Use this for last-mile review. AI-900 is mostly a classification exam: what kind of problem is this, what concept or service family fits it, and which tempting distractor actually solves a different problem? Keep the current Microsoft naming shift in mind: the certification page now says Azure AI Foundry is becoming Microsoft Foundry, but the live AI-900 study guide still uses the older wording.
Grounding: Supplying relevant source context so a generative model can answer from trusted material instead of only from its pretrained knowledge.
Inference: Using a trained model or prebuilt AI service to produce output for new input.
Validation data: Data used to check whether a model generalizes beyond the examples it learned during training.
AI-900 answer sequence
Use this when the stem mixes workload classification, responsible AI, ML basics, or service family choice.
flowchart TD
S["Scenario"] --> W["Classify the workload"]
W --> R["Check responsible AI or ML fit"]
R --> F["Pick the correct Azure service family"]
F --> E["Reject answers that solve a different problem"]
Fast lane picker
| If the question is really about… |
Focus first on… |
Strongest first move |
| what kind of problem is being solved |
workload classification |
decide vision, language, speech, document, ML, or generative AI before naming a service |
| ethics, bias, trust, or human oversight |
responsible AI |
map the scenario to fairness, safety, privacy, inclusiveness, transparency, or accountability |
| labeled vs unlabeled data |
ML basics |
decide classification, regression, or clustering first |
| image, document, OCR, or face clue |
vision lane |
identify the output type before the Azure service |
| speech, translation, sentiment, or entities |
NLP lane |
separate text-analysis tasks from audio tasks |
| chat, summarization, copilots, or grounded answers |
generative-AI lane |
decide whether the task truly creates content or only extracts or classifies it |
Workload chooser
| Requirement |
Strongest first fit |
Why |
| classify or detect things in images |
computer vision |
the output is about visual content |
| extract text or structure from forms, invoices, or receipts |
document processing |
the clue is about reading an existing document, not generating content |
| detect sentiment, entities, or key phrases in text |
NLP |
the system is analyzing human language |
| convert audio into text or text into speech |
speech |
the input or output starts with audio |
| predict a label or number from historical data |
machine learning |
the task is predictive rather than prebuilt perception |
| draft, summarize, rewrite, or chat |
generative AI |
the system creates new content |
Responsible-AI quick rules
| Principle |
Fast anchor |
Look for clues like… |
| fairness |
avoid unjust bias in outcomes |
different groups receiving systematically weaker decisions |
| reliability and safety |
behave dependably and reduce harm |
dangerous failure modes, unsafe outputs, weak fallback paths |
| privacy and security |
protect data and access |
sensitive prompts, exposed data, weak controls, unsafe sharing |
| inclusiveness |
work for varied users and conditions |
accent bias, disability barriers, language coverage gaps |
| transparency |
make AI use and limits understandable |
users do not know AI is involved or cannot interpret the result |
| accountability |
humans remain responsible |
no owner, no escalation path, no review step for risky output |
Machine-learning quick rules
| Concept |
Use it when… |
Do not confuse it with… |
| classification |
the output is a category such as approve or deny |
regression |
| regression |
the output is a number such as sales or delay |
classification |
| clustering |
you want unlabeled grouping |
classification |
| supervised learning |
labels already exist |
unsupervised learning |
| deep learning |
the pattern is complex, often in images, audio, or text |
all ML being generative by default |
| transformer architecture |
modern language and generative models rely on it |
a separate exam lane about low-level model internals |
| training |
the model learns from historical examples |
inference |
| inference |
the trained model produces output on new input |
training |
Computer-vision quick rules
| Scenario clue |
Strongest first answer |
| assign the whole image to a category |
image classification |
| identify what objects are present and where they appear |
object detection |
| read printed or handwritten text from an image |
OCR |
| locate a face or analyze face-related attributes |
facial detection or facial analysis |
| read fields and layout from a form or receipt |
document processing first, not generic image tagging |
NLP and speech quick rules
| Scenario clue |
Strongest first answer |
| detect sentiment, entities, key phrases, or summary |
Azure AI Language capability |
| convert spoken words into text |
Azure AI Speech capability |
| create spoken audio from text |
Azure AI Speech capability |
| translate text between languages |
translation capability |
| use a model to predict the next words in a sequence |
language modeling |
Generative-AI quick rules
| Requirement |
Strongest first fit |
Why |
| draft, summarize, rewrite, or answer conversationally |
generative AI |
new content is being created |
| answer from trusted internal content |
grounded generation |
retrieval context reduces unsupported answers |
| extract exact fields from an invoice |
document processing |
the task is reading existing structure, not generating it |
| predict churn, fraud, or pass/fail |
classification model |
the output is a label, not generated language |
| compare candidate models at a high level |
model catalog or evaluation capability |
AI-900 stays conceptual, not implementation heavy |
Azure service-fit quick rules
| Requirement |
Strongest first fit |
| custom training, experimentation, deployment, lifecycle management |
Azure Machine Learning |
| image analysis and OCR |
Azure AI Vision |
| face-related detection and analysis |
Azure AI Face detection service |
| text analytics and language understanding |
Azure AI Language |
| speech recognition, synthesis, or audio-first translation |
Azure AI Speech |
| large language model access on Azure |
Azure OpenAI Service |
| high-level generative-AI workspace, model access, and orchestration surface |
Azure AI Foundry / Microsoft Foundry naming transition |
| compare or browse available model choices |
Azure AI Foundry model catalog |
Commonly confused pairs
| Pair |
Keep this distinction clear |
| document processing vs generative AI |
extract existing information vs create new content |
| document processing vs generic computer vision |
form and layout reading vs general scene analysis |
| classification vs regression |
category vs number |
| classification vs clustering |
labeled prediction vs unlabeled grouping |
| training vs inference |
learning stage vs usage stage |
| Azure Machine Learning vs prebuilt Azure AI service |
custom model lifecycle vs ready-made AI capability |
| Azure AI Language vs Azure AI Speech |
text-first workload vs audio-first workload |
| grounding vs fine-tuning |
runtime context vs changing model behavior through training |
| Azure AI Foundry vs Azure OpenAI Service |
broader generative-AI platform surface vs model service access |
Last 15-minute review
| Review this |
Because it fixes… |
| workload classification order |
early service-choice mistakes |
| responsible-AI principles |
abstract-but-testable concept misses |
| regression vs classification vs clustering |
core ML confusion |
| training, validation, overfitting, and inference |
ML vocabulary drift |
| Vision vs Face vs Language vs Speech vs Azure Machine Learning |
service-family mispicks |
| grounding vs generic generation |
generative-AI distractors |
Keep going