Roughly 65% of large enterprises are now actively testing or deploying multimodal AI in production, according to McKinsey’s latest State of AI report. That shift is exactly why AI-102 computer vision solutions skills carry real weight this year. Teams increasingly need to decide between dedicated vision services and general-purpose multimodal models like GPT-4o for the same task, and the exam tests that judgment directly.
This second set in our AI-102 series digs into that decision-making. It also covers a genuinely current curveball: Microsoft recently announced the retirement of Azure Custom Vision. Therefore, the right answer for anyone planning a new custom image model today has actually changed, and question 13 below walks through exactly what that means in practice.
This set covers the “Implement computer vision solutions” domain, worth 10-15% of the AI-102 exam. Specifically, you’ll work through OCR versus document extraction, content moderation, face detection, video insight extraction, spatial analysis, and the classification-versus-object-detection choice inside Custom Vision. For example, several scenarios below test something that wasn’t even true a year ago, so don’t lean on older study material for those.
Each scenario mirrors the real exam’s format. A specific requirement rules out two or three otherwise-reasonable options, leaving one clearly correct answer. Read the scenario, choose your answer, then reveal the explanation to see exactly why the distractors fail. If you missed the first set, start with AI-102 Questions 1-10 before working through these AI-102 computer vision solutions questions, since a few concepts carry over.
One more thing worth flagging: Azure’s computer vision lineup looks deceptively similar at first glance. Vision, Custom Vision, Video Indexer, and Content Understanding all touch images or video in some way. However, each one solves a narrower problem than it might first appear, and the exam leans heavily on that narrowness. Getting comfortable with where one service’s job ends and another’s begins is, honestly, most of what this domain tests.
Question 11: Extracting Text From Photographed Whiteboards
Fabrikam’s field technicians photograph whiteboards after every planning meeting. The photos mix printed sticky notes with messy handwriting. There’s no fixed layout or document type to anchor against.
A) Azure AI Document Intelligence prebuilt invoice model
B) Azure Vision in Foundry Tools OCR (Read feature)
C) Azure AI Custom Vision object detection
D) Azure Content Safety image moderation
👁 Reveal Answer
Correct Answer: B
Explanation: Azure Vision’s OCR capability extracts both printed and handwritten text from general, unstructured images. Therefore, it fits a whiteboard photo with no fixed layout perfectly. Document Intelligence’s prebuilt models expect a known document type instead, such as an invoice or receipt, with predictable fields. That’s not what a whiteboard photo offers. Custom Vision detects objects or classifies images; it doesn’t read text at all. Content Safety moderates for harmful content, not text extraction. So, it doesn’t apply here either.
Question 12: Moderating User-Uploaded Product Photos
Tailwind Traders lets shoppers upload their own photos alongside product reviews. Some uploads occasionally contain inappropriate or unsafe content. As a result, the team needs an automated way to flag it before publication.
A) Azure Vision in Foundry Tools image analysis
B) Azure AI Content Safety
C) Azure AI Video Indexer
D) Azure AI Custom Vision classification
👁 Reveal Answer
Correct Answer: B
Explanation: Content moderation is explicitly Content Safety’s job within Microsoft Foundry, not Azure Vision’s. Official guidance is direct on this point: don’t use Azure Vision to moderate content. Video Indexer analyzes video and audio content instead, not standalone product photos. So, it’s the wrong tool for this job. Custom Vision classifies images against labels you define yourself. As a result, it isn’t built for general-purpose harmful-content detection out of the box.
Question 13: Planning a Brand-New Custom Image Model Today
Northwind Traders is kicking off a brand-new defect-classification project this month. The architecture team wants to follow Microsoft’s current guidance rather than an outdated tutorial. After all, they know Azure Custom Vision’s roadmap has changed recently.
A) Create a new Azure Custom Vision resource, since it’s still the standard recommendation
B) Use Azure Machine Learning AutoML or Content Understanding for new custom classification work
C) Use Azure Vision’s general image analysis without any custom training
D) Use Azure AI Video Indexer trained on labeled still images
👁 Reveal Answer
Correct Answer: B
Explanation: Microsoft has announced the planned retirement of Azure Custom Vision. Full support for existing customers continues only until September 2028. For new custom classification or object-detection work, Microsoft now points teams toward Azure Machine Learning AutoML instead. Alternatively, Content Understanding handles generative, schema-based classification across image, document, audio, and video content. Building on the older service for a brand-new project ignores that guidance. General image analysis without custom training can’t recognize project-specific defects, either. Video Indexer processes video and audio, not a still-image classification workload.
Question 14: Confirming Identity at a Secure Door
Woodgrove Bank wants to confirm an employee’s identity at a secure server room door. A camera mounted above the entrance compares the live image against an approved staff photo in real time.
A) Azure AI Video Indexer face identification on recorded footage
B) Azure Vision face detection and analysis capabilities
C) Azure AI Search semantic ranking
D) Azure Content Understanding document extraction
👁 Reveal Answer
Correct Answer: B
Explanation: Detecting, recognizing, and analyzing human faces in images is a core Azure Vision capability. It’s built for exactly this kind of real-time, single-image comparison. Video Indexer’s face identification works against recorded video content instead, not a live door-camera comparison against one reference photo. Semantic ranking is a search-relevance feature with nothing to do with faces. Content Understanding extracts structured fields from documents and media. However, it doesn’t perform identity comparison.
Question 15: Turning a Recorded Town Hall Into a Searchable Transcript
Litware recorded a two-hour town hall. The team needs a searchable transcript, speaker labels, and scene breaks, so employees can jump straight to the parts that matter to them.
A) Azure Vision in Foundry Tools image analysis
B) Azure AI Video Indexer
C) Azure AI Custom Vision object detection
D) Azure Vision Spatial Analysis
👁 Reveal Answer
Correct Answer: B
Explanation: Azure AI Video Indexer is purpose-built to extract deep insights from recorded or live video and audio. This includes transcription, speaker identification, and scene segmentation. Azure Vision’s image analysis processes still images instead, not a two-hour recording. Custom Vision detects objects in individual frames rather than producing a transcript. Spatial Analysis tracks the presence and movement of people in a camera feed. As a result, it doesn’t generate transcripts or speaker labels.
Question 16: Counting Shoppers in a Live Store Camera Feed
Continuing this set, Contoso wants to know how many shoppers linger in each aisle. The retailer already has ceiling cameras installed in every store, and it wants to reuse that existing hardware.
A) Azure AI Video Indexer scene segmentation on stored recordings
B) Azure Vision Spatial Analysis to detect presence and movement of people
C) Azure AI Custom Vision image classification
D) Azure AI Document Intelligence layout analysis
👁 Reveal Answer
Correct Answer: B
Explanation: Spatial Analysis is the Azure Vision feature specifically designed to detect the presence and movement of people from a live camera feed. It’s a distinct capability from Video Indexer. In contrast, Video Indexer focuses on insights from recorded or live content, like transcripts and scene changes, rather than real-time foot-traffic counting. Custom Vision classifies whole images against labels; it doesn’t track movement over time. Document Intelligence analyzes documents, which have nothing to do with a retail camera feed.
Question 17: Answering Open-Ended Questions About Any Photo
Your team is building a chat assistant that should answer open-ended, conversational questions about whatever photo a user happens to paste in. There’s no predefined list of visual features to check against.
A) Azure Vision in Foundry Tools with a fixed set of visual features
B) A GPT-4o class model through Azure OpenAI in Foundry Models
C) Azure AI Custom Vision trained on a narrow label set
D) Azure AI Video Indexer
👁 Reveal Answer
Correct Answer: B
Explanation: Official guidance recommends Azure OpenAI for broad, nonspecific image analysis and open-ended conversation. Specifically, multimodal models like GPT-4o can reason about an image without a predefined feature list. Azure Vision returns a fixed, structured set of visual features per request instead, which doesn’t stretch to free-form conversation. Custom Vision only recognizes the specific labels it was trained on. Video Indexer is built for video and audio content, too, not a single pasted photo.
Question 18: Extracting Structured Fields From Mixed Claim Media
Adventure Works processes insurance claims that arrive as both photos and short video walkarounds of vehicle damage. The team wants one schema-defined extraction pipeline that pulls fields like damage type and estimated cost from both media types.
A) Azure AI Document Intelligence prebuilt receipt model
B) Azure Content Understanding in Foundry Tools
C) Azure Vision image analysis alone
D) Azure AI Custom Vision object detection alone
👁 Reveal Answer
Correct Answer: B
Explanation: Content Understanding uses generative AI to extract structured, schema-defined fields across images, video, audio, and documents in a single pipeline. This matches the mixed-media claim scenario exactly. Document Intelligence’s prebuilt models target documents specifically instead, not video walkarounds. Vision image analysis alone returns generic visual features rather than a custom claim schema. Custom Vision object detection can locate damage in a photo. However, it has no video support and no structured field schema of its own.
Question 19: Locating Every Defective Part, Not Just Flagging the Image
Midway through this set, Fabrikam’s quality team needs to know exactly where each defective bolt sits on a conveyor-belt image. Simply flagging that the image contains a defect somewhere isn’t precise enough.
A) Custom Vision image classification
B) Custom Vision object detection
C) Azure Vision automatic captioning
D) Content Safety image categories
👁 Reveal Answer
Correct Answer: B
Explanation: Object detection returns coordinates for each object it finds. As a result, it can pinpoint every defective bolt individually on the belt. Image classification only assigns one label to the whole image instead, such as “defective” versus “acceptable,” without locating anything. Automatic captioning generates a general description of the scene, not defect coordinates. Content Safety categorizes harmful content types; it has nothing to do with manufacturing defects.
Question 20: Generating Accessible Alt Text at Scale
Finally, Tailwind Traders wants automatic, accessible alt-text generated for thousands of existing product photos in its catalog. Training a custom model for every product category simply isn’t practical at that scale.
A) Azure AI Custom Vision, trained on labeled examples per category
B) Azure Vision in Foundry Tools automatic captions and dense captions
C) Azure AI Video Indexer scene descriptions
D) Azure AI Document Intelligence layout model
👁 Reveal Answer
Correct Answer: B
Explanation: Azure Vision’s image analysis generates human-readable captions and dense captions out of the box, with no custom training required. Therefore, it’s ideal for accessible alt text across a large catalog. Custom Vision would require labeled training data for every product category instead, which is exactly the effort this scenario wants to avoid. Video Indexer describes video scenes, not standalone product photos. Document Intelligence analyzes document layout and fields, not product imagery.
Study Tips for AI-102 Computer Vision Solutions
A few habits will help with AI-102 computer vision solutions questions specifically. First, memorize the four-way split between Azure Vision, Custom Vision, Video Indexer, and Content Understanding. Each one owns a distinct slice: general image features, custom-labeled recognition, video and audio insight, and schema-defined multi-media extraction. Getting this map straight in your head resolves at least half of the scenario-based questions on this domain.
Second, don’t study Custom Vision as if nothing changed. Microsoft’s retirement announcement is recent enough that older courses and question banks may still present it as the default answer for any custom classification task. For new projects, Azure Machine Learning AutoML or Content Understanding is now the recommended path instead. Expect the real exam to reflect this shift, since Microsoft updates AI-102 content on a rolling basis.
Third, get comfortable with the Azure OpenAI versus Azure Vision boundary. Use Azure OpenAI when the task is open-ended or conversational. In contrast, use Azure Vision when you need a fixed, structured set of visual features returned quickly and cheaply.
Fourth, read the Microsoft Learn guide to choosing an image and video processing technology in full. It lays out exactly which service to use, and which to avoid, for nearly every scenario in this set. This Microsoft Learn architecture guide is worth bookmarking alongside your other AI-102 computer vision solutions notes.
Finally, practice spotting the one differentiating detail in each scenario. Specifically, watch for words like “live feed,” “structured schema,” “custom label,” or “open-ended,” since each one points at a different service. Building a quick mental checklist of these trigger words before exam day pays off far more than memorizing every API parameter.
Keep Practicing Your AI-102 Computer Vision Solutions Skills
That’s 20 questions down across two sets, and the natural language processing domain is up next. Therefore, keep this AI-102 computer vision solutions momentum going and come back for the next set, which moves into text analytics, translation, and speech.
In addition, if you haven’t worked through it yet, start with AI-102 Questions 1-10 to cover the generative AI domain first. Our AZ-305 mixed domain review and AZ-305 cost optimization set are also solid companions if you’re studying multiple Microsoft certifications at once. Good luck, and see you in the next set.
