AI-901 Scenario Practice: Read Facts and Trace Results
Work through AI-901 requirements, client behavior, and extraction evidence without relying on service-name shortcuts.
Use IT Mastery to practice choosing an answer from the requirements and visible evidence. These short walkthroughs show how to review a decision; they are not official Microsoft exam questions.
Read the task before choosing a capability
Identify four things: the input, the required output, the constraint, and the step being tested. A question about why a client displays no text is different from one about which model to choose, even if both mention the same Foundry deployment.
For code or configuration, trace the values that are actually passed. For a table or result, use its evidence. A product name may identify the setting, but it rarely proves that an option satisfies every requirement.
Walkthrough: a good average hides a fairness problem
A hiring-support model identifies most qualified applicants overall, but its recall is substantially lower for one demographic group. The evaluation groups are similarly sized, and an independent review established which applicants were qualified.
The deciding evidence is the group-level difference in missed qualified applicants. Investigate the errors and relevant data representation, choose a mitigation supported by that investigation, and retest group-level performance. A larger model, removal of demographic input fields, or a common threshold change does not establish the cause or prove that the disparity is resolved.
If the prompt instead described people being unable to use the interface, accessible interaction would become the immediate concern. Match the responsible AI principle and control to the observed problem.
Walkthrough: the agent’s tool is not used
A support team tests a saved agent with refund-policy instructions and an order-lookup tool. The application instead sends the customer’s question directly to the model deployment used by that agent. The answer contains no current order information.
The decisive fact is the application’s invocation target. A model deployment call does not inherit a saved agent’s configuration. First compare the request with the supported saved-agent invocation path. Increasing the output length would not apply the absent tool configuration.
Do not infer that changing the target guarantees a successful refund. The tool still needs the required input and access, and any consequential action still needs the application’s authorization controls.
Walkthrough: a transcription is not an answer
A demonstration accepts a spoken question about a timetable and displays the recognized words. The requirement is to answer the question aloud, using the timetable provided to the application.
The displayed transcript shows that recognition worked. It does not show that the question was answered. Follow the remaining path: provide the request and timetable to a suitable model or agent, then deliver the spoken response through the selected audio-output workflow.
Azure Speech Voice Live can manage conversational audio. Its internal path depends on the chosen model or agent configuration: do not assume every voice interaction passes a separate transcript into the reasoning model. The test is whether the complete application delivers the required behavior, not whether it contains a service with “Speech” in its name.
Walkthrough: values lose their relationships
An application needs one record per invoice line, with the description, quantity, and unit price kept together. The source contains these two lines:
| Description | Quantity | Unit price |
|---|---|---|
| Notebook | Missing | 12.50 |
| Pen | 2 | 2.00 |
The application flattens the extracted values into separate lists and omits the missing quantity. This is illustrative application data, not a Content Understanding response schema.
descriptions = [
"Notebook", "Pen"
]
quantities = [2]
unit_prices = [12.50, 2.00]
for row in zip(
descriptions,
quantities,
unit_prices
):
print(row)
Before opening the diagnosis, trace what this code prints. Which source line owns the surviving quantity, and how many records reach the application?
Check the mapping diagnosis
The code prints ('Notebook', 2, 12.5) and stops. The quantity belongs to the pen, but pairing values by list position assigns it to the notebook. Python's zip stops at the shortest list, so the pen record disappears entirely.
Preserve each source line's fields together before flattening loses their relationships. A collection of line-item objects can retain both records and represent the missing quantity as null when the application contract permits it:
[
{
"description": "Notebook",
"quantity": null,
"unitPrice": 12.50
},
{
"description": "Pen",
"quantity": 2,
"unitPrice": 2.00
}
]
Equal list lengths alone would not prove that the values belong together. Keep the source-line association through extraction and mapping, and validate the resulting records against the source.
More OCR alone does not define those relationships. Inspect source quality, the analyzer’s field schema, and the final mapping. If a mandatory value is unreadable, mark it for the required review or recovery step instead of inventing a value. See Microsoft’s Content Understanding overview for the extraction workflow.
Compare plausible alternatives
For each option, state the requirement it satisfies and any requirement it fails. Extra capability is not automatically a defect: opinion mining, for example, adds target-level detail to sentiment analysis without removing overall sentiment results. A rejection needs a relevant limitation, missing condition, or conflict with the task.
For “Select TWO,” judge every choice and verify that the selected pair provides the two required contributions. Do not assume two options conflict merely because they concern the same service.
Review a miss so it transfers
Write one sentence describing the deciding fact, then change that fact. If the application already invoked the saved agent, what evidence would you inspect next? If every invoice line were guaranteed complete, which mapping concern would remain?
This is a stronger review habit than memorizing a phrase such as “use an agent” or “choose structured output.” Use the current blueprint to keep the review at AI-901 depth.