Reviewing the AI-901 outline

Hana's customer-service assistant pilot has been running in a sandbox for three weeks when she starts her final review for AI-901. It answers questions about store hours and return policies from a set of Harrow & Finch documents, and it has already taught her more than any reading could: why the model sometimes invents a policy that does not exist, why the tone changes when she adjusts one setting, and why the legal team asked to see how it handles a customer who mentions an injury. "That is the exam in miniature," Renata tells her. "Concepts, configurations and responsibility, then building with Foundry." This lesson reviews the AI-901 outline dated April 15, 2026, assuming you have studied MSAZ-110 and MSAZ-111.

What is different about AI-901

AI-901 replaced AI-900 as the exam for Azure AI Fundamentals in 2026; Microsoft's certification page now lists AI-901 as the required exam, and third-party guides report that AI-900 was retired at the end of June 2026. Compared with the other three fundamentals exams, AI-901 leans further toward practice. Its audience profile describes candidates at the start of their AI solution development careers who are familiar with Python syntax and Azure resources, and it says familiarity with REST APIs, SDKs and command-line tools is expected. Its second skill area, weighted 55 to 60, uses the verb implement. You will not write long programs, but you should be able to read a short Python snippet that calls a model or a service, recognize which setting does what, and choose the right Foundry capability for a scenario.

The study guide also says that most questions cover features that are generally available, but that preview features may appear if they are commonly used. In a field that changes as quickly as AI, that matters. Product names have changed several times in recent years, and older study material may use names that Microsoft no longer uses. When your material and Microsoft's current documentation disagree, follow the documentation, and note the old name in your mistakes log so that you recognize it if it appears in a distractor.

Area 1: Identify AI concepts and capabilities (40 to 45)

The responsible AI objectives ask you to describe Microsoft's principles for responsible AI. Microsoft names six: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Be ready to match each to a scenario. A loan model that approves applicants at different rates for reasons unrelated to risk raises fairness. An assistant that gives dangerous advice raises reliability and safety. A system trained on customer messages without consent raises privacy and security. An app that does not work with screen readers raises inclusiveness. A system whose users cannot tell they are talking to AI, or why it decided something, raises transparency. And a deployment with no one responsible for reviewing its behavior raises accountability.

The model components and configurations objectives cover what a model is made of and the settings you control. Know the difference between training a model and using a pre-trained one; the idea of tokens, the pieces of text a language model reads and writes, and of a context window, the limit on how many tokens a model can consider at once; the role of a system message, which sets the assistant's behavior and boundaries; and common settings such as temperature, which makes output more varied when higher and more predictable when lower, and a maximum number of output tokens. Embeddings, which represent text as lists of numbers so that similar meanings sit close together, underpin search and grounding. Grounding means giving the model relevant source material, such as Harrow & Finch's return policy, so that answers are based on it rather than on whatever the model learned in training.

The AI workloads objectives ask you to recognize the main kinds of AI work: generative AI that creates text, code or images; agents that use a model plus tools to carry out tasks; natural language processing such as sentiment analysis, key phrase extraction, entity recognition and translation; speech recognition and synthesis; computer vision such as image analysis, object detection and optical character recognition; and information extraction from documents and forms. A good review exercise is to list ten business problems and name the workload for each.

A worked example: why embeddings work

Hana's pilot finds the right policy document by comparing embeddings. Each document and each question becomes a vector of numbers, and the system picks the document whose vector points in the most similar direction to the question's. A common measure is cosine similarity, which is 1 for vectors pointing the same way and close to 0 for unrelated ones. Real embeddings have hundreds or thousands of numbers; the toy vectors below have three, but the arithmetic is the same.

import math
def cosine(a, b):
    dot = sum(x * y for x, y in zip(a, b))
    return dot / (math.sqrt(sum(x * x for x in a)) * math.sqrt(sum(y * y for y in b)))

question = [0.9, 0.1, 0.3]            # "Can I return a drill after 40 days?"
docs = {
    "Return policy": [0.8, 0.2, 0.4],
    "Store hours":   [0.1, 0.9, 0.2],
    "Loyalty points": [0.3, 0.3, 0.9],
}
for name, vec in docs.items():
    print(f"{name:15} {cosine(question, vec):.3f}")

Return policy   0.984
Store hours     0.271
Loyalty points  0.601

The return policy scores highest, so it is passed to the model as grounding. On the exam you will not calculate cosine similarity, but you may be asked what embeddings are for, or why grounding reduces made-up answers. The answer to the second is that the model is given the relevant facts and instructed to base its answer on them.

Area 2: Implement AI solutions by using Microsoft Foundry (55 to 60)

Microsoft Foundry is Microsoft's platform for building AI applications and agents on Azure. The exam's subskills cover four kinds of solution. Generative AI apps and agents: choosing and deploying a model from the model catalog, testing prompts in a playground, writing a system message, grounding the model in your own data, and building an agent that can call tools. Text and speech: analyzing text, translating it, converting speech to text and text to speech. Computer vision and image generation: analyzing images, reading text in them and generating images from descriptions. Information extraction: pulling structured fields from documents, forms and other content. For each, know what the capability does, when to use it, and what a short code sample that calls it looks like.

Several distinctions recur in practice items. A playground is for experimenting; an application calls a deployed model through an endpoint with a key or an identity. Prompt changes and system messages adjust behavior without retraining; fine-tuning changes the model itself and is rarely the first choice. Content filters and safety settings reduce harmful output but do not replace human review for high-risk uses. And an agent differs from a simple chat app because it can decide to call tools, such as looking up an order, rather than only producing text.

Reading code on the exam

Because the audience profile expects Python familiarity, be ready for items that show a short code fragment and ask what it does or what is missing. Typical fragments create a client with an endpoint and a credential, send a list of messages with a system message and a user message, set a temperature, and print the response. The exam is testing whether you understand the pieces: which line sets the behavior, which line sends the request, where the key or identity comes from, and why a secret should not be written into the code. Hana's practice is to explain each line of the samples in MSAZ-111 aloud. If she cannot, she adds the line to her mistakes log.

Original practice items

Try these before reading the answers. (1) A retailer wants to know whether product reviews are positive or negative. Which workload? Natural language processing, specifically sentiment analysis. (2) True or false: a system message is visible to and written by the end user. False; the developer writes it to set the assistant's behavior, and users normally do not see it. (3) A model gives answers that ignore the company's own policies. Which technique most directly addresses this? Grounding the model in the policy documents. (4) A support line wants callers' speech turned into text so it can be searched later. Which capability? Speech to text. (5) Which principle is raised if nobody can explain why a model recommended a particular price? Transparency.

(6) A developer tests prompts interactively before writing any code. Where in Foundry would they do this? In a playground. (7) An application must stop producing violent or hateful content. Which feature helps? Content filters, alongside human review for sensitive uses. (8) A team wants a model that writes product descriptions in the company's distinctive style, and prompting has not worked after many attempts. Which option might they consider next? Fine-tuning, accepting its extra cost and effort. Several of these test the order in which you would try things, which is a common pattern in AI-901 items.

Machine learning basics still matter

Although the current outline centers on generative AI and Foundry, the idea of a model as something learned from data is still in scope under model components. Know that a model is trained on examples, that training data quality shapes what a model can do and the biases it may carry, and that a model is evaluated on data it did not see during training. Know the difference between predicting a number, such as next month's sales, and predicting a category, such as whether a message is a complaint. These ideas also help with the responsible AI objectives, because many fairness problems begin with unrepresentative training data.

Hana's final mistakes log for AI-901 has nineteen entries. Six are product names she had learned in older forms. "The concepts didn't change," she says. "The labels did."