AB-100 exam tips: how to answer architect questions
The best AB-100 tip is to answer as an architect rather than a builder: identify the requirement and the constraint, then choose the lightest design that satisfies both, and reject answers that add build effort, risk or cost the scenario never asked for. The exam is about judgement across Microsoft’s AI and business-application platforms, so several options usually work; one fits best.
Read the constraint first
Almost every scenario carries a limiting phrase: shortest time to value, lowest running cost, no human in the loop, data must stay in a region, users stay in Microsoft 365. Underline it before reading the options. Then eliminate any answer that breaks it, even if it is technically excellent.
Prefer the lightest option that fits
A recurring pattern across the outline is a ladder from light to heavy:
| Lighter | Heavier |
|---|---|
| Use a prebuilt agent | Build a custom agent |
| Extend Microsoft 365 Copilot | Build a standalone agent |
| Ground and prompt a catalogue model | Create a custom model |
| Use a connector or MCP tool | Use computer use |
| Standard topics or intents | Generative orchestration for everything |
Move to the heavy end only when the scenario says the light end cannot meet a stated need.
Evidence before action
For monitoring and tuning, do not change anything the data does not support. Normal latency with falling quality points at content, not capacity. High escalations on one topic point at that topic’s knowledge. Answers that retrain or replace models without evidence are usually distractors.
Governance is rarely optional
Where a scenario involves sensitive data, assume it needs access control on grounding data, audit trails and residency checks. Where it involves autonomy, assume it needs limits and escalation. An answer that delivers the function but ignores these is usually wrong on the Deploy domain, which is 40–45% of the exam.
Know the vocabulary exactly
| Term | Means |
|---|---|
| A2A | Open protocol for agents to talk to other agents |
| MCP | Open protocol for connecting models and agents to tools and data |
| Grounding | Supplying approved data so the agent answers from it |
| ALM | Lifecycle management: environments, versioning, promotion |
| ROI and TCO | Benefit against total cost of ownership |
| Prompt manipulation | Input that tries to override instructions |
Mixing up A2A and MCP is a cheap mistake and a common one.
Use the outline update
The English outline is dated October 14, 2026. Product names move to Microsoft Foundry and Foundry Tools. If your practice material uses older names, translate them rather than assuming a different service. See the exam format for the update details.
Time and pacing
Microsoft’s exam page does not publish the duration, so check it at booking and plan your pace from the real number. Answer straightforward items first, and flag long scenarios. Never spend long on a single item. For the booking and day-of rules see the exam day guide.
Common traps
- Choosing a custom model when grounding would do.
- Choosing an autonomous agent when approval is required.
- Fixing quality with a larger model before checking the data.
- Editing production directly instead of promoting a tested package.
- Ignoring who is allowed to see the retrieved data.
Practise with the free sample questions and the practice test.