Question 1

Fabrikam is building an internal AI assistant that can answer questions over sensitive finance and HR data. The CISO wants to avoid a situation where any authenticated user can see all content once inside the app.Fill in the blank:From a security perspective, the AI solution should ______.

Options :

A : rely on network location instead of user identity

B : use strong identity, least privilege, and encryption

C : disable all logging to avoid capturing sensitive data

D : give broad admin roles to speed up delivery

Answer: B

Question 2

An AI transformation lead tells executives:“When we use Azure OpenAI or enterprise Copilot experiences, customer prompts and completions are not used to train the foundation models by default, and our data remains isolated to our tenant under Microsoft’s enterprise data privacy commitments.”Is this statement correct?

Options :

A : True

B : False

Answer: A

Question 3

An AI transformation lead notices that their Azure OpenAI–based assistant often jumps straight to a recommendation without explaining why. For risk and compliance reasons, the business needs the model to show its intermediate reasoning steps so humans can review the logic.Which prompt engineering change best addresses this requirement?

Options :

A : Add instructions like “Think step by step and show your reasoning before the final answer”

B : Only increase the maximum tokens allowed for the response

C : Remove all system instructions and rely on the model defaults

D : Ask users to add extra questions manually when they want more detail

Answer: A

Question 4

Tailwind Traders is piloting a generative AI assistant for internal policy questions. Early feedback is that answers are fluent but sometimes off-topic or missing critical rules. You explain that improving the prompts will likely fix much of this without changing the model.To improve answer quality with minimal architecture changes, the most effective next step is to focus prompt engineering on ______.

Options :

A : Adding more GPUs to the cluster

B : Rewriting prompts with clear roles and constraints

C : Migrating the app to a different Azure region

D : Shortening all responses to one sentence

Answer: B

Question 5

A telco is training a churn model. Midway through, you discover that:10% of labels are wrong due to a bug in how “churn” was recordedSome important features have inconsistent units across systemsMany recent records are missing key fieldsWhat is the most accurate description of how this data quality will affect the model?

Options :

A : The model will be largely unaffected because modern algorithms are explicitly designed to ignore label noise and unit inconsistencies during training.

B : Poor labels, inconsistent units, and missing values can distort the patterns the model learns, reducing accuracy and making outputs harder to trust, especially for borderline customers.

C : Only missing values matter; label errors and unit issues are irrelevant because the optimisation process doesn’t look at them directly.

D : Data quality problems affect only training speed, not model behaviour, so your main risk is longer training time rather than worse predictions.

Answer: B