Another exam

A big change since moving from research to IT has been the focus on external certifications for learning. I’m a big proponent and I’ve written before about taking the AZ-900 and the AI-102. I’ve sat many more in the three-odd years since — from security to FinOps.

The AZ-900 taught me how Microsoft exams work. The AI-102 taught me how to study for them with limited time. The AB-731 — Microsoft’s AI Transformation Leader certification — taught me something different: how to make the AI do the studying with me.

The AB-731 is a newer exam. It targets business decision-makers who are guiding AI transformation within their organisations. No code required. The focus is on recognising opportunities for AI, understanding Microsoft’s AI ecosystem (Copilot, Foundry, Azure AI services), and leading adoption responsibly.

The problem with study material

The AB-731 went generally available in February 2026. That means the ecosystem of third-party study material is thin. A few Udemy courses, some blog posts, a handful of practice test sites. Nothing with the depth I wanted.

With previous exams I relied heavily on practice questions to test my understanding. For the AI-102 I used a combination of Microsoft Learn, Udemy, and John Savill’s exam crams. For the AB-731, the pickings were slim.

So I made my own.

Using AI to generate mock exams

The approach was straightforward. I pointed an AI assistant at the official Microsoft study guide and asked it to generate a practice exam based on the skill areas and weightings it found. I used Claude Code, but any AI tool that can read a URL and produce structured text would work.

The first attempt was functional but had problems. The answer to almost every question was “B”. The question headings gave away answers — “Q19. Copilot in Excel” when the correct choice was Copilot in Excel.

This is where the iterative process became valuable. Each round of feedback made the exams better:

  1. Flatten the format. Replace nested sub-headings with a simple Answer: line. Keep the structure minimal — sections and questions only.
  2. Randomise answer positions. Distribute correct answers roughly evenly across A, B, C, D. No more “when in doubt, pick B”.
  3. Neutral headings. Use “Scenario: Architecture Decision” instead of revealing the answer in the heading.
  4. Increase difficulty progressively. First exam: clear correct answers with obvious distractors. Second: all four options sound plausible. Third: multi-step scenarios requiring synthesis of multiple concepts.

The format

Pick whatever format works for you — markdown, Word, plain text, Org mode. The format doesn’t matter as long as it’s something you can edit and your AI tool can read back. Each question follows a simple structure:

Q1. Scenario: Knowledge Management
A global consulting firm has 10 years of project reports...

A. [Option]
B. [Option]
C. [Option]
D. [Option]

Answer:

You fill in your answers, pass the file back to the AI, and it grades the lot — overall score, per-section breakdown, and explanations for every wrong answer. The whole cycle takes minutes.

I ended up generating three exams of increasing difficulty: 44 questions, 42 questions, 42 questions. 128 questions total, weighted to match the official exam blueprint.

What I learned about the approach

The obvious benefit is volume. 128 tailored practice questions in an evening, with no subscription fees and no dodgy exam dump sites.

But the less obvious benefit is the feedback loop. When I scored 100% on the second exam, I could immediately ask for harder material. When I missed a question about Copilot Chat vs the free Copilot on the first exam, the grading report explained exactly why — and subsequent exams tested that distinction more aggressively.

A few caveats

This approach works well for conceptual exams like the AB-731 where the questions test understanding of principles, architectures, and decision-making. It would be less effective for exams that test specific API syntax or require hands-on labs.

The AI is also generating questions based on its understanding of the study guide, not from a bank of actual exam questions. The questions are representative, not identical. That said, for an exam where the goal is applied understanding, representative is exactly what you want.

The result

I passed with 93%. I don’t think that’s down to the practice exams alone — a lot of the material overlaps with what I’m already doing day-to-day, working with AI tools and thinking about how organisations adopt them. But the targeted prep filled the gaps, particularly around specific Azure AI service capabilities and Microsoft’s frameworks for responsible AI adoption. The two reinforced each other well.

If you’re preparing for a Microsoft certification and want to try this approach, the workflow is simple: give the AI the exam code or study guide URL, ask for practice questions in your preferred format, take the exam, get it graded, ask for a harder version. Repeat until the material sticks.

Preparing for a certification exam? I’d be interested to hear how others are using AI tools in their study workflows.