AI-901 Practice Questions — Free

Question 1 of 25

Which capability identifies that two differently worded sentences mean the same thing?

• A.Speech synthesis rendering both of the sentences aloud
• B.Semantic understanding using embeddings or a language modelCorrect
• C.Semantic segmentation of an image of the two sentences
• D.Optical character recognition applied to both sentences

Question 2 of 25

An organisation is deciding between fine-tuning a model and providing context in prompts. Which statement is correct?

• A.Prompt context permanently updates the deployed model
• B.Fine-tuning only changes the prompt template that is used
• C.Prompt context adapts per request; fine-tuning changes the weightsCorrect
• D.Both approaches modify the underlying model identically

Question 3 of 25

A logistics firm wants to read licence plate numbers from camera images. Which capability is this?

• A.Optical character recognition within computer visionCorrect
• B.Speech recognition applied to the camera feed
• C.Image generation from the camera metadata
• D.Text summarization of the captured images

Question 4 of 25 · multiple response

Which of the following are examples of computer vision capabilities? (Select TWO)

• A.Detecting and locating objects within a photographCorrect
• B.Converting a spoken sentence into written text
• C.Producing an audio narration from an article
• D.Reading printed text from a scanned pageCorrect

Question 5 of 25

A team must pick a model for real-time chat where responses must start quickly. Which factor matters most?

• A.The date the model card was last updated
• B.The number of Azure regions offering the model
• C.Latency and support for streaming responsesCorrect
• D.The size of the dataset used to train the model

Question 6 of 25

Which factor limits how much conversation history an application can send?

• A.The length of the model deployment's display name
• B.The number of Azure subscriptions in the tenant
• C.The model's context window measured in tokensCorrect
• D.The physical screen size of the client's device

Question 7 of 25

A team must explain why the same AI feature costs more some months. What is the likely cause?

• A.The model's parameter count is changed on a monthly basis
• B.Request volume and tokens consumed vary with actual usageCorrect
• C.Azure automatically changes the region every single month
• D.The content filter applies a charge per blocked response

Question 8 of 25

A generative AI feature must not reveal one customer's data to another. Which principle applies?

• A.Privacy and securityCorrect
• B.Fairness across groups
• C.Transparency to users
• D.Inclusiveness of design

Question 9 of 25

A business wants AI to draft responses that staff then edit and send. Which framing is accurate?

• A.The AI is fully responsible for anything that it drafts
• B.Staff review is unnecessary because the model is accurate
• C.The AI assists a human who remains responsible for the contentCorrect
• D.The model becomes accountable once the staff approve it

Question 10 of 25

Which describes the role of evaluation data in an AI project?

• A.It sets the temperature parameter used in production
• B.It is used to train the model's underlying parameters
• C.It measures whether the system meets quality before releaseCorrect
• D.It determines the Azure region used for the deployment

Question 11 of 25

An AI solution must continue working correctly when it receives unusual or malformed input. Which principle applies?

• A.Inclusiveness of the design
• B.Fairness across groups
• C.Transparency to end users
• D.Reliability and safetyCorrect

Question 12 of 25

A company wants to extract product details from photographs taken by field staff. Which capability applies?

• A.A text-only model reading the image file names
• B.Content Understanding applied to the imagesCorrect
• C.A relational database query run over the images
• D.Speech synthesis applied to the photograph files

Question 13 of 25

A team must process ten thousand invoices per night with extraction. Which consideration matters most for the application design?

• A.The number of developers assigned to the project team
• B.Batching, throughput limits, and retry handling for the callsCorrect
• C.The colour scheme of the extraction results dashboard
• D.The length of the file names used for the invoices

Question 14 of 25

What distinguishes an agent from a simple chat completion call in Foundry?

• A.An agent runs only in the portal and never from code
• B.An agent cannot be given a system prompt or instructions
• C.An agent can use tools and take multiple steps to complete a taskCorrect
• D.An agent always runs on a larger model than chat completions

Question 15 of 25

An application must support users who ask questions in a language the business does not support. What should it do?

• A.Detect the language and respond or explain the limitationCorrect
• B.Return an unexplained error with no guidance for the user
• C.Silently drop the request without any response at all
• D.Attempt an answer in a language nobody has verified

Question 16 of 25

An application takes spoken customer questions and must answer them aloud. Which Foundry Tools service handles the final step?

• A.Azure Speech recognition transcribing the generated answer
• B.Azure Speech synthesis converting the generated answer to audioCorrect
• C.An image-generation model visualising the spoken answer
• D.Content Understanding extracting fields from the answer text

Question 17 of 25

An agent needs to look up live order status from an internal system. What must be configured?

• A.A second copy of the model deployment for lookups
• B.A higher temperature so the agent guesses the status
• C.A larger context window configured on the agent's model
• D.A tool or action the agent can call against the order systemCorrect

Question 18 of 25

A developer must handle a model response that is longer than the application's display area. What is the appropriate approach?

• A.Increase the temperature in order to shorten the answers
• B.Send the request twice and display the shorter of the two
• C.Set an appropriate token limit and instruct the desired lengthCorrect
• D.Truncate the response mid-sentence with no indication to users

Question 19 of 25

A developer must decide where to put a rule that the assistant should never give legal advice. Where does it belong?

• A.In the description field of the Azure resource group
• B.In an email circulated to the development team
• C.In a comment within the application's source code files
• D.In the system prompt, reinforced by output validation in codeCorrect

Question 20 of 25

An application's responses must never include a competitor's name. Which approach is most reliable?

• A.Reduce the maximum output tokens allowed per response
• B.Rely on the model to infer the restriction on its own
• C.Instruct the model in the system prompt and validate outputCorrect
• D.Increase the temperature to avoid repeating brand names

Question 21 of 25

An application must summarize documents and highlight risks for a legal team. Which design fits?

• A.Replace the legal team entirely with the AI feature
• B.Generate the summary with no defined risk criteria
• C.Generate with explicit risk criteria and mark for reviewCorrect
• D.Publish the generated summary directly as legal advice

Question 22 of 25

A team must hand a Foundry application over to an operations team. What should the handover include?

• A.Only the display name of the model deployment used
• B.Only the names of the developers who originally built it
• C.Only the location of the source code repository
• D.Monitoring, alerting, known failure modes, and escalationCorrect

Question 23 of 25

An application must analyse the sentiment of reviews in eight languages. What should the team verify?

• A.That the model has the largest available context window
• B.That every review is manually translated to English first
• C.That the chosen capability supports all eight languagesCorrect
• D.That all of the reviews originate from a single country

Question 24 of 25 · multiple response

Which of the following are valid considerations when choosing a deployment region for a Foundry workload? (Select TWO)

• A.Model availability and latency to the usersCorrect
• B.Data residency requirements for the organisationCorrect
• C.The number of characters in the region's name
• D.The alphabetical order of the region names

Question 25 of 25

A developer must decide how to handle a request the model answers slowly. Which approach fits?

• A.Send the identical request repeatedly in parallel
• B.Set a timeout, stream where possible, and show progressCorrect
• C.Cancel the request and present the user a blank screen
• D.Wait indefinitely with no feedback given to the user