CertArea
All Microsoft exams

Free AI-102 practice test: 30 questions with answers

30 original questions for Designing and Implementing a Microsoft Azure AI Solution, drawn from every topic area. Answer each one to see whether you got it right and why. No sign-up.

Free questions
30
Full bank
112 questions

AI-102 exam guide: topics, cost and format

All 30 AI-102 questions with answers

The same questions as the test above, grouped by exam topic. Open any question to see the answer and explanation.

Implement generative AI solutions (5 questions)

  1. Question 1

    What is the purpose of retrieval-augmented generation (RAG)?

    1. ATo train a new foundation model from scratch using only the organisation's internal documents and no base model
    2. BTo remove the need for any prompt when the model answers questions from users of the application
    3. CTo ground a model's answers in retrieved content from the organisation's own data sources
    4. DTo convert all answers into images so that the model can respond visually to users without any text
    Show answer and explanation
    Correct answer: C

    RAG grounds answers in retrieved organisational content. It does not remove prompts, train a foundation model from scratch or convert answers to images.

  2. Question 2

    Which Azure resource hosts a Microsoft Foundry project that holds the models and tools for a generative AI solution?

    1. AA virtual machine that runs the operating system for the model and is not linked to any project at all
    2. BA storage account with no compute, which holds only the raw files that the team uploads for later use
    3. CA Foundry hub and project, which group the models, data connections and tools for the solution
    4. DA DNS zone, which maps the names of the AI endpoints to their IP addresses on the public network
    Show answer and explanation
    Correct answer: C

    A Foundry hub and project group models, connections and tools. A storage account, a bare virtual machine and a DNS zone do not organise the solution.

  3. Question 3

    What is a prompt flow used for in Azure AI development?

    1. AStoring images in a database so that they can be retrieved by a computer vision model later in the process
    2. BSending email notifications to the users of the solution whenever a model returns a response to them
    3. CBuilding, testing and evaluating a flow of prompts, tools and model calls that forms the solution logic
    4. DMonitoring the physical temperature of the servers that host the model across the data centre network
    Show answer and explanation
    Correct answer: C

    Prompt flow builds, tests and evaluates chains of prompts, tools and model calls. Email, image storage and server temperature are unrelated.

  4. Question 4

    Which evaluation measure checks whether a RAG answer is supported by the retrieved context?

    1. AFrame rate, which measures how many frames per second a video stream delivers to the viewer's screen
    2. BGroundedness, which measures whether the answer is supported by the source content provided to it
    3. CWord count, which measures the length of each answer in words without considering its supporting content
    4. DContrast ratio, which measures the visual difference between the text and background colours on the page
    Show answer and explanation
    Correct answer: B

    Groundedness measures support from the retrieved context. Frame rate, word count and contrast ratio are unrelated.

  5. Question 5

    Which parameter controls how random a generated response is in a chat model?

    1. AThe resource group name, which determines the region in which the model's responses are generated for users
    2. BThe maximum response length, which sets the number of tokens the model may use in a single response in total
    3. CTemperature, which sets the randomness of token selection, with lower values giving more predictable output
    4. DThe number of deployments, which sets how many identical copies of the model answer requests in parallel
    Show answer and explanation
    Correct answer: C

    Temperature controls randomness in token selection. Maximum length, resource group and deployment count do not control randomness.

Plan and manage an Azure AI solution (5 questions)

  1. Question 6

    Which Azure AI service suits a solution that must extract key phrases and detect sentiment from customer reviews?

    1. AAzure AI Vision, which analyses images to detect objects, tags and text found inside pictures
    2. BAzure AI Language, which provides text analytics features such as key phrase extraction and sentiment
    3. CAzure AI Speech, which converts spoken audio into text and synthesises speech from written text
    4. DAzure AI Translator, which translates text between languages but does not analyse the meaning of the text
    Show answer and explanation
    Correct answer: B

    Azure AI Language provides key phrase extraction and sentiment analysis for text. Vision, Speech and Translator address different data types or tasks.

  2. Question 7

    Which deployment option lets a team run an Azure AI container on its own edge device?

    1. AA deployment that stores the model weights in a spreadsheet so that the device can read them when needed
    2. BA serverless deployment that exists only in the public cloud and cannot be run on any local hardware at all
    3. CA deployment that sends every request to a single shared endpoint in one region with no local copy kept
    4. DA containerised deployment of the service that is pulled to the device and run there
    Show answer and explanation
    Correct answer: D

    Many Azure AI services can be deployed as containers for local or edge use. The other options describe cloud-only or non-viable approaches.

  3. Question 8

    Why should an AI resource key be stored in Azure Key Vault rather than in application code?

    1. ASo that the key can be printed on the application's user interface for the users to copy manually when needed
    2. BSo that the key runs faster when the application calls the AI service during a busy period of the day
    3. CSo that the key is shared with every developer in the company by default without any further access control
    4. DSo that the key can be rotated and access to it audited without redeploying the application code
    Show answer and explanation
    Correct answer: D

    Key Vault allows rotation and auditing without changing code. Speed, sharing and display are not reasons to keep keys there.

  4. Question 9

    Which authentication approach avoids storing a static key in an application that runs in Azure?

    1. AManaged identity with Microsoft Entra ID role-based access to the AI resource
    2. BEmbedding the resource key as a constant string inside the client application's source code
    3. CSharing one administrator password between all application instances in the deployment environment
    4. DDisabling authentication on the AI endpoint so that every caller can use it without any credential
    Show answer and explanation
    Correct answer: A

    Managed identity with Entra ID roles avoids static keys. Embedded keys, shared passwords and disabled authentication are unsafe.

  5. Question 10

    What is the main purpose of a content filter in an Azure OpenAI deployment?

    1. ATo detect and block harmful input or output categories according to configured severity levels
    2. BTo compress the model's output so that each response uses fewer tokens and costs less per request
    3. CTo translate every response into a second language before it is returned to the user of the application
    4. DTo cache responses so that identical prompts return the stored answer without calling the model again
    Show answer and explanation
    Correct answer: A

    Content filters detect and block harmful categories at configured severity. Compression, translation and caching are unrelated.

Implement knowledge mining and information extraction solutions (5 questions)

  1. Question 11

    Which THREE components are part of an Azure AI Search indexing pipeline? (Choose three.)

    1. AA data source that connects the indexer to the content to be indexed
    2. BAn indexer that runs on a schedule or on demand to load the index from the data source
    3. CA skillset that applies AI enrichment steps such as entity recognition to the content during indexing
    4. DA tracker that monitors the colour theme of the search results page seen by the end user of the site
    5. EA gateway that encrypts every query result for the mobile applications that read the index
    Show answer and explanation
    Correct answer: A, B, C

    Data sources, indexers and skillsets form the indexing pipeline. A colour tracker and a result-encrypting gateway are not part of Azure AI Search.

  2. Question 12

    Which TWO are valid Azure AI Search index features? (Choose two.)

    1. ASemantic ranking that reorders results by meaning using a language model
    2. BVector search over embeddings stored in a vector field of the index
    3. CAutomatic conversion of every index into a relational database with foreign keys for joins
    4. DBuilt-in playback of audio files stored in the index for listening during a search session
    5. EDirect printing of search results to a physical printer connected to the search service
    Show answer and explanation
    Correct answer: A, B

    Semantic ranking and vector search are index query features. Relational conversion, audio playback and printing are not Azure AI Search features.

  3. Question 13

    Which Azure AI Search component is used to pull data from a source into an index on a schedule?

    1. AA scoring profile, which ranks search results by relevance once the index already holds the content
    2. BA data source only, which stores the data but never moves it into an index without a separate process
    3. CA synonym map, which lists alternative words for search terms and does not load any content into the index
    4. DAn indexer, which reads from a data source and writes documents into the search index
    Show answer and explanation
    Correct answer: D

    An indexer reads from a data source and populates the index. A data source alone, a synonym map and a scoring profile do not load content.

  4. Question 14

    What is a knowledge store in Azure AI Search?

    1. AA storage destination that receives enriched content from a skillset as files, objects or tables
    2. BA cache of recent queries that the search service keeps in memory to speed up common searches for users
    3. CA list of approved users who may read the search index, maintained by the security team for each index
    4. DA catalogue of the vendors that supply the search service's underlying hardware in each Azure region
    Show answer and explanation
    Correct answer: A

    A knowledge store receives enriched content as files, objects or tables. Query caches, user lists and vendor catalogues are different things.

  5. Question 15

    Which Document Intelligence model extracts fields from common forms such as invoices without custom training?

    1. AA prebuilt model, such as the invoice model, which extracts standard fields from that document type
    2. BA custom neural model trained from scratch on a single unlabelled sample that the user uploads for the job
    3. CA composed model that combines a dozen unrelated document types into one model with no labelled data at all
    4. DA layout model that only detects the colours of the page background and not the text or fields on it
    Show answer and explanation
    Correct answer: A

    Prebuilt models extract standard fields from common document types without custom training. Custom, composed and colour-only options do not fit that requirement.

Implement computer vision solutions (4 questions)

  1. Question 16

    Which Azure AI capability reads printed text from an image?

    1. ASpeech synthesis, which converts written text into a spoken audio file for the user to listen to
    2. BOptical character recognition in the Azure AI Vision read feature
    3. CLanguage detection, which identifies the language of text that has already been typed into the application
    4. DAnomaly detection, which finds unusual values in a numeric time series stored in a database table
    Show answer and explanation
    Correct answer: B

    The Vision read feature performs optical character recognition. Speech synthesis, language detection and anomaly detection do different jobs.

  2. Question 17

    Which task is an object detection model designed to perform?

    1. ALocate and label each instance of objects within an image, usually with bounding boxes
    2. BAssign a single label to a whole image, such as indicating whether it shows a cat or a dog overall
    3. CConvert an image into a vector of numbers that can be compared with other images in a similarity search
    4. DTranslate the text in an image into another language for display on the screen of the user's device
    Show answer and explanation
    Correct answer: A

    Object detection locates and labels instances with bounding boxes. Whole-image classification, embedding and translation are different tasks.

  3. Question 18

    What is the purpose of a labelled dataset when training a custom vision model?

    1. AIt replaces the need for any image at all, since the labels alone are enough to train the model
    2. BIt provides the examples that the model learns from, with the correct classes or boxes marked
    3. CIt stores the model's predictions so that they can be shown to customers as the final product output
    4. DIt defines the colours that the interface uses when the model is displayed to the operator of the system
    Show answer and explanation
    Correct answer: B

    Labelled examples are what a custom vision model learns from. Labels alone cannot replace images, and predictions and interface colours are not training data.

  4. Question 19

    Which Azure AI feature detects the presence and movement of people in a video stream?

    1. ADocument Intelligence, which extracts fields from forms and invoices uploaded to the document storage
    2. BSpeech-to-text, which converts the audio track of a video into a transcript that can be searched by the user
    3. CKey phrase extraction, which finds the main topics in the text of a document uploaded to the solution
    4. DSpatial Analysis, which detects presence and movement of people in video from camera feeds
    Show answer and explanation
    Correct answer: D

    Spatial Analysis detects presence and movement of people in video. Speech-to-text, key phrases and form extraction do not.

Implement natural language processing solutions (4 questions)

  1. Question 20

    Which Azure AI Language feature identifies people, places and organisations in text?

    1. ANamed entity recognition, which detects and categorises entities such as people and organisations
    2. BSentiment analysis, which scores the tone of a piece of text as positive, negative, neutral or mixed overall
    3. CLanguage detection, which identifies which language the text is written in, not what the entities are
    4. DSpeech recognition, which converts audio into text but does not analyse the entities that the text contains
    Show answer and explanation
    Correct answer: A

    Named entity recognition detects and categorises entities. Sentiment, language detection and speech recognition do different tasks.

  2. Question 21

    Which feature detects personal information such as phone numbers in text?

    1. APersonally identifiable information detection in Azure AI Language
    2. BImage tagging in Azure AI Vision, which labels the objects that appear inside the photographs
    3. CAnomaly detection in Azure AI Metrics Advisor, which finds unusual values in a time series of data
    4. DDocument translation in Azure AI Translator, which converts entire files between languages
    Show answer and explanation
    Correct answer: A

    PII detection in Azure AI Language finds personal information in text. The other services address images, time series and translation.

  3. Question 22

    Which component of the Speech service converts written text into spoken audio?

    1. ASpeech-to-text, which transcribes spoken audio into written text for the application to process further
    2. BText-to-speech, which synthesises audio from text using a chosen voice
    3. CIntent recognition, which maps spoken or typed phrases to the actions defined in a language model
    4. DKeyword recognition, which listens for a specific wake word in audio and does not produce any speech output
    Show answer and explanation
    Correct answer: B

    Text-to-speech synthesises audio from text. The other components convert speech to text or recognise intents and keywords.

  4. Question 23

    What does SSML allow a developer to control in speech synthesis?

    1. AThe language model used to generate the text that is then spoken aloud by the synthesis service
    2. BThe number of users who can listen to the same audio file at the same time on the platform
    3. CPronunciation, pauses, speaking rate and emphasis through markup in the input text
    4. DThe physical volume of the speaker in the room where the application's audio output is being played
    Show answer and explanation
    Correct answer: C

    SSML controls pronunciation, pauses, rate and emphasis. User counts, the text model and physical speaker volume are not controlled through SSML.

Implement an agentic solution (4 questions)

  1. Question 24

    Which component manages the runs and threads of an agent in the Microsoft Foundry Agent Service?

    1. AThe billing dashboard, which displays the charge for each agent run in the subscription by resource group
    2. BThe DNS record of the endpoint, which maps the agent's public name to the IP address of its hosting server
    3. CThe agent's thread and run objects, which hold the conversation state and the steps taken by the agent
    4. DThe colour theme of the agent's interface, which the service stores as a configuration file for each user
    Show answer and explanation
    Correct answer: C

    Threads and runs hold conversation state and agent steps. DNS records, billing dashboards and themes are not agent run components.

  2. Question 25

    Why would a multi-agent solution use an orchestrator agent?

    1. ATo replace the user interface of the solution with a single static page that lists the answers in a table
    2. BTo remove the need for any specialised agent, since the orchestrator can perform every task alone without help
    3. CTo store the conversation history for all agents on a single disk so that the history is never lost in an outage
    4. DTo route tasks to specialised agents and combine their results into one coherent response
    Show answer and explanation
    Correct answer: D

    An orchestrator routes tasks and combines results from specialised agents. It does not remove the need for them, and storage and UI are separate concerns.

  3. Question 26

    What should happen before an autonomous agent performs an irreversible action such as deleting a record?

    1. AThe agent runs the action at once and reports the result afterwards, since the action has already been decided by the model
    2. BA human approval step, or a policy check that the action is allowed, before it runs
    3. CThe action is recorded in a log only and is never reviewed by anyone in the organisation during the normal process
    4. DThe action is delegated to the user's browser so that the user's session carries out the deletion on their behalf
    Show answer and explanation
    Correct answer: B

    Irreversible actions should be approved or policy-checked before execution. Running first, logging only, and browser delegation all remove that control.

  4. Question 27

    Which TWO practices help keep an agent's behaviour safe and predictable? (Choose two.)

    1. ASetting clear instructions and boundaries for the agent's role and the actions it may take
    2. BLetting the agent use any credential it finds in the environment so that it can complete tasks without interruption
    3. CLogging each tool call and its result so that the agent's actions can be reviewed afterwards
    4. DAllowing the agent to run without a time limit or step limit so that long tasks are never stopped
    5. EChanging the agent's instructions during each conversation based on what the user asks for in the moment
    Show answer and explanation
    Correct answer: A, C

    Clear boundaries and logged tool calls keep an agent predictable and reviewable. Using any credential found, unlimited runs and per-conversation instruction changes make behaviour unsafe.

Implement agentic solutions (3 questions)

  1. Question 28

    What is the role of a tool in an AI agent?

    1. AA user interface element that displays the agent's name and avatar on the screen of the chat window only
    2. BAn action or function the agent can call to retrieve information or perform a task
    3. CA licence that allows the agent to be deployed in a particular region of the cloud for the customer
    4. DA log file that records every token the agent generated during the whole of a conversation with the user
    Show answer and explanation
    Correct answer: B

    Tools are actions or functions an agent can call. Interface elements, licences and log files are different things.

  2. Question 29

    Why should an agent's tool calls be restricted to the minimum set it needs?

    1. ATo reduce the number of words in the agent's name, which makes the agent easier to find in the directory
    2. BTo make the agent respond more slowly, so that the user has time to read each answer before the next
    3. CTo reduce the risk that a manipulated or faulty agent performs an action it should not perform
    4. DTo ensure the agent can call every tool in the organisation whenever it chooses, without any approval
    Show answer and explanation
    Correct answer: C

    Restricting tools limits what a manipulated or faulty agent can do. Slowing responses, naming and unrestricted access do not serve that purpose.

  3. Question 30

    Which practice helps test an agent before deployment?

    1. AEvaluating it against a set of representative tasks and checking its tool use and outputs
    2. BDeploying it to all users at once and reading the complaints as the only form of testing before any change
    3. CRemoving all the tools so that the test is simpler and the agent has no way to act during the evaluation
    4. DAsking the agent to test itself without any reference answers or expected outcomes against which to compare
    Show answer and explanation
    Correct answer: A

    Evaluating against representative tasks checks tool use and outputs before rollout. Deploying first, removing tools and self-testing without references are weak practices.