Overview
Identify AI concepts and capabilities is 40-45% of AI-901 and covers three skill groups: Microsoft's six responsible AI principles, the components and configuration of AI models, and the recognition of AI workloads such as generative and agentic AI, text analysis, speech, computer vision and information extraction. The emphasis is on recognition and judgment: reading a short business scenario and naming the principle at stake, the model type that fits, or the workload being described.
This domain is the vocabulary layer for the rest of the exam. Domain 2 asks you to build lightweight apps in Microsoft Foundry, and each of those tasks starts with a decision made here: is this a text analysis job or a generative one, should the model be multimodal, which deployment type matches the traffic pattern, and what does a temperature change do to the answer. AI-901 replaced the older AI-900 outline, so classic machine learning theory (regression, clustering, Azure Machine Learning) is no longer a listed skill.
How it's tested
Questions are short scenarios with one decisive clue. Expect to match a situation to a responsible AI principle, a workload to a capability, a requirement to a model type, and a symptom to a configuration parameter. Distractors are usually the neighboring concept, so precise definitions beat general impressions, and the scenario's wording about outcome or harm decides the answer.
-
Name the principle
A scenario describes a harm or a safeguard, such as a loan model that rejects one group more often or a chatbot that discloses it is AI, and you name which of the six responsible AI principles applies.
-
Match the workload
A business need is described in plain language, and you identify the workload or technique, for example sentiment analysis rather than entity detection, or object detection rather than image classification.
-
Pick the model or setting
A requirement such as image input, low cost at scale, predictable output, or steady high-volume traffic points to a model type, a configuration parameter, or a deployment option.
A retailer wants a support bot whose answers about return policy are consistent every time the same question is asked. Options list raising temperature, lowering temperature, increasing max tokens, and switching to an image model. Consistency means less randomness in token selection, so lowering temperature is correct. Max tokens only limits answer length, and raising temperature makes wording vary more, the opposite of the goal.
Skills 14
Official skills from the Microsoft outline, in outline order.
1.1 Describe considerations for fairness in an AI solutionTreating comparable people the same, regardless of group membership.
- What it means
- A fair AI system allocates opportunities, resources and quality of service without disadvantaging people because of characteristics such as gender, ethnicity, age or disability.
- What to understand
- Unfairness usually enters through training data that under-represents or historically disadvantages a group, so mitigation starts with examining data and measuring outcomes per group, not just overall accuracy. A model can be accurate on average and still unfair to a subgroup.
- How it shows up
- A hiring or lending model approves one demographic group at a lower rate than another with similar qualifications, and you identify fairness as the principle at risk.
- Common confusion
- Candidates pick inclusiveness, but inclusiveness is about designing so everyone can use and benefit from the system; fairness is about outcomes that differ by group.
1.2 Describe considerations for reliability and safety in an AI solutionPerforming as intended, including under unexpected conditions.
- What it means
- A reliable and safe AI system behaves consistently with its design, handles unusual inputs without causing harm, and is tested against the conditions it will meet in production.
- What to understand
- This principle drives rigorous testing, defined operating limits, monitoring after deployment and, for generative AI, content filters that block harmful output. High-stakes uses such as medical or autonomous systems need explicit thresholds and a fallback when confidence is low.
- How it shows up
- A scenario describes an AI system that must be tested on edge cases before release, or that must hand off to a human when it encounters an input it was not designed for.
- Common confusion
- Reliability is often confused with accountability; testing and safe failure behavior are reliability and safety, while who answers for the outcome is accountability.
1.3 Describe considerations for privacy and security in an AI solutionProtecting personal data and the system itself.
- What it means
- AI systems should respect privacy by collecting and using personal data only with appropriate consent and controls, and should be secured against misuse and attack.
- What to understand
- Models are trained on and fed large amounts of data, so concerns include personal data in training sets, sensitive data sent in prompts, data that could be revealed in responses, and attacks such as prompt injection. Controls include data minimization, redacting personally identifiable information (PII), access control and encryption.
- How it shows up
- A company wants to analyze customer call transcripts but must remove names and account numbers first, which points to privacy and security and often to PII detection.
- Common confusion
- Transparency is about explaining how the system works; telling users how their data is stored and protected is a privacy and security consideration, not a transparency one.
1.4 Describe considerations for inclusiveness in an AI solutionDesigning so that everyone can use and benefit from the system.
- What it means
- Inclusive AI empowers and engages people across abilities, languages and circumstances, with accessibility and diverse user needs considered from the start of design.
- What to understand
- Typical inclusive measures are speech interfaces for people with limited vision, captions for people who are deaf or hard of hearing, support for many languages and accents, and involving diverse users in testing. Inclusiveness widens who can use the system at all.
- How it shows up
- An app adds text-to-speech output and voice commands so users with visual impairments can use it, and you identify inclusiveness.
- Common confusion
- Because both mention groups of people, candidates mix inclusiveness with fairness; if the issue is access or usability, it is inclusiveness, and if it is unequal outcomes, it is fairness.
1.5 Describe considerations for transparency in an AI solutionMaking the system understandable, including its limits.
- What it means
- Transparency means users and stakeholders understand that they are dealing with AI, what the system is for, how it reaches its results in broad terms, and where it is likely to be wrong.
- What to understand
- Practical transparency includes disclosing AI-generated content, publishing intended uses and limitations (as in Microsoft transparency notes), showing confidence scores, and explaining which factors influenced a prediction. It helps people decide how much to trust an output.
- How it shows up
- A chatbot must tell customers it is an automated assistant and that its answers can contain errors, which is transparency.
- Common confusion
- Candidates choose accountability because disclosure sounds like responsibility; disclosure and explanation are transparency, while ownership and oversight are accountability.
1.6 Describe considerations for accountability in an AI solutionPeople remain answerable for what the AI system does.
- What it means
- Accountability means the people who design, deploy and operate an AI system are responsible for its behavior, supported by governance, review processes and the ability for humans to intervene.
- What to understand
- An AI system is never the responsible party. Accountability shows up as named owners, internal review boards, audit trails, compliance with standards and laws, and human-in-the-loop review for consequential decisions.
- How it shows up
- An organization sets up a review board that must approve AI use cases and can override automated decisions, which is accountability.
- Common confusion
- Monitoring a model's performance in production is reliability and safety; deciding who is answerable and who can override it is accountability.
1.7 Describe how generative AI models workTokens, embeddings, transformers and next-token prediction.
- What it means
- Large language models split text into tokens, represent them as embedding vectors, and use a transformer architecture with attention to predict the most likely next token repeatedly until a response is complete.
- What to understand
- The model generates output from patterns learned during training, not by looking up facts, so it can produce fluent but wrong answers (fabrications or hallucinations). Its knowledge stops at a training cutoff, and it can only consider what fits in its context window, which is why grounding with your own data improves accuracy.
- How it shows up
- A scenario shows a model confidently giving an incorrect answer about company data it never saw, and you identify that the model predicts likely text and must be grounded in that data.
- Common confusion
- Candidates think a model retrieves answers from a database or learns from each conversation; a deployed model's weights do not change when users chat with it.
1.8 Identify an appropriate AI model, based on capabilitiesMatching model type to inputs, outputs, cost and quality needs.
- What it means
- The Foundry model catalog offers language models, reasoning models, small language models, multimodal models that accept images or audio, embedding models, image-generation models and speech models, each suited to different tasks.
- What to understand
- Choose by required input and output modalities first, then by quality, latency, context window and cost. Embedding models return vectors for search and similarity rather than text, reasoning models spend more time and tokens on multi-step problems, and small language models trade some capability for speed and lower cost.
- How it shows up
- A requirement such as analyzing photos submitted with a question, or building semantic search over documents, points to a multimodal model or an embedding model respectively.
- Common confusion
- Picking the largest general model for every job; an embedding model, not a chat model, is the right choice for vector search, and a smaller model may meet a simple classification need at lower cost.
1.9 Identify appropriate model deployment options and configuration parametersDeployment types plus parameters such as temperature and max tokens.
- What it means
- A model must be deployed before an app can call it, and the deployment type controls where processing happens and how capacity is billed, while runtime parameters shape each response.
- What to understand
- Standard deployments are pay-per-token, with Global, Data Zone or regional processing options that matter for data residency; provisioned deployments reserve throughput units for steady high-volume or latency-sensitive traffic; batch deployments process large asynchronous jobs at lower cost. Temperature and top_p control randomness, max tokens caps output length, and stop sequences end generation at a chosen string.
- How it shows up
- A scenario gives a constraint such as data that must stay in the EU, predictable cost for constant heavy load, or more creative wording, and you match it to a deployment type or parameter.
- Common confusion
- Increasing max tokens does not make answers better or more detailed by itself; it only allows longer output and raises the possible cost of each call.
1.10 Identify scenarios for common AI workloads, including generative and agentic AI, text analysis, speech, computer vision, and information extractionRecognizing which workload a business need describes.
- What it means
- Each workload solves a different kind of problem: generative AI creates new content, agentic AI pursues a goal by reasoning and calling tools, text analysis extracts meaning from language, speech converts between audio and text, computer vision interprets images and video, and information extraction turns unstructured content into structured fields.
- What to understand
- The key distinction for agents is action: an agent combines a model, instructions and tools so it can look things up, call functions and complete multi-step tasks, whereas a plain chat app only returns text. Information extraction is about producing a defined set of fields, while text analysis produces general insights such as sentiment or entities.
- How it shows up
- A description like reading invoices and saving vendor name and total to a database points to information extraction, while booking a meeting room after checking availability points to an agent.
- Common confusion
- Labeling any chatbot as agentic; without tools or actions that change something or fetch live data, it is a generative chat workload.
1.11 Describe common text analysis techniques, including keyword extraction, entity detection, sentiment analysis, and summarizationFour techniques for extracting meaning from text.
- What it means
- Keyword extraction pulls out the main talking points, entity detection finds and categorizes named items such as people, places, organizations, dates and PII, sentiment analysis scores text as positive, negative, neutral or mixed, and summarization condenses text into a shorter version.
- What to understand
- Summarization can be extractive (selecting the most representative existing sentences) or abstractive (writing new sentences). Sentiment results come with confidence scores, and opinion mining links sentiment to specific aspects such as food or service in a review.
- How it shows up
- A company wants to know which products are mentioned in reviews and whether customers are happy, which needs entity detection plus sentiment analysis, not keyword extraction alone.
- Common confusion
- Keyword extraction returns phrases without categorizing them; if the scenario needs to know that a phrase is a person, location or date, the technique is entity detection.
1.12 Identify features and capabilities of speech recognition and speech synthesisSpeech to text and text to speech.
- What it means
- Speech recognition converts spoken audio into text, and speech synthesis converts text into natural-sounding spoken audio using a chosen voice.
- What to understand
- Recognition can run in real time for live captions or in batch for recorded files, and can add features such as speaker diarization and custom vocabulary. Synthesis uses neural voices, and Speech Synthesis Markup Language (SSML) controls pronunciation, speaking rate, pitch and pauses. Speech translation combines recognition with translation.
- How it shows up
- A call center wants written transcripts of recorded calls, which is speech recognition, while an app that reads notifications aloud is speech synthesis.
- Common confusion
- Mixing up the direction: recognition is audio in, text out; synthesis is text in, audio out.
1.13 Identify features and capabilities of computer vision and image-generation modelsAnalyzing images versus creating them.
- What it means
- Computer vision models analyze visual input through tasks such as image classification, object detection, image captioning and tagging, optical character recognition (OCR) and face detection, while image-generation models create new images from text prompts.
- What to understand
- Image classification assigns a label to the whole image, object detection locates each object with a bounding box, and segmentation outlines objects at the pixel level. Multimodal language models can also describe or answer questions about an image. Face identification and verification are restricted Limited Access features, while face detection is broadly available.
- How it shows up
- A warehouse needs to count and locate each box in a camera frame, which is object detection; a marketing team wants new product images from a description, which is image generation.
- Common confusion
- Choosing image classification when the scenario needs the position or count of items; only object detection returns bounding boxes.
1.14 Identify techniques to extract information from text, images, audio, and videosTurning unstructured content into structured data.
- What it means
- Information extraction applies OCR and layout analysis to documents and images, transcription to audio, and frame, scene and transcript analysis to video, then maps the results to a defined set of fields.
- What to understand
- The output is structured data, such as JSON with named fields, that downstream systems can store or act on. In Foundry, Azure Content Understanding handles all four modalities with analyzers that either extract values that appear in the content or generate and classify values, such as a call summary or a sentiment label.
- How it shows up
- A scenario asks how to capture invoice numbers and totals from scanned PDFs, or a summary and topics from recorded meetings, and you identify information extraction rather than general text analysis.
- Common confusion
- Assuming OCR alone is extraction; OCR returns text, while extraction also identifies which text is the invoice total or the customer name.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Six responsible AI principles
Microsoft's principles are fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability, and each scenario question tests one of them.
- Remember
- Fairness concerns unequal outcomes, inclusiveness concerns access, transparency concerns understanding, and accountability concerns human ownership.
- Tested as
- Matching a short scenario to exactly one principle is one of the most predictable question types in this domain.
- Trap
- Choosing the principle whose name sounds closest to the scenario wording instead of identifying the actual harm or safeguard described.
-
Tokens
Models read and write text as tokens, which are word pieces, and both pricing and context limits are counted in tokens.
- Remember
- Input tokens and output tokens both count toward cost, and the context window limits how many tokens a single request can include.
- Tested as
- Explains why long prompts cost more and why a very long document may need to be chunked.
- Trap
- Treating a token as exactly one word; common words can be one token, but longer or rare words split into several.
-
Embeddings
An embedding is a vector of numbers that represents the meaning of text, so semantically similar text produces vectors that are close together.
- Remember
- Embedding models output vectors, not readable text, and they power semantic search and retrieval for grounding.
- Tested as
- Appears when a scenario needs search by meaning rather than exact keywords.
- Trap
- Selecting a chat completion model for a vector search requirement.
-
Transformer and attention
The transformer architecture uses attention to weigh how much each token in the context relates to every other token when predicting the next one.
- Remember
- Attention lets the model use context, such as knowing which noun a pronoun refers to, across a long passage.
- Tested as
- Underpins questions on how generative models produce coherent text.
- Trap
- Believing the model understands or verifies facts; it predicts likely continuations based on learned patterns.
-
Hallucination and grounding
A hallucination is fluent output that is factually wrong, and grounding reduces it by supplying trusted data in the prompt, often through retrieval.
- Remember
- Grounding changes what the model sees at request time; it does not retrain the model.
- Tested as
- Common answer when a model gives wrong answers about private or recent information.
- Trap
- Choosing fine-tuning to add frequently changing facts, when retrieval-based grounding is cheaper and stays current.
-
Temperature and top_p
Both parameters control randomness in token selection, with lower values giving more focused, repeatable output and higher values giving more varied output.
- Remember
- Adjust one of them at a time; lower temperature for factual or consistent answers, higher for creative drafts.
- Tested as
- Frequent in questions about inconsistent answers or dull creative output.
- Trap
- Raising max tokens to fix inconsistency, which only changes the allowed response length.
-
Max tokens
Max tokens sets the upper limit on how many tokens the model may generate in one response.
- Remember
- Too low a value truncates answers mid-sentence; a higher value permits longer output and higher cost but does not force it.
- Tested as
- Appears when responses are cut off or when cost per call must be capped.
- Trap
- Assuming a higher limit improves quality or detail.
-
Standard vs. provisioned vs. batch deployment
Standard deployments bill per token on shared capacity, provisioned deployments reserve throughput units for predictable performance, and batch deployments process large asynchronous jobs at a discount.
- Remember
- Global, Data Zone and regional variants decide where data is processed, which matters for residency requirements.
- Tested as
- Scenario clues such as steady high volume, EU-only processing, or overnight bulk jobs map directly to a deployment type.
- Trap
- Picking provisioned throughput for low or unpredictable traffic, where pay-per-token standard is cheaper.
-
Agentic AI
An agent combines a model, instructions and tools so it can reason about a goal, call tools or data sources, and complete multi-step tasks.
- Remember
- Tools are what separate an agent from a chat app; they let the agent retrieve knowledge or take actions.
- Tested as
- Workload questions now include agentic scenarios alongside classic workloads.
- Trap
- Calling any chatbot an agent even when it cannot use tools or take actions.
-
Entity detection vs. keyword extraction
Entity detection finds named items and assigns categories such as Person, Location, DateTime or PII, while keyword extraction lists the main phrases without categories.
- Remember
- If the scenario needs the type of the item, it is entity detection.
- Tested as
- These two are deliberately paired as distractors in text analysis questions.
- Trap
- Using keyword extraction to find dates or phone numbers.
-
Object detection vs. image classification
Image classification labels the whole image, while object detection finds each object and returns its class and bounding box.
- Remember
- Counting or locating items requires object detection.
- Tested as
- The most common computer vision distinction on the exam.
- Trap
- Choosing classification for a counting or location requirement.
-
SSML
Speech Synthesis Markup Language is XML that controls how synthesized speech sounds, including voice, rate, pitch, pauses and pronunciation.
- Remember
- SSML applies to text to speech, not speech recognition.
- Tested as
- The answer when a scenario needs to fine-tune pronunciation or pacing of a spoken response.
- Trap
- Assuming SSML improves transcription accuracy.
-
Information extraction output
Extraction produces structured fields, usually JSON, defined by a schema, rather than free text or general insights.
- Remember
- Extraction answers what is the value of this specific field, such as invoice total or caller intent.
- Tested as
- Distinguishes extraction scenarios from text analysis and OCR scenarios.
- Trap
- Stopping at OCR, which returns text but does not label which text is which field.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Fairness vs. inclusiveness vs. transparency vs. accountability
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Fairness | Comparable people get comparable outcomes | About unequal results across groups | Labeling an accessibility gap as fairness |
| Inclusiveness | Everyone can use and benefit from the system | About access, ability and language coverage | Labeling biased approval rates as inclusiveness |
| Transparency | People understand the system, its purpose and its limits | About disclosure and explanation | Confusing disclosure with ownership |
| Accountability | Humans and organizations answer for the system | About governance, oversight and the ability to intervene | Thinking monitoring performance is accountability rather than reliability |
Keyword extraction vs. entity detection vs. sentiment analysis vs. summarization
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Keyword extraction | Lists the main phrases in text | No categories attached | Using it to find dates or names by type |
| Entity detection | Finds named items and assigns a category | Returns type, such as Person or Location, and can flag PII | Treating it as a topic list |
| Sentiment analysis | Scores text as positive, negative, neutral or mixed | Measures opinion, with confidence scores | Expecting it to say what the opinion is about without opinion mining |
| Summarization | Condenses text into a shorter form | Extractive selects sentences; abstractive writes new ones | Assuming every summary quotes the source verbatim |
Standard vs. provisioned vs. batch deployment
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Standard (Global, Data Zone, regional) | Pay-per-token access on shared capacity | Best for variable or modest traffic; processing location set by the variant | Ignoring the data residency implication of Global |
| Provisioned | Reserved throughput units billed for capacity | Predictable latency and cost for steady high volume | Choosing it for light, unpredictable use |
| Batch | Asynchronous processing of large request files | Lower cost, results within a completion window, not real time | Using it for interactive chat |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Disclosure labeled as accountability
- What candidates pick
- Accountability for a requirement that users be told they are talking to an AI.
- Why it looks right
- Being upfront feels like taking responsibility.
- Why it's wrong
- Telling users what the system is and where it may fail is transparency; accountability is about who owns the system and can intervene.
Clue to watch for Words like disclose, inform, explain or limitations point to transparency.
-
Accessibility labeled as fairness
- What candidates pick
- Fairness for adding voice control and captions for users with disabilities.
- Why it looks right
- Both principles involve treating groups of people well.
- Why it's wrong
- Making a system usable by people with different abilities is inclusiveness; fairness concerns unequal outcomes from the model.
Clue to watch for If the change widens who can use the system, it is inclusiveness.
-
Max tokens as a quality dial
- What candidates pick
- Increase max tokens to get more accurate or consistent answers.
- Why it looks right
- A larger limit sounds like more capability.
- Why it's wrong
- Max tokens only caps response length; randomness is controlled by temperature or top_p, and accuracy by grounding and prompt design.
Clue to watch for The symptom is inconsistency or inaccuracy, not truncated output.
-
Chat model for semantic search
- What candidates pick
- Deploy a chat completion model to find documents similar in meaning to a query.
- Why it looks right
- Chat models understand language well.
- Why it's wrong
- Similarity search compares vectors, which an embedding model produces; a chat model returns text, not comparable vectors.
Clue to watch for Phrases like similar meaning, vector, or semantic search.
-
Provisioned throughput for occasional use
- What candidates pick
- A provisioned deployment for an internal tool used a few times a day.
- Why it looks right
- Reserved capacity sounds more reliable.
- Why it's wrong
- Provisioned capacity is billed whether or not it is used, so low or bursty traffic is cheaper on a standard pay-per-token deployment.
Clue to watch for Low, unpredictable or occasional traffic and a cost focus.
-
OCR mistaken for extraction
- What candidates pick
- OCR alone to capture invoice totals into a database.
- Why it looks right
- The totals are printed text, and OCR reads printed text.
- Why it's wrong
- OCR returns all the text but does not identify which value is the total; field extraction maps values to named fields.
Clue to watch for The output must populate specific named fields.
Answer framework
Use these steps to turn any short AI-901 concept scenario into the single answer it is testing.
- Underline the outcome the scenario wants, such as a label, a location, a field value, new content, or an action taken.
- Map that outcome to a workload: classification or detection for images, entity or sentiment for text, recognition or synthesis for speech, extraction for fields, generation or agents for content and actions.
- If the scenario describes a harm or safeguard, name who is affected and how, then pick the matching responsible AI principle.
- For model questions, check required input and output modalities first, then cost, latency and data residency constraints.
- For configuration questions, match the symptom to the one parameter or deployment type that controls it, and eliminate options that change something else.
Practice question
Want more like this? The AI-901 practice test has questions for every domain.
Checklist
0 of 12
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Start with the responsible AI principles because they are the most predictable questions in this domain and the easiest to lock in. Write one realistic example of a violation and one safeguard for each principle, then practice pairing them, focusing on the two pairs that cause most errors: fairness versus inclusiveness and transparency versus accountability.
Next, study how generative models work, then model selection and configuration. Learn the parameters by their effect, so that a symptom such as inconsistent answers or cut-off responses immediately suggests the right control. Treat deployment types as a short decision table built on traffic pattern, cost model and data residency. This group is the most underestimated part of the domain because it looks technical, yet the questions stay at the level of choosing an option.
Finish with the workload skills. Build a one-page grid that lists each workload, its typical input, its output, and a scenario keyword that signals it. Pay special attention to the new agentic and information extraction workloads, which did not feature in the same way on AI-900.
Afterwards, move straight to Domain 2 and deploy a model in the Foundry portal, because changing temperature in the playground and watching the output shift will make the configuration questions here feel obvious.
FAQ
How much of AI-901 is Domain 1?
Identify AI concepts and capabilities is weighted at 40-45% of the exam, with the remaining 55-60% in Implement AI solutions by using Microsoft Foundry.
Is AI-901 the same as AI-900?
No. AI-901 replaced AI-900, which Microsoft retired on June 30, 2026, as the exam for the Azure AI Fundamentals certification. The new outline drops classic machine learning topics such as regression, clustering and Azure Machine Learning, and adds agentic AI, model deployment and hands-on Foundry skills.
Do I need to memorize the responsible AI principles word for word?
You need to know all six names and, more importantly, the specific kind of concern each covers, because questions describe a situation and expect you to identify the principle from it.
How technical are the generative AI questions in this domain?
They stay conceptual: tokens, embeddings, the transformer's next-token prediction, hallucinations and grounding. You will not calculate anything, but you must know what each concept explains and what it does not.
What is the difference between temperature and max tokens?
Temperature controls how random the model's token choices are, which affects consistency and creativity. Max tokens only limits how long the response can be.
What makes a scenario agentic rather than just generative?
An agent uses tools, such as knowledge retrieval, code execution or calls to other systems, to work toward a goal and can take actions. A generative chat app only produces text from the prompt it receives.
How long should I study this domain?
As a planning estimate, about 10 to 15 hours suits most beginners, with extra time if responsible AI principles or text analysis techniques are new to you.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 1 knowledge
You have worked through the responsible AI principles, how generative models produce output, model and deployment choices, and the workloads AI-901 expects you to recognize. Check whether you can pick the right principle, workload and setting under exam timing.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the AI-901 AI tutor Get any concept explained your way