AB-620 vs AI-103: which agent exam?
Both exams are about building AI agents on Microsoft technology, and both sit at associate level. The difference is how you build. AB-620 is the low-code path through Copilot Studio and the Power Platform. AI-103 is the code-first path through Microsoft Foundry and Python.
Pick the one that matches the platform your organisation actually uses. They are not a sequence, and neither is a prerequisite for the other.
Side by side
| AB-620 | AI-103 | |
|---|---|---|
| Certification | AI Agent Builder Associate | Azure AI Apps and Agents Developer Associate |
| Build surface | Copilot Studio, Power Platform | Microsoft Foundry, SDKs |
| Primary language | Power Fx, low-code + APIs | Python |
| Largest domain | Integrate and extend agents (40–45%) | Generative AI and agents (30–35%) |
| Distinctive content | MCP, A2A, connectors, computer use, Fabric, ALM | RAG, evaluation, vision, speech, extraction |
| Assumes you know | Dataverse, Power Platform, M365 Copilot | Python, Azure, REST/SDK basics |
| Typical role | App developer, consultant, ISV partner | Azure developer, AI engineer |
What only AB-620 covers
- Copilot Studio specifics: topics, agent flows, generative answers nodes, adaptive cards, variables.
- Enterprise connectors: Copilot connectors, Power Platform connectors, and knowledge sources such as ServiceNow and SAP.
- Computer use — agents that operate an application interface directly when no API exists.
- A2A protocol for agent-to-agent collaboration, and MCP tools.
- Fabric data agents.
- Application lifecycle management: solutions, environment variables, Power Platform Pipelines. This is Power Platform discipline and has no equivalent on AI-103.
What only AI-103 covers
- Deploying and consuming models directly, with deployment and capacity decisions.
- Building RAG pipelines yourself: chunking, indexing, hybrid and vector search.
- Formal evaluation — groundedness, relevance, fluency, coherence.
- Computer vision, speech and information extraction as first-class domains.
- Securing and monitoring the Azure infrastructure underneath.
The overlap
Real, but smaller than it looks. Both expect you to understand agents conceptually — tools, knowledge, orchestration, multi-agent patterns — and both care about responsible AI and monitoring. Both now touch Foundry: AB-620 has an entire section on integrating Copilot Studio agents with Foundry, including the model catalog and Application Insights.
So passing one genuinely helps with the other. It does not come close to covering it.
Which should you take?
Take AB-620 if your organisation runs on Microsoft 365 and the Power Platform, your agents need to reach ServiceNow or SAP, you work with Dataverse, or you are a consultant delivering agents to business users. Low-code is a delivery advantage here, not a limitation.
Take AI-103 if you write Python, your work sits in Azure rather than Microsoft 365, you need control over retrieval and evaluation, or you are building a product rather than an internal solution.
Take both if you are a consultant who needs to meet clients wherever they are. Do AI-103 first if you are a developer, AB-620 first if you come from the Power Platform — in each case you start from the ground you already know.
One practical difference
AI-103 has an official Microsoft practice assessment. AB-620 does not yet: Microsoft states it is “not currently available” and that these usually appear within eight weeks of an exam leaving beta. If you value official practice material as part of your preparation, that is a real, if temporary, point in AI-103’s favour.