Azure AI Fundamentals (AI-901) common mistakes
Azure AI Fundamentals (AI-901) · Understand the traps and how to avoid them.
Treating generative AI, machine learning, and rule-based automation as interchangeable labels, so a scenario about a chatbot that composes new text gets classified as classic classification ML.
Marketing language around AI products blurs the line, and candidates map familiar terms like 'model' or 'training' onto any AI scenario without checking whether the output is generated content or a predicted label.
What to do instead
Before answering, ask two questions: does the system produce new content (text, images, code) or assign a label/score to existing data, and does it learn patterns from examples or follow hand-written rules? Generative AI maps to content creation, ML to prediction from learned patterns, and rule-based systems to deterministic logic with no learning.
Confusing computer vision capabilities, for example describing optical character recognition as detecting objects or facial recognition as image classification, when a scenario asks which workload fits.
All vision tasks accept images as input, so the input type feels like the deciding factor and candidates stop reading for the output type, which is what actually distinguishes the workloads.
What to do instead
Classify by output: text extracted from an image is OCR, bounding boxes with labels are object detection, one label per image is classification, and identity or attribute inference from a face is facial recognition. Write the output type next to each option before choosing.
Mixing up natural language processing subtasks, such as calling sentiment analysis 'key phrase extraction' or treating translation as an entity recognition task, in scenario questions.
These workloads are often delivered by the same service and described in similar language, and candidates rely on the scenario's topic rather than the exact transformation requested.
What to do instead
Anchor on the transformation: sentiment returns a polarity, key phrase extraction returns salient terms, entity recognition returns typed entities like people or places, translation returns text in another language, and question answering returns an answer span. Match the requested output, not the domain of the text.
Assuming responsible AI principles are only about avoiding offensive output, so a scenario about a hiring model with skewed training data gets answered with a content-filtering control instead of fairness mitigation.
Safety filters are the most visible responsible AI feature, and candidates collapse all six principles into one mental bucket, losing the ability to map a specific harm to the matching principle.
What to do instead
Memorize the six principles as a checklist (fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability) and attach one concrete example to each. Then identify which harm the scenario describes and select the principle and mitigation that address that harm specifically.
Choosing a model deployment option by capacity rather than by the workload's latency and data-handling needs, for example picking a provisioned deployment for an intermittent low-volume prototype.
Candidates see throughput numbers and assume bigger is better, without weighing cost model, traffic pattern, and whether the scenario needs guaranteed capacity or pay-per-use.
What to do instead
Read the scenario for traffic pattern and budget signals: steady high volume with latency guarantees points to provisioned throughput, while spiky or experimental usage points to a standard pay-as-you-go deployment. State the deciding constraint explicitly before selecting.
Confusing the roles of a model, a deployment, and an application in Foundry, such as expecting to call a model directly without a deployment or expecting a deployment to include prompt logic.
The portal presents models and deployments on adjacent screens, and candidates who have not built an end-to-end flow assume the model endpoint and the application are the same object.
What to do instead
Draw the chain once: model catalog entry, then deployment that exposes an endpoint, then application code or an agent that sends prompts and handles responses. When a question asks what is missing from a failing call, check whether a deployment exists and whether the client targets the deployment name.
Treating prompt engineering as a substitute for grounding, so a scenario requiring answers from an internal document set is solved by rewording the system prompt instead of adding retrieved context.
Prompting is taught first and feels universally applicable, and candidates underestimate how reliably a model hallucinates on private or recent facts without supplied context.
What to do instead
Ask whether the required knowledge is in the model's training data. If the answer depends on private, current, or verifiable sources, the fix is retrieval or grounding that injects the source text, with prompt wording used only to shape format and tone.
Selecting an evaluation approach that does not match the task, for example judging a summarization solution only by exact-match scoring or skipping evaluation because the output 'looks good' in one manual test.
Candidates default to the metric they remember and do not consider that generative outputs vary, so single-sample inspection and string-matching metrics both give misleading confidence.
What to do instead
Match metric to task: classification uses accuracy, precision, recall, or F1; generation quality uses similarity or model-assisted scoring plus human review; grounded systems also need groundedness and relevance checks. Plan a small labeled test set rather than one anecdote.
Ignoring content safety and abuse controls when designing a user-facing solution, then being unable to answer which layer blocks harmful input versus harmful output.
Safety is treated as an automatic platform property, so candidates do not learn that filters can be configured and that separate controls exist for input, output, and repeated abuse.
What to do instead
Learn the layers distinctly: input filtering screens prompts, output filtering screens completions, and abuse monitoring throttles or blocks patterns over time. For each scenario, identify whether the risk enters through the user or leaves through the model, then choose the corresponding control.
Answering cost and quota questions by guessing at token economics, such as assuming output tokens are free or that a larger model always costs the same per call regardless of prompt length.
Pricing details feel like trivia, so candidates skip them, but exam scenarios embed cost constraints in the prompt and expect the candidate to reason about input versus output usage.
What to do instead
Remember that billing scales with tokens processed and that input and output are priced separately, so trimming context and limiting response length both reduce cost. When a scenario states a budget or latency limit, use those numbers to eliminate options before choosing.