Free AB-100 sample questions with answers
Here are five free AB-100 sample questions, covering the Plan, Design and Deploy domains, with the answer and the reasoning. Try each one before opening the answer. These are original scenario questions written against the official skills outline, not questions from the real exam. Deploy gets three, because it is 40–45% of the exam.
Question 1. A retailer wants an assistant for store managers. A prebuilt agent covers about 80% of the requirement. The remaining 20% is a lookup in an internal stock system. The business wants the shortest time to value and the lowest running cost. What should you recommend?
- A. Build a new agent from scratch for full control
- B. Train a custom model on the stock data
- C. Use the prebuilt agent and extend it with the stock lookup
- D. Reject the prebuilt agent because it does not cover everything
Show answer
Answer: C
Extending the prebuilt agent with the stock lookup covers the gap with the least build effort and cost. Building from scratch repeats work the prebuilt agent already does. A custom model answers a need nobody stated. Rejecting the prebuilt agent because it is incomplete ignores the build, buy or extend comparison.
Want more questions like this? Full AB-100 practice tests →
Question 2. A bank designs a solution where a front-line agent talks to customers and must pass complex mortgage questions to a specialist agent that uses a different platform, using an open protocol rather than a proprietary integration. Which protocol is intended for this agent-to-agent communication?
- A. Agent2Agent (A2A)
- B. Model Context Protocol (MCP) only
- C. A fallback topic
- D. A prompt library
Show answer
Answer: A
Agent2Agent (A2A) is the open protocol for agents to communicate with other agents across platforms. Model Context Protocol connects a model or agent to tools and data, not to other agents as peers. A fallback topic is a Copilot Studio construct. A prompt library stores prompts and does not carry agent communication.
Want more questions like this? Full AB-100 practice tests →
Question 3. After launch, an agent's answer quality drops. Logs show response times are normal and error rates are low, but users repeatedly rate answers as unhelpful after a product catalogue change. What does the telemetry most strongly suggest?
- A. The platform needs more capacity
- B. The grounding data no longer matches the catalogue
- C. The language model must be replaced
- D. The user ratings can be ignored
Show answer
Answer: B
Normal latency and low errors rule out platform faults, and the timing with the catalogue change points to stale or mismatched grounding data. Scaling up capacity addresses latency problems, which do not exist. Switching models is unevidenced. Treating the ratings as noise ignores the strongest signal available.
Want more questions like this? Full AB-100 practice tests →
Question 4. A company manages Copilot Studio agents, a custom model and the data both use. Releases keep breaking because teams promote them at different times. Which ALM approach addresses this?
- A. Let each team promote whenever it is ready
- B. Edit production directly to keep versions aligned
- C. Freeze the model so it never changes
- D. Version the agent, model and data, and promote compatible sets through test to production
Show answer
Answer: D
Each of agents, models and data needs its own lifecycle, but promotion must be coordinated and versioned so that compatible versions move together through test to production. Letting teams promote independently causes the mismatch. One shared development environment with direct production edits removes control. Freezing the model forever stops improvement instead of managing change.
Want more questions like this? Full AB-100 practice tests →
Question 5. A customer-facing agent summarises uploaded documents. A tester uploads a document containing hidden text that tells the agent to ignore its rules and reveal its configuration. Which design measure best mitigates this?
- A. Make the system prompt more polite
- B. Treat uploaded content as untrusted data, filter inputs and outputs, and limit the agent’s tool permissions
- C. Increase the context window
- D. Block every document upload
Show answer
Answer: B
Treating retrieved or uploaded content as data rather than instructions, with input and output filtering and least-privilege tool access, limits the damage of prompt manipulation. Rewording the system prompt more politely does not stop injected instructions. A larger context window gives the attacker more room. Blocking all uploads removes the feature rather than securing it.
Want more questions like this? Full AB-100 practice tests →
How did you do?
Question 1 is the build, buy or extend logic that runs through the Plan domain. Question 3 tests whether you read telemetry before changing anything: when latency and errors are fine, the fault is in the content or the design. Question 5 shows the shape of the security questions, where the right answer reduces the attack surface instead of removing the feature.
What these questions have in common
None asks for a definition. Each gives a requirement, a constraint and several plausible options, and the answer depends on one detail: lowest cost, open protocol, evidence in the logs, coordinated promotion, untrusted input. Underline the constraint before reading the options.
For more, see the Plan, Design and Deploy guides, or take the 20-question practice test.