AI-901 Complete Guide: Azure AI Fundamentals New Exam

AI-901: Microsoft Azure AI Fundamentals went generally available in June 2026 and has replaced AI-900, which was retired on June 30, 2026. It is positioned as the successor to AI-900, but it is far more than a rename. Where the old AI-900 introduced individual Cognitive Services at the conceptual level, AI-901 pivots hard toward writing actual code on top of the Microsoft Foundry unified platform.

Three questions tend to come up first: AI-900 or AI-901? Is Python required? What changes with Microsoft Foundry? This article starts from the official Skills measured (as of 2026-04-15) and walks through what Microsoft Foundry really is, how Responsible AI is tested, the concrete differences from AI-900, and how AI-901 leads into AI-103.

AI-901 Exam Specs

The certification you earn from AI-901 is Microsoft Certified: Azure AI Fundamentals — the same badge AI-900 used to grant. Microsoft has not retired that certification; it refreshed the requirement to earn it, so AI-901 is now the exam you pass to get it. What differs is the exam content and difficulty: AI-900 was about conceptual understanding, AI-901 is about implementation on Microsoft Foundry. If you already hold the certification from AI-900, it stands as it is — there is nothing to retake.

The passing score is the Fundamentals standard of 700 / 1000. Duration and question count are not published on the official exam page. Pricing only says "varies by country/region" — no fixed value is published either, so check the amount on the official exam page with your country/region selected, or in the Pearson VUE booking flow, before you book.

The certification is valid for 1 year. You can renew it via the free renewal assessment on Microsoft Learn. Unlike AZ-900 and other Fundamentals certifications, which don't expire, AI-901 requires annual renewal — Microsoft's acknowledgement that AI moves fast enough to need it.

The Decisive Differences Between AI-900 and AI-901

The biggest difference between AI-900 and AI-901 is whether coding is required. AI-900 was a concept exam open to technical and non-technical candidates alike — zero Python knowledge was fine. AI-901 Domain 2 (55-60% of the exam) is almost entirely Foundry SDK implementation, and basic Python syntax plus pip install-level operations are assumed knowledge.

The domain count itself has changed. AI-900 had 5 domains (each 15-25%) that walked through Cognitive Service capabilities one by one (Vision / Speech / Language / Decision / Generative AI). AI-901 collapses these into just 2 domains: Domain 1 covers general AI concepts and Responsible AI principles, and Domain 2 covers implementation on Microsoft Foundry. Individual Cognitive Services have been unified as Foundry Tools — for example, the old Form Recognizer is now folded into Azure Content Understanding.

The weight of generative AI has also grown sharply. AI-900 packed generative AI into a single 20-25% domain; in AI-901, nearly all of Domain 2 is generative AI and agent implementation on Foundry. Modern mainstream topics — prompt engineering, multimodal models, single-agent design and testing — are tested head-on.

The crucial point: AI-900 retired on June 30, 2026 and can no longer be taken, but the Azure AI Fundamentals certification itself was not retired. Microsoft replaced the exam behind it, so candidates who say "I just want the concepts" or "I don't do Python" no longer have an AI-900 escape hatch — AI-901, Python and all, is the only route to this certification now. Anyone who already passed AI-900 keeps the certification they earned.

What Microsoft Foundry Really Is

To understand AI-901 you first have to nail down what Microsoft Foundry is. Foundry is the unified enterprise AI platform formalized in the January 2026 Microsoft Product Terms. The naming history is messy: Azure AI Studio → Azure AI Foundry → Microsoft Foundry, three renames to land on the current name. Any AI-900 material that says "Azure AI Studio" should be mentally translated to "Microsoft Foundry."

The essence of Foundry is unifying Azure OpenAI + Cognitive Services + Hub resources. Previously you had to create Azure OpenAI Service, the various Cognitive Services, and Hub resources separately and wire them together. With Foundry, a single Foundry resource handles everything. The model catalog lists 1,900+ models: Microsoft's Phi-4, OpenAI's GPT-5 / GPT-4.1, Anthropic Claude, xAI Grok, Mistral, DeepSeek-R1, Meta Llama, and more — all callable through a unified interface.

The API surface has also evolved. The old Assistants API (v0.5 / v1) has been replaced by the Responses API (Agents v2), and the SDK has converged on azure-ai-projects 2.x plus a unified OpenAI() client. Existing Azure OpenAI resources can be upgraded to Foundry resources in place, keeping the same endpoint, API key, and state.

For exam prep, opening the Foundry portal (https://ai.azure.com), deploying a model, and trying out Foundry Tools (Azure Speech / Azure Content Understanding) hands-on is essential. Without knowing the portal layout, screenshot-style questions ("from which tab in the Foundry portal do you select what?") become hard to navigate.

Domain 1: Identify AI Concepts and Capabilities (40-45%)

This domain rests on three pillars: the Responsible AI principles, AI model components, and scenario identification across AI workloads.

The 6 Responsible AI Principles

You're tested on whether you can explain each of the six principles defined by Microsoft's Responsible AI Standard at the level of "why it matters" and "what implementation-level considerations it implies." The order is Fairness → Reliability and safety → Privacy and security → Inclusiveness → Transparency → Accountability. AZ-900 also asks 1-2 questions here, but AI-901 goes deeper, with scenario-based items asking which principle is most at stake in a given situation.

AI Model Components and Configuration

You're tested on how generative AI models work (Transformer, tokens, context window), capability-based model selection, and deploy options plus configuration parameters (temperature / top_p / max_tokens and so on). Opening the Foundry Model Garden and comparing each model's characteristics (fast vs high-accuracy, cost) makes implementation-scenario questions far easier to handle.

Scenario Identification Across AI Workloads

Generative AI / Agentic AI, text analysis (keyword extraction, entity detection, sentiment, summarization), speech recognition / speech synthesis, computer vision / image generation, and information extraction from text / image / audio / video. Questions ask which technology best fits a given business problem and whether multimodal is required. Similar in spirit to AI-900, but the granularity is dialed up to concrete Foundry service selection.

Domain 2: Implement AI Solutions Using Microsoft Foundry (55-60%)

The largest-weighted domain demonstrates implementation skills with the Foundry SDK. It splits into four sub-areas.

Implementing Generative AI Apps and Agents

Crafting effective system and user prompts, deploying and interacting with models in the Foundry portal, building a lightweight chat client with the Foundry SDK, and constructing and testing single-agent solutions. Prompt engineering is one area where lived experience tweaking prompts in the Foundry portal pays off far more than reading a study guide.

Text and Speech Implementation

Building lightweight apps that include text analysis, responding to audio prompts with multimodal models, and constructing speech apps using Azure Speech in Foundry Tools. What AI-900 treated separately as Speech to Text / Text to Speech / Speech Translation is now consolidated under Azure Speech inside Foundry Tools.

Computer Vision and Image Generation

Interpreting visual input with multimodal models, generating visual output with generative models, and building lightweight apps that include vision capabilities. Instead of the old Computer Vision API, the central pattern is using Foundry multimodal models (e.g., GPT-4.1) to interpret images as input.

Information Extraction

Information extraction from documents and forms with Azure Content Understanding (under Foundry Tools), extraction from images / audio / video, and building lightweight apps that lean on Content Understanding. The old Document Intelligence (Form Recognizer) has been folded into Content Understanding, so anything you studied under the Form Recognizer name needs to be re-mapped.

How to Memorize the 6 Responsible AI Principles

The six Responsible AI principles also appear on AZ-900 and SC-900, but AI-901 probes the order and the considerations behind each principle more deeply. A practical mnemonic is F-R-P-I-T-A (Fairness → Reliability → Privacy → Inclusiveness → Transparency → Accountability) — keeping that letter order in your head removes the hesitation when scanning answer choices.

By "considerations" we mean concrete concerns. For Fairness, "could biases in the training data disadvantage a particular group?" For Privacy, "could user-provided data be re-used in training?" For Accountability, "who is the human ultimately responsible for the AI's decisions?"Just remembering the names of the principles will leave you guessing on scenario-based answer choices.

Official Study Resources

Because AI-901 is still a new exam, the official resource catalog isn't as rich as AZ-900's — but the essentials are all there.

If the official Practice Assessment isn't in your hands yet, you may be tempted to reach for third-party question banks. But for a brand-new exam like AI-901, the question-trend analysis hasn't matured yet, and you're risking time on inaccurate banks. For now, focus on the official Study Guide and hands-on practice in the Foundry portal; once you can access the Practice Assessment, drill it repeatedly.

How AI-901 Compares to G-ken / E-shiken

In the Japanese market, the best-known AI certifications are JDLA's G-ken (G-test) and E-shiken (E qualification). AI-901 sits in a clearly different place. G-ken is a concept exam aimed at business users — six sittings per year, no Python required, and still highly recognized at Japanese SIers and end-user companies. AI-901 goes beyond business concepts into Azure/Foundry implementation, making it more developer-leaning. For evaluating technical staff, AI-901 has the edge in direct on-the-job relevance.

E-shiken sits even higher — JDLA-accredited coursework is a prerequisite, and the exam tests deep-learning math and implementation at the highest difficulty tier. AI-901 plays in a different game, with the relative ordering roughly E-shiken ≫ AI-103 > AI-901 > G-ken. For consultants and PMs, "G-ken (business understanding) + AI-901 (Azure implementation)" works well in the field; for ML engineers, "E-shiken + AI-103" is the stronger combination.

Next Steps: AI-103 and Beyond

After AI-901 establishes the basics, the standard next stop is Exam AI-103: Developing AI Apps and Agents on Azure, which went GA in June 2026. AI-103 is the successor to AI-102 (Azure AI Engineer Associate), a certification that retired in full on June 30, 2026. The certification you earn from AI-103 is Microsoft Certified: Azure AI Apps and Agents Developer Associate, making the AI-app/agent-developer positioning explicit.

AI-103 treats the Foundry SDK as prior knowledge from AI-901 and pushes into more advanced topics: deeper prompt engineering, multi-agent orchestration, and responsible AI governance. The lightweight chat client you wrote for AI-901 grows into multi-agent collaboration, memory management, and tool use in AI-103. AI-102 retired on June 30, 2026, so anyone starting the AI Engineer track from here on takes AI-103.

AI-901 → AI-103 is the AI developer route, but you can also combine AZ-104 (Administrator) or DP-700 (Fabric Data Engineer) to aim for AI-infrastructure operations or AI data-pipeline roles. Note that the older DP-203 (Azure Data Engineer Associate) was retired — certification and all — on March 31, 2025; on this route the certification you now earn is Microsoft Certified: Fabric Data Engineer Associate via Exam DP-700: Implementing Data Engineering Solutions Using Microsoft Fabric. As Microsoft Foundry spreads as the enterprise AI substrate, three roles each have their own certification track: "those who use AI models," "those who run the AI infrastructure," and "those who prepare data for AI."

Frequently Asked Questions

Terminology and weighting in this article are based on the official Microsoft Learn AI-901 exam page and the official Study Guide. This article is not an official Microsoft Corporation product and has no affiliation or sponsorship relationship. Microsoft, Azure, Microsoft Foundry, and Microsoft Entra are trademarks of the Microsoft group of companies. AI-901 went GA in June 2026; exam scope and specifications may still change, so always confirm the latest information on the official page. Information current as of May 24, 2026, updated on August 30, 2026 to reflect the retirement of the AI-900 exam (June 30, 2026) and the GA of AI-901.