In May 2024 I sat exam AI-102 and earned the Microsoft Certified: Azure AI Engineer Associate credential — the hands-on, build-it counterpart to the AI-900 fundamentals I'd passed a few months earlier. I'm Thimo Buchheister, founder and CEO of ThreeBIT GmbH in Ibbenbüren, and where AI-900 asked whether I could describe AI on Azure, AI-102 asked whether I could actually implement it: provision the services, wire them into an application through SDKs and REST APIs, ground a model in a customer's own data, and stand behind the result in production.
What Azure AI Engineer Associate is
This is an associate-level, role-based certification — Microsoft pitches it at the Azure AI engineer role, not at someone who needs a vocabulary briefing. The whole posture is different from fundamentals. Where AI-900 has no prerequisites and explicitly welcomes non-technical people, AI-102 assumes you can already write software.
Microsoft's audience profile is blunt about it: as an Azure AI engineer you "build, manage, and deploy AI solutions that leverage Azure AI," and you participate in all phases of that work — requirements and design, development, deployment, integration, maintenance, performance tuning and monitoring. You're expected to have experience developing in Python or C#, and to be comfortable using REST APIs and SDKs to build "secure image processing, video processing, natural language processing, knowledge mining, and generative AI solutions on Azure." You should understand the Azure AI portfolio and its data-storage options, and be able to apply responsible-AI principles in practice rather than recite them.
In other words: it's a credential for the person who has to make the thing work, ship it, and keep it running — which, in a small company, is frequently me.
It's also a credential with a clock on it. Microsoft's associate, expert and specialty certifications expire after 12 months and must be renewed — though renewal is free, done through an online assessment on Microsoft Learn rather than by re-sitting the full proctored exam. I'll come back to why I think that annual renewal is a feature, not a chore. (One note for accuracy: Microsoft has announced that the AI-102 exam and this certification path will retire on 30 June 2026. My credential stands; I'm writing about the exam as I earned it.)
What it actually certifies
The exam doesn't ask you to recognise AI workloads. It asks you to implement them across six skill areas, and the weightings show where Microsoft puts the centre of gravity for an AI engineer today:
- Plan and manage an Azure AI solution (20–25%) — selecting the right Azure AI service for a given problem (generative AI, vision, language, speech, information extraction, knowledge mining); creating and deploying AI resources; installing the right SDKs; determining endpoints; wiring AI into a CI/CD pipeline; monitoring, cost management, and — critically — securing the solution: protecting account keys, managing authentication, and implementing AI responsibly with content moderation, content filters, blocklists, prompt shields and harm detection.
- Implement generative AI solutions (15–20%) — the headline domain. Provisioning and deploying an Azure OpenAI resource, selecting and deploying the right model, submitting prompts to generate code and natural-language responses, generating images, and — the part that matters most in real projects — implementing a RAG pattern by grounding a model in your own data. It also covers prompt flow, prompt templates, evaluating models and flows, applying prompt-engineering techniques, controlling generative behaviour through parameters, and fine-tuning.
- Implement an agentic solution (5–10%) — building custom agents, including with the Azure AI Foundry Agent Service, and orchestrating multi-agent workflows. This domain was added as the syllabus moved with the field.
- Implement computer vision solutions (10–15%) — analysing images (object detection, tagging, OCR, reading handwriting), training and publishing custom vision models, and extracting insights from video.
- Implement natural language processing solutions (15–20%) — key-phrase and entity extraction, sentiment, language detection, PII detection, translation, text-to-speech and speech-to-text, and custom language and question-answering models.
- Implement knowledge mining and information extraction (15–20%) — building Azure AI Search solutions with indexes, indexers, custom skills, and semantic and vector search; and document intelligence — using prebuilt and custom models to extract structured data from documents.
The thing worth dwelling on is that generative AI is now a first-class, examined engineering discipline here — not a bullet point. When the AI-900 fundamentals syllabus was refreshed, generative AI became its largest describe-it section. AI-102 is the corresponding build-it commitment: it expects you to stand up Azure OpenAI, ground it in real data with RAG, evaluate it, and operationalise it responsibly. That progression — from "what is generative AI" to "deploy it, ground it, secure it" — is exactly the journey our customers have been on, and exactly why I wanted both credentials.
On the mechanics: it's a proctored exam with 100 minutes to complete, the passing score is 700 on Microsoft's scaled 1,000-point scale, and it's offered in ten languages including English and German. If you fail, you can retake after 24 hours. It is, in every sense, a serious exam.
Why I bothered to certify
I run a software company. I could — and do — read the documentation. So why book a proctored, associate-level engineering exam as the founder rather than delegating it to a developer?
Three honest reasons.
First, AI-900 left a gap that only a hands-on exam could close. Fundamentals gave me the map: the vocabulary, the responsible-AI principles, a sober sense of which workload wants which service. But a map isn't the territory. There's a real difference between knowing that RAG grounds a model in your data and having actually provisioned the search index, written the retrieval step, and watched what happens when the model confidently cites a document that doesn't say what it claims. AI-102 forced me through the territory — the SDKs, the deployment options, the failure modes — not just the legend on the map.
Second, credibility that's specific, not generic. "We do AI" is noise; every shop says it. Being able to point to a Microsoft credential that certifies I can design and implement an Azure AI solution — Azure OpenAI, Azure AI Search, the responsible-AI tooling — is a concrete, verifiable signal to a customer who is, rightly, sceptical of AI hype. It's the same instinct behind every badge we hold: show it, don't ask to be taken on faith.
Third — the founder's reason — I won't ask my team to climb a ladder I'm exempt from. If certification is part of how ThreeBIT demonstrates competence, the person at the top belongs on the same rungs. Earning the associate credential myself, exam pressure and all, is a cheap and honest way to keep that real. And the annual renewal keeps it honest over time: in a field that re-draws its own map every few months, a credential that quietly expires unless you re-engage every year is precisely the right shape. It's not a trophy on a shelf; it's a standing commitment to keep up.
How it shows up in our work at ThreeBIT
This is where the syllabus stops being abstract. ThreeBIT builds software for industries where a defect isn't a cosmetic glitch — it's a missed export, a failed payment, or a compliance finding. Our customers are largely German and frequently regulated, and they increasingly want AI inside those systems: a document that classifies and extracts itself, a search box that understands intent rather than keywords, a support flow that drafts a grounded reply. The interesting question is never "can it?" — it's "should it, and can we prove it behaves?"
That's exactly the territory AI-102 certifies, and here's how the domains map onto what we actually build:
- Generative AI with grounding (RAG), not free-floating chat. When we put a language model near a customer's workflow, we don't let it improvise from its training data — we ground it in the customer's own documents with retrieval, so answers trace back to a source the customer controls and can audit. The exam's RAG and Azure AI Search objectives aren't theory to me; they're the default architecture for any AI feature that has to be trustworthy.
- Document intelligence on real paperwork. German business runs on documents — invoices, delivery notes, contracts, forms. The document-intelligence and knowledge-mining skills map straight onto extracting structured, validated data from that paperwork instead of asking a human to re-key it, with the model's output checked rather than trusted blindly.
- Computer vision where it earns its place. Not every problem wants a large language model. AI-102's vision domain — and the discipline of choosing classical vision or a custom-trained model over a generative one when that's the better, cheaper, more auditable fit — makes me a better steward of a customer's budget and a better judge of which tool actually belongs.
- Data staying in-tenant. This is the non-negotiable one. For a regulated German customer, where the data goes is the whole conversation. The exam's emphasis on planning, securing and managing the solution — key protection, authentication, content filtering, monitoring — is the part that lets us build AI that keeps a customer's data inside their own Azure tenant, under their governance, rather than shipping it off to somewhere we can't account for.
That last point is the real return on this certification. AI-900 made me fluent in the why of responsible AI. AI-102 made me competent in the how of building it — provisioning Azure OpenAI, grounding it in tenant data, securing the keys, filtering the output, monitoring the behaviour. For the customers we serve, that combination isn't a nice-to-have. It's the difference between an AI feature they can put in front of an auditor and one they can't risk turning on.
Foundations told me what good looks like. The engineering credential is the part that lets me ship it — the first time, for people who can't afford for it to go wrong.
Sources & further reading
- Microsoft Learn — Microsoft Certified: Azure AI Engineer Associate (certification overview, audience profile, renewal, retirement notice): https://learn.microsoft.com/en-us/credentials/certifications/azure-ai-engineer/
- Microsoft Learn — Study guide for Exam AI-102: Designing and Implementing a Microsoft Azure AI Solution (skills measured, domain weightings, passing score, change log): https://learn.microsoft.com/en-us/credentials/certifications/resources/study-guides/ai-102
- Microsoft Learn — Certification renewal (associate/expert/specialty certifications expire annually; free online renewal): https://learn.microsoft.com/en-us/credentials/certifications/renew-your-microsoft-certification
- Microsoft Learn — Course AI-102T00-A: Develop AI solutions in Azure (C#/Python, Azure AI Foundry, REST/SDK): https://learn.microsoft.com/en-us/training/courses/ai-102t00
- Microsoft Learn — Microsoft Certified: Azure AI Fundamentals (AI-900) (the foundations counterpart): https://learn.microsoft.com/en-us/credentials/certifications/azure-ai-fundamentals/
Image credits
All images are used under their respective Creative Commons or public-domain terms; we are grateful to the creators.
- Electronic Bible publishing station — © kahunapulej, CC BY-SA 2.0, via Flickr (source).
- Game Development — © qubodup, CC BY 2.0, via Flickr (source).
- Team Meeting — © Startup Stock Photos, CC CC0 1.0, via Stocksnap (source).