AI-103: Azure AI Apps and Agents Developer — What Changed from AI-102

If you earned or studied for AI-102 (Azure AI Engineer Associate) before 2026, the exam number you see on the schedule now — AI-103 — looks like a typo. It is not. Microsoft retired AI-102 and replaced it with AI-103: Azure AI Apps and Agents Developer. The subject shifted meaningfully, and if you show up to the new exam expecting the old one you will feel it on question one.

This is the post I wish I had before sitting the new exam.

---

What Changed — In One Sentence

AI-102 was about building AI solutions on Azure: pick a model, deploy it, call it, monitor it. AI-103 is about building applications and agents that use AI — with a much heavier emphasis on agent frameworks, tool-calling, evaluation, and the operational reality of shipping something that does not hallucinate in front of a customer.

The migration guide from Microsoft puts it this way: AI-103 measures whether you can "design, build, evaluate, and deploy AI applications and agents that are grounded, performant, and safe." That sentence is doing a lot of work. Let me unpack it.

---

The Five Places AI-103 Is Not AI-102

1. Agents Are Now Core, Not a Chapter

On AI-102, "agent" was a chapter near the end. On AI-103, it is the spine of the whole exam. You are expected to understand:

Concept	What you need to know
Agent architecture	The loop: plan → call tool → observe → respond, and where each step lives in code
Tool definitions	How to describe a tool to a model so it actually calls the right one — unambiguous function schemas
Orchestration patterns	Single-agent vs multi-agent: when to reach for each, and what you pay in latency and cost for each
Grounding	Retrieval-Augmented Generation (RAG), citations, and the difference between a model that looks confident and one that is tethered to source material

A minimal agent loop in code looks like this:

textUser asks a question
  ↓
Model decides: can I answer from my own knowledge?
  ├── Yes → generate the answer
  └── No  → choose a tool (search, database, API, …)
                ↓
           Call the tool with arguments the model produced
                ↓
           Read the tool result
                ↓
           Model incorporates the result into the final answer

If you worked with Azure AI Studio's agent capabilities or the newer agent SDK surfaces, this is familiar territory. If your AI-102 knowledge stopped at "deploy a model to an endpoint and call it with REST," this is the biggest gap.

2. Evaluation Is Explicitly on the Exam

AI-102 touched evaluation lightly — mostly "did you set up monitoring?" AI-103 expects you to know how to measure whether an AI application is any good before you ship it:

• RA GSL (generative AI evaluation) metrics: groundedness, relevance, coherence, fluency, similarity — and when each one matters.
• Adversarial and safety evaluation: what you test for when the cost of a bad answer is embarrassment vs. when the cost is a compliance incident.
• Programmatic evaluation vs. human evaluation: when you can automate and when a human in the loop is the only acceptable answer.

This is the part of the exam where people who only know how to call an endpoint tend to fail. You need to know how to tell whether your thing works, not just how to build it.

A simple evaluation harness might look like this conceptually:

For each test query:
  1. Run the agent and capture the answer
  2. Compare the answer against a reference (ground truth or retrieved source)
  3. Score on groundedness, relevance, coherence
  4. Aggregate: pass/fail threshold, average scores, worst-case examples

Report: which queries failed, why, and what would fix them.

3. Safety and Responsible AI Moved From "Nice to Have" to "Examinable"

AI-102 mentioned content filtering. AI-103 goes further:

• Prompt injection and jailbreak patterns: what they look like, categories of injection, and the mitigations that actually reduce risk (not just the ones that look good in a demo).
• Content safety filters: how they are configured, what they catch, and the tradeoff between over-filtering (a helpful answer blocked) and under-filtering (a dangerous one through).
• Responsible AI principles in practice: fairness, reliability, privacy, inclusiveness, transparency, accountability — not as abstract values but as decisions you make when you pick a model, a grounding strategy, and an evaluation plan.

4. The Azure Services Have Been Renamed and Re-shuffled

Several services and SDK surfaces changed names or rolled into new ones. The exam expects you to know the current names and when to use which:

Old / colloquial name	What the exam expects you to know now
Azure OpenAI Service (model deployment)	Azure AI Foundry / model deployment surfaces — the current vocabulary
Cognitive Services (legacy)	Consolidated into Azure AI services; know what moved where
Azure AI Studio (agent workflows)	Agent building, evaluation, and monitoring surfaces

You do not need to memorize every service name change. You do need to know enough to read a scenario and pick the right current service without hesitating.

5. Deployment and Operations Got More Real

AI-103 is less about "deploy a model" and more about "ship and run an AI application":

• Latency and cost tradeoffs: smaller models for cheaper/faster paths, larger models for quality, when to cache, when to stream.
• Monitoring an AI workload: what you monitor (token usage, latency, error rates, evaluation drift) and how that differs from monitoring a regular REST API.
• Grounding and retrieval in production: chunking strategies, index refresh, and the failure modes that show up only after real users start asking real questions.

---

What Did Not Change

Not everything moved. The durable parts of AI engineering on Azure are still on the exam:

• Selecting the right model for the job: knowing when a small model with grounding beats a large model without it.
• Fine-tuning vs. prompt engineering vs. RAG: the decision space is unchanged even if the exam vocabulary around it got sharper.
• Integration patterns: calling models from code, handling responses, retry and error handling, batching where it makes sense.
• Basic Azure plumbing: resource groups, authentication, networking considerations for AI services, keys vs. managed identity.

If AI-102 gave you a foundation, AI-103 builds on it. It does not throw it away.

---

How to Study for AI-103 Now

1. Start with the official learning path for AI-103, not the AI-102 one. The module list will reflect the agent-and-evaluation emphasis.
2. Build one agent end to end. Something small: a tool-calling agent that queries a dataset or API and answers from it. The act of wiring tools, handling the loop, and evaluating the output teaches more than any number of video hours.
3. Run an evaluation. Use Azure AI Studio evaluation or a local harness to measure your agent on groundedness and relevance. The exam will ask you to reason about evaluation results.
4. Read the responsible AI and safety documentation for the current services, not the 2024 versions. The content safety surfaces changed.
5. Practice with recent practice questions. The question style around agents, grounding, and evaluation is new enough that older question banks will mislead you.

---

The Mindset Shift

The cleanest way to think about AI-103 is this:

AI-102 tested whether you could build an AI solution on Azure. AI-103 tests whether you can ship and operate an AI application that a real user can trust.

That shift — from "can you wire it up" to "can you ship something safe and useful" — is the through-line of the whole 2026/2027 certification reset. It shows up in AI-103, in AI-901, and in the new agent track. Once you see it, the exam stops feeling like a collection of facts and starts feeling like a coherent bet about where the job went.

And that bet is: the career is no longer in deploying models. It is in building applications and agents on top of them that do not embarrass you in production.