Microsoft AI-500 is an expert-level exam that tests whether you can build several AI agents that work together in production without falling over. The full title is Exam AI-500: Designing and Implementing Multi-Agent AI Solutions, and passing it on its own earns the Microsoft Certified: Multi-Agent AI Solutions Expert credential.
It is still a beta exam at the time of writing. That matters more than it sounds, because beta status changes what you pay, when you get your result, and how much prep material exists. This post covers the AI-500 exam topics and their weightings straight from the Microsoft Learn study guide, the things Microsoft has deliberately not published yet, and whether the exam is worth booking now or after it goes generally available.
What the AI-500 exam actually covers
The distinction that trips people up: AI-500 is not about building an AI application, and it is not about building one agent. AI-102 and its replacement AI-103 cover that ground. AI-500 assumes you already have several agents and asks whether you can make them coordinate, share memory, hand off tasks, stay inside their permission boundaries, and keep running when one of them fails.
Microsoft's audience profile for the exam is specific. You are expected to be proficient in Python, to have deployed agentic systems in production, and to have orchestrated agent logic using Microsoft Foundry. You also need working familiarity with Microsoft Agent Framework, Model Context Protocol (MCP), retrieval-augmented generation (RAG) and LangGraph, plus experience with Azure compute, network, storage and data services.
That is an unusually concrete audience profile by Microsoft's standards, and it is a warning. This is not a certification you can read your way through.
Exam Tip: The passing score for AI-500 is 700, the same as every other Microsoft role-based exam. Microsoft scores on a scaled 1000-point system, so 700 is not 70% of the questions answered correctly.
AI-500 exam topics and domain weightings
Microsoft publishes four skill areas with percentage bands rather than fixed weightings. Here they are exactly as listed in the official study guide.
| Domain | Weighting | What it tests |
|---|---|---|
| Architect multi-agent solutions | 15 to 20% | Decomposing goals into workflows and agents, agent personas and autonomy levels, memory architecture, protocol choice, Zero Trust and per-agent identity scoping |
| Develop multi-agent solutions in Azure | 30 to 35% | Advanced prompt engineering, context management, multi-agent RAG, building MCP servers and clients, orchestration patterns, Agent Framework, LangChain and LangGraph |
| Evaluate, optimise and monitor multi-agent solutions | 20 to 25% | Evaluation design in Foundry, context window diagnostics, LLM-as-a-judge loops, tracing, token and cost monitoring |
| Secure, govern and deploy multi-agent solutions | 20 to 25% | Identity-based access and RBAC, OAuth 2.0 and on-behalf-of flows, Key Vault secrets, guardrails, AI Red Teaming Agent, blue/green and canary release |
Two things stand out when you plan study time around that table.
The development domain is the biggest single block at up to 35%, and it is the most hands-on. Orchestration patterns alone cover hub-and-spoke, sequential, parallel, peer-to-peer and orchestrator-subagent, and you are expected to implement them rather than recognise them.
The two operational domains together account for 40 to 50% of the exam. Candidates coming from a development background consistently underweight these. If you can build a working agent swarm but have never traced token usage across it, set a per-agent cost quota, or written a rollback plan for an agent release, that half of the exam will hurt.
The specific subtopics worth flagging
A few bullets in the study guide describe failure modes that only appear at multi-agent scale, and they are the ones with the least third-party coverage:
- Context window diagnostics: sliding-window amnesia, summary drift, vector-only recall and entity continuity issues
- Multi-tier state persistence covering session state, shared team state and long-term semantic memory, with tenant-isolation policies
- Lateral movement prevention between agents, and compliance control mapping for regulated deployments
- Agent replay capture for reproducible debugging, and behavioural drift detection
If your notes do not have entries for those four, you have a gap.
AI-500 prerequisites
Microsoft lists no prerequisite certification for AI-500. The exam page shows "Related exams: none", and AI-500 is the sole requirement for the Multi-Agent AI Solutions Expert certification. You can book it cold.
Whether you should is a different question. Unlike the Azure administrator and security tracks, where an expert certification sits on top of a named associate exam, Microsoft has left this one open. In practice the readiness bar is the audience profile, not a badge. Someone holding AI-103 or AI-200 has roughly the right Azure AI grounding, but neither exam goes near multi-agent orchestration, agent memory architecture or per-agent identity scoping.
Exam Tip: No prerequisite is not the same as no assumed knowledge. Microsoft expects Python proficiency and production experience with Microsoft Foundry. Neither is tested indirectly, both are assumed.
AI-500 cost, beta status and what Microsoft has not published
This is where you need to be careful with what you read elsewhere. Several prep sites quote a firm US price and a firm general availability date. Microsoft's own exam page does neither.
Here is what is confirmed and what is not:
| Detail | Status |
|---|---|
| Passing score | Confirmed: 700 |
| Domain weightings | Confirmed: four domains, bands as listed above |
| Prerequisite exam | Confirmed: none |
| Retirement date | Confirmed: none |
| Languages | Currently English only |
| List price | Not published on the exam page, which states only that price depends on the country or region where the exam is proctored |
| Question count and time limit | Not published |
| General availability date | Not published on the exam page. The instructor-led course AI-500T00 is listed as available from 30 September 2026 |
| Practice Assessment | Not yet available. Microsoft says these usually appear within eight weeks of an exam leaving beta |
Beta registration opened in July 2026 at a heavily discounted rate. If you sit a beta exam, your result is not returned immediately, because Microsoft is still gathering data on question quality. Expect a wait of several weeks rather than a score at the test centre.
Budget for the same price band as Microsoft's other role-based exams and confirm the exact figure at checkout before you book. Prices vary by country and Microsoft changes them without notice.
One detail people miss: Microsoft associate, expert and specialty certifications expire after a year. Renewal is a free online assessment on Microsoft Learn, so the annual expiry is an admin task rather than a repeat exam fee.
Where AI-500 sits in the Microsoft AI certification path
Microsoft rebuilt its AI certification ladder through 2026, and AI-500 is now the top of it.
| Exam | Level | Focus |
|---|---|---|
| AI-901 | Fundamentals | Azure AI concepts, no coding |
| AI-103 | Associate | Building AI apps and single agents on Azure |
| AI-200 | Associate | Azure AI cloud developer work |
| AI-300 | Associate | Azure machine learning operations |
| AI-500 | Expert | Production multi-agent systems, orchestration, governance and deployment |
The gap between the associate rung and AI-500 is wider than the gap between fundamentals and associate. AI-103 and AI-200 teach you to ship an AI feature. AI-500 asks how you operate a fleet of agents that make decisions on their own, with an audit trail, a cost ceiling and a rollback path.
Who should sit AI-500, and who should not
Sit it if you are already running agentic workloads in production and want the credential to match. The exam maps closely to real operational work, so if the job is your day job, the prep is mostly formalising what you know.
Sit it after general availability if you want a scored result quickly, official practice assessments, and a settled question pool. Beta exams reward early movers with a lower price and a longer wait.
Skip it for now if you are building your first agent, if Python is not comfortable, or if you have not worked in Microsoft Foundry. AI-103 or AI-200 is the correct next step, and both have a far denser body of prep material behind them.
Skip it entirely if you want an AI certification for a CV rather than for a role. An expert-level exam on multi-agent orchestration is a poor way to signal general AI literacy, and AI-901 does that job in a quarter of the time.
Frequently Asked Questions
What is the AI-500 exam?
AI-500 is Microsoft's exam titled Designing and Implementing Multi-Agent AI Solutions. It tests architecture, development, evaluation, security and deployment of production multi-agent AI systems on Azure using Microsoft Foundry. Passing it earns the Microsoft Certified: Multi-Agent AI Solutions Expert certification on its own, with no second exam required.
Are there prerequisites for AI-500?
No. Microsoft lists no prerequisite certification and no related exams for AI-500. The practical bar is the audience profile: Python proficiency, production experience with agentic systems, and hands-on work with Microsoft Foundry, MCP, RAG and LangGraph.
How much does the AI-500 certification cost?
Microsoft has not published a list price on the AI-500 exam page. It states only that the price depends on the country or region in which the exam is proctored. Beta candidates paid a discounted rate. Check the figure at checkout on Microsoft Learn before booking, since prices vary by country and change without notice.
Is Microsoft AI certification free?
The exams are not free, but the training is. Microsoft Learn hosts the full learning paths for AI-500 at no cost, and certification renewal is a free online assessment rather than a paid retake. The only unavoidable cost is the exam fee itself.
What is the passing score for AI-500?
700 out of 1000, the same as every Microsoft role-based exam. The score is scaled, so it does not correspond to answering 70% of questions correctly.
Ready to Start Practising?
AI-500 rewards people who have already done the work, which makes the associate rung beneath it the highest-value place to spend study time if you are not there yet. CertCrush has full practice question banks and study guides for AI-200 and AI-103, the two exams that build the Azure AI foundation AI-500 assumes you have.
Create a free CertCrush account and start with timed practice on the exam you are actually ready to sit.