Azure AI-900 common mistakes
Azure AI-900 · Understand the traps and how to avoid them.
Reads a scenario such as 'identify whether a photo of a manufactured part shows a defect' and selects the anomaly detection workload because the sentence contains the word 'detect', rather than selecting a computer vision workload.
Exam stems are written with action verbs (detect, extract, predict, classify) that appear across every workload family, so candidates pattern-match on the verb instead of on the input data type and the required output.
What to do instead
Before answering, write down two things: the input format (image, text, audio, numeric/tabular) and the output format (a label, a number, a location, a grouping). Map input to the workload family first, then use the output to pick the specific technique. Do not use the verb in the stem as evidence.
Treats the responsible AI principles as interchangeable, for example stating that 'inclusiveness' means the model produces equally good results for every demographic group, or that 'transparency' means the system is secure.
Several of the six principles are near-synonyms in everyday English (fairness vs inclusiveness, transparency vs accountability), so the wording feels familiar and the candidate answers from general intuition rather than from the formal definitions.
What to do instead
Memorise one distinguishing test word per principle: fairness = parity of outcomes across groups; inclusiveness = usable by people with the widest range of abilities and needs; reliability and safety = behaves correctly under unexpected or adverse conditions; privacy and security = data is protected and controlled; transparency = people can understand how and why a decision was made; accountability = humans remain responsible and there is governance. If the stem is about who is answerable, it is accountability; if it is about who is able to use it, it is inclusiveness.
Recommends Azure Machine Learning for every scenario, including ones that only need summarising text, translating speech, or detecting faces in an image.
Azure Machine Learning is the most prominent service name in the domain and feels like the 'safe' answer for anything involving AI, while the prebuilt Azure AI services are taught more briefly.
What to do instead
Apply a two-question rule. Does an existing prebuilt model already perform this task (language, vision, speech, document extraction)? If yes, the answer is an Azure AI service / prebuilt model. Does the organisation need to train a model on its own labelled data and manage the model lifecycle? If yes, the answer is Azure Machine Learning. Only pick Azure Machine Learning when the scenario mentions custom training, datasets, compute, or model evaluation.
Labels a scenario as clustering when it is classification, or as classification when it is regression, for example calling 'predict whether a customer will leave' clustering, or 'predict the price of a house' classification.
The everyday meaning of 'group', 'categorise' and 'predict' blurs together, and the candidate does not check whether labelled answer values exist in the historical data.
What to do instead
Ask whether the training data already contains the answer (a label). No labels, and the goal is to discover natural groupings: clustering (unsupervised). Labels present: look at the label's data type - a category, yes/no, or class name is classification; a numeric quantity is regression; a numeric value over time is forecasting. Anchor each option to the label type before choosing.
Chooses image classification when the stem requires locating several separate objects and returning coordinates, or chooses object detection when the stem asks for a per-pixel outline of each object.
All three techniques take images and produce object-related output, so they are remembered as one bucket; candidates rarely rehearse the output format that distinguishes them.
What to do instead
Decide by output shape, not by the object count. One label for the whole image: image classification. Many labels for the whole image with no location: multilabel classification. Bounding boxes with coordinates for each object: object detection. A mask that assigns every pixel to a class: semantic segmentation. If the stem mentions drawing a box, it is detection; if it mentions outlining the exact shape of each object, it is segmentation.
Answers with Azure AI Vision OCR for a scenario about pulling vendor name, invoice date and line-item totals out of scanned invoices, or answers with Face for a scenario about extracting printed text from a photograph.
Both services are described as 'reading things from images', and the general-purpose Vision resource does contain a Read capability, so the more specific document service is overlooked.
What to do instead
Separate 'where is the text' from 'what does the field mean'. If the task is to read text from a photo, sign, or scene: Vision OCR / Read. If the task is to extract named fields and values from structured or semi-structured forms, invoices, or receipts: Azure AI Document Intelligence. If the task is to detect or analyse human faces: Face. Check the stem for the words forms, fields, invoices, or receipts to trigger Document Intelligence.
Mismatches the NLP feature to the requested output, for example choosing key phrase extraction when the stem asks for the people, places and organisations named in the text, or choosing sentiment analysis when the stem asks what the user wants the application to do.
All of these features take text and return short text-based tags or scores, so the candidate remembers the family but not the specific return type.
What to do instead
Memorise the exact return shape of each feature: sentiment = positive/negative/neutral with confidence scores; key phrase extraction = the main concepts as a list; named entity recognition = tokens tagged by type such as person, location, organisation, datetime, quantity; language detection = a language name and code; intent recognition in conversational language understanding = the action the user wants plus extracted entities. Match the noun in the question to the return shape.
Treats conversational language understanding, question answering and translation as the same service, recommending question answering for a scenario where a user issues commands like 'book a flight to Paris tomorrow' that must be mapped to an action, or recommending conversational language understanding for an FAQ knowledge base.
All three consume natural language and appear in conversational solutions, and the marketing descriptions of each mention 'chatbots' and 'conversation'.
What to do instead
Ask what the system does with the utterance. Map it to an action or command with parameters: conversational language understanding (intents and entities). Retrieve the best answer from a curated knowledge base of question-and-answer pairs: question answering. Convert text between languages: translation. If the stem mentions utterances mapped to intents, it is conversational language understanding; if it mentions an FAQ or knowledge base, it is question answering.
Describes retrieval-augmented generation as fine-tuning, or says fine-tuning is the way to give a model access to a company's current internal documents.
Both are described as 'making the model know more about your data', and both appear near each other in generative AI guidance.
What to do instead
Use the weight-change test. If the model's parameters are unchanged and grounding documents are retrieved and inserted into the prompt at request time, that is retrieval-augmented generation - it is the correct choice for frequently changing or private factual content. If additional training on example data changes the model's behaviour, tone, or format, that is fine-tuning. Fine-tuning is not the tool for injecting current facts.
States that a well-written prompt makes a generative model's output reliable and factual, or that prompt engineering changes or retrains the underlying model.
Prompt engineering is presented as the primary skill for working with generative AI, so it is credited with guarantees it cannot provide; and because prompt quality visibly changes output, candidates infer that the model itself has been modified.
What to do instead
Treat generative output as unverified by default. Prompt engineering only shapes the input at inference time; it does not alter model weights and does not guarantee correctness. Reduce risk with grounding the prompt in retrieved, authoritative content, adding content filters and moderation, and keeping a human review step for consequential outputs. On the exam, when a stem asks how to make output trustworthy, prefer grounding and verification over 'write a better prompt'.