typed fields and tables from documents such as invoices and receipts. Translator converts languages, Speech handles audio, and Content Safety filters harmful content.

Correct answer: A, B — Intent · Entity

The user's goal is the intent and the extracted details are entities, and together they drive the bot's response. Epochs, gradients, and checkpoints are model training concepts rather than conversational language understanding concepts.

Correct answer: A — Use a standard pay-as-you-go deployment with quota limits and monitor token usage

Consumption-based deployment with quota limits matches cost to actual usage and quotas cap runaway spend, while usage monitoring shows where tokens go. Permanent maximum provisioning pays for idle capacity, disabling logging removes visibility, and raising max tokens increases cost.

Correct answer: A — Generative AI creates new content such as text or images rather than only predicting a label or value

The defining difference is output: generative models produce new content, while predictive models output a class or a number. Generative models are trained on very large corpora, they usually run in the cloud, and their quality can be evaluated with human review and automated metrics.

Correct answer: A — Retrieval Augmented Generation using an index over the documents

RAG retrieves the relevant passages at query time and grounds the answer in them, so refreshing the index is all a weekly document update requires. Weekly fine-tuning is slow and expensive for changing facts, pre-training is out of reach for this purpose, and temperature only affects output variety.

Correct answer: A — The input text and instructions given to the model to guide its output

A prompt is the instruction and context supplied at inference time that shapes what the model generates. Learning rate, hardware, and training corpora are all aspects of building the model rather than using it.

Correct answer: A — Evaluations with built-in quality metrics

Foundry evaluations run a dataset through a prompt or flow and score outputs with metrics such as groundedness, relevance, and coherence. Budgets track spend, Advisor gives platform recommendations, and resource locks prevent accidental deletion.

Correct answer: A — A private endpoint on the resource with public network access disabled

A private endpoint gives the service a private IP inside the virtual network, and disabling public access ensures traffic cannot arrive any other way. A client-side network security group does not change the destination path, key rotation is an authentication practice, and audit policies only report.

Correct answer: A — Agents with tools or function calling

Agents and function calling let the model choose a defined tool, receive its result, and compose an answer from it. Batch pipelines score data offline, labelling projects annotate training data, and compute autoscaling manages capacity.

Correct answer: A — Azure AI Language

Azure AI Language exposes prebuilt natural language features including entity recognition, PII detection, key phrases, and summarisation. Machine Learning compute is infrastructure for custom models, and Data Factory and Synapse are data integration and analytics services.

Correct answer: A, B — Transparency · Accountability

Being able to explain a decision makes the system understandable, which is transparency, and it lets people be held answerable for outcomes, which is accountability. Inclusiveness is about serving everyone, reliability is about consistent safe behaviour, and privacy is about protecting data.

Correct answer: A — A project within a Foundry resource

A Foundry project is the workspace that holds a team's models, connections, indexes, evaluations, and deployments together. Resource locks prevent deletion, availability sets are a virtual machine construct, and management groups organise subscriptions.

Correct answer: A — The user sees the first words sooner, improving perceived responsiveness

Streaming is a user experience improvement: time to first token drops and long answers render progressively. It does not change token billing, does not improve accuracy, and does not remove the need for content filtering.

Correct answer: A — Revise the system prompt to instruct the model to use all supplied context and increase the output token limit

If retrieval is returning the right content, the gap is in generation, so the prompt must instruct the model to synthesise across all passages and the response must have room to include them. Fewer results removes information, deleting the index removes grounding, and maximum temperature reduces faithfulness.

Correct answer: A — Input tokens are billed, so a large context increases the cost and latency of every call

Generative model pricing counts both input and output tokens, so repeatedly sending a large document is billed every time and also adds processing latency. The model does not retain the document between calls, long input does not disable anything, and request size directly affects cost.

Correct answer: A — A text-to-image model deployed from the model catalog

Generating an image from a written description requires a text-to-image generative model, which is deployed from the catalog like any other model. Speech, translation, and document extraction address entirely different modalities.

Correct answer: A — Log request and response metadata, token usage, latency, and content filter events to Azure Monitor

Structured telemetry on usage, latency, and safety events is what makes regressions and cost surprises visible before users are affected. Waiting for complaints, disabling diagnostics, and monthly manual checks all leave the application effectively unmonitored.

Correct answer: A — Fairness

Fairness is the principle about avoiding systematic disadvantage to particular groups. Reliability and safety concerns consistent, safe operation, transparency concerns understandability, and privacy and security concerns protecting data.

Correct answer: A, B — The output must follow a very specific format that prompts do not enforce reliably · A large, high-quality labelled dataset for the task already exists

Fine-tuning is worth its cost when behaviour must be baked in and there is enough labelled data to teach it. Rapidly changing facts belong in retrieval, avoiding training cost argues for prompting, and five examples is a few-shot prompt rather than a training set.

Correct answer: A, B — The model and version deployed to each endpoint · Evaluation results and the prompts used to produce them

Knowing which model version is serving traffic and holding the evaluation evidence behind that decision is what makes an AI deployment auditable. Front-end styling, user browsing habits, and developer hardware have no bearing on the audit trail.

Correct answer: A — Binary classification

Choosing between exactly two labels is binary classification. Regression predicts numbers, forecasting predicts future values in a time-ordered sequence, and clustering has no labels at all.

Correct answer: A — Azure AI Content Safety filters on the deployment

Content Safety evaluates both input and output across harm categories and severity levels and blocks content that exceeds the configured thresholds. Firewalls and network security groups filter network traffic, and Key Vault policies control access to secrets.

Correct answer: A — Hybrid search with semantic ranking

Hybrid search runs both keyword and vector queries and fuses the results, and semantic ranking then reorders them by relevance. Geo-replication is about availability, indexer scheduling controls ingestion timing, and synonym maps expand query terms without combining retrieval modes.

Correct answer: A — Few-shot prompting

Including a small number of demonstrations in the prompt is few-shot prompting and it steers format and style without changing the model. Zero-shot supplies no examples, fine-tuning updates model weights, and chunking splits documents for indexing.

Correct answer: A — The model may produce plausible but factually incorrect statements

Hallucination, meaning confident output that is not true, is the headline risk for a public-facing generative assistant. Models do respond, they accept system instructions, and their output varies with both the prompt and sampling settings.

Correct answer: A — Speech to text

Speech to text transcribes audio into written words. Text to speech is the reverse, speech translation converts speech from one language to another and is a different service capability, and language detection identifies the language of written text.

Correct answer: A — They set persistent role, tone, and behavioural rules for the conversation

The system message carries standing instructions about the assistant's role and constraints, applied to every turn. History is application state, authentication uses Entra ID or keys, and region is a deployment property.