AI-103 and AI-300: a guide and my take
I sat two new Microsoft certifications back to back this month: AI-103 and AI-300. Passed both on the first attempt, about two weeks of prep behind each. This is half guide - what the exams actually cover, how I prepared - and half a straight account of what I think about them, which parts of my own preparation I’d change, and where I’m genuinely unsure I did well.
What these two certs actually are
The naming trips people up, so let’s clear it up first.
AI-103, Developing AI Apps and Agents on Azure, is the exam behind the Azure AI Apps and Agents Developer Associate certification. It’s the builder’s exam: planning an Azure AI solution, then implementing generative AI and agentic solutions, computer vision, text analysis, and information extraction on top of it.
AI-300, Operationalizing Machine Learning and Generative AI Solutions, is the exam behind the Machine Learning Operations Engineer Associate certification. It’s the operator’s exam: the full machine learning model lifecycle, designing and implementing GenAIOps infrastructure, and generative AI quality assurance and observability once something is already built. Honestly, the certification name is a bit misleading - “Machine Learning Operations” reads like it’s about classic MLOps alone, but in practice the exam splits its weight evenly across both worlds: traditional ML model lifecycle and pipelines on one side, and GenAIOps - prompts, agents, generative AI quality and observability - on the other. It’s not a GenAIOps exam with a legacy MLOps section bolted on; both halves carry real weight.
Put simply - AI-103 asks “can you build it,” AI-300 asks “can you run it once it’s live.” Different exams, but they sit on the same underlying platform (Microsoft Foundry), which is exactly why I ended up sitting them together.
Skills measured
AI-103 - Developing AI Apps and Agents on Azure
| Domain | Weight |
|---|---|
| Implement generative AI and agentic solutions | 30-35% |
| Plan and manage an Azure AI solution | 25-30% |
| Implement computer vision solutions | 10-15% |
| Implement text analysis solutions | 10-15% |
| Implement information extraction solutions | 10-15% |
AI-300 - Operationalizing Machine Learning and Generative AI Solutions
| Domain | Weight |
|---|---|
| Implement machine learning model lifecycle and operations | 25-30% |
| Design and implement a GenAIOps infrastructure | 20-25% |
| Design and implement an MLOps infrastructure | 15-20% |
| Implement generative AI quality assurance and observability | 10-15% |
| Optimize generative AI systems and model performance | 10-15% |
(Full domain lists and sub-skills, straight from Microsoft’s official study guides: AI-103, AI-300. Microsoft does revise these periodically, so check the current version before you plan your prep around these numbers.)
What the exam actually feels like
I sat both exams online through Pearson OnVUE, proctored the normal way - nothing unusual there, but worth naming for anyone who hasn’t sat a Microsoft exam remotely before.
The skill bullets above are the syllabus, but they undersell how the questions are actually framed. Read through either study guide and almost every bullet is scenario language - “implement a solution that…”, “choose the appropriate…”, “build a solution that analyzes…” - not “define X” or “name the parameter for Y.” That carries straight through into the exam. In my sitting, the large majority of questions were scenario-based: a short business or technical situation, then a decision to make, rather than a bare recall question.
Both exams also included a case study - a longer scenario with a block of questions attached to it rather than one-off items - and each case study I got had 6 questions tied to it. Total question count: 63 for AI-103, 60 for AI-300. Microsoft doesn’t publish exact question counts or case-study structure anywhere in the study guides (both can vary between sittings and exam versions), so take these as a real data point from my own exam rather than a guaranteed number.
How I actually prepared
Base layer was the two official Microsoft Learn video series - one for AI-103, one for AI-300 - plus the matching Microsoft Learn training paths and modules for each. The standard route. Nothing unusual there.
Those paths aren’t just reading, either - they’re heavy on labs, and I did most of them rather than skimming past to the module quiz. That mattered less as “exam prep” and more because it lined up with what I already do day to day: my actual job runs mostly around Azure and Foundry, so a lot of this wasn’t new hands-on time so much as it was putting a name and a structure to things I already work in regularly.
Once I’d covered the material, instead of reaching for one of the many generic exam-dump or practice-test sites out there, I built my own question banks - small HTML artifacts, generated with Claude - targeted specifically at the areas I knew I was weakest on. I never touched the generic practice sites. The tailored version let me keep narrowing in on exactly what I was shakiest on, instead of grinding through questions on things I’d already got solid.
My take
The syntax-question problem
Both exams leaned harder on literal SDK/syntax recall than I expected - “what does this specific call look like,” not just “what’s the right approach.” I’m genuinely not sure I did well on that slice of either exam.
I think this is worth naming honestly, because it’s not really a “study harder” problem - it’s a GenAI-era one. Day to day, I don’t hand-write every code block anymore. I architect around what’s available and let the model fill in the exact syntax. That’s the right way to work now, but it means the specific muscle these exams test - precise recall of SDK calls without assistance - is exactly the muscle that gets less exercise. Worth knowing going in, so you can deliberately drill it rather than assume your day-to-day fluency will carry over.
Sit them together
If you’re doing both, I’d recommend sitting AI-103 and AI-300 close together rather than spacing them out. There’s real overlap - same platform, overlapping concepts around agents and generative AI - and studying one reinforces the other while it’s still fresh. That said, neither one is a walk in the park on its own; “do them together” is about efficiency, not about either exam being easy.
Prompt versioning: fine for one agent, not for a fleet
One thing from the AI-300 material I want to push back on a little: the GenAIOps guidance leans on file-based prompt versioning - prompts committed alongside code, versioned through GitHub. That’s fine for a single, light agent. It gets tricky fast once you’re running multi-agent orchestration - many prompts, many agents, and you want to iterate on one without redeploying the rest.
What I’d actually reach for instead is a managed, lightweight prompt library: prompts maintained as their own artifact, fetched at runtime rather than baked into a deploy. That also makes prompts something non-technical people can review and edit directly, and it ties much more naturally into custom evaluation interfaces - you can swap a prompt version and re-run evals against it without touching code at all. File-based versioning is the path of least resistance, not necessarily the right one once the system gets past a single agent.
That said, a managed prompt library isn’t free. It’s another service to run, another place a deploy can drift from what’s actually live, and now prompt changes can ship outside your normal review and rollback path entirely - which is exactly the safety net file-based versioning gives you for free. So this isn’t a strict upgrade, it’s a trade-off, and which side wins depends on the project: a handful of agents with infrequent prompt changes, stick with files; a fleet of agents with prompts iterated on constantly by people outside the codebase, that’s when the extra infrastructure starts paying for itself.
What’s next
Next up on the build side: putting the managed-prompt-library idea into practice on my own multi-agent work.