AB-100 Plan AI solutions explained
The Plan AI-powered business solutions domain is 25–30% of AB-100 and tests whether you can decide what to build before anyone builds it. It has three parts: analysing requirements and data, designing the overall AI strategy, and evaluating costs and benefits. The questions describe a business situation and ask which approach, platform or investment case is the most sensible.
Analyse requirements
The official outline asks you to assess where agents help with task automation, data analytics and decision-making, to review data used for grounding, and to organise business data so other AI systems can use it.
Fit for an agent
Not every process wants an agent. A useful test is to look at the work itself:
| Signal | Points to |
|---|---|
| Repetitive, rule-based, well-defined steps | Task automation, possibly a task agent |
| Judgement needed over many documents or records | Retrieval-grounded assistance |
| Event-driven and ongoing, with limited human input | Autonomous agent, with guardrails |
| One-off questions answered from a knowledge base | Prompt-and-response agent |
Grounding data quality
The outline names five properties for grounding data: accuracy, relevance, timeliness, cleanliness and availability. A common scenario gives an agent that answers incorrectly and asks for the first fix. Usually the answer is the data (outdated policy documents, duplicated records, content the agent cannot reach), not a larger model.
Design the overall AI strategy
This is the longest list in the domain. The ideas that matter most:
- Adoption process. The outline refers to the AI adoption process in the Cloud Adoption Framework for Azure: a staged approach from strategy and planning through readiness, adoption and governance.
- Multi-agent solutions. You design across platforms such as Microsoft 365 Copilot, Copilot Studio and Microsoft Foundry. The skill is choosing where each agent lives and how they cooperate.
- Prebuilt agents. Start with what exists. Build custom only when a requirement is not met.
- Build custom or extend Microsoft 365 Copilot. If users already work in Microsoft 365 and need company data and actions inside it, extending Copilot is usually lighter than a new standalone agent.
- Custom models. Create a custom model only when prompting, grounding and a catalogue model cannot meet accuracy, cost or latency needs. Small language models can serve narrow, high-volume tasks.
- Prompt library and guidelines. A shared library of tested prompts, with naming, versioning and ownership, so teams do not reinvent them.
- AI Center of Excellence. A central team for standards, reusable assets, training and governance, sitting next to the business units.
- Dynamics 365 apps. Solutions may combine several Dynamics 365 apps, so plan the data and agent boundaries between them.
Evaluate costs and benefits
The exam wants decisions with numbers behind them.
- ROI criteria include benefits such as time saved, error reduction and revenue effect, set against total cost of ownership: licences, consumption, build effort, integration, support and governance.
- Build, buy or extend compares those costs and the time to value for each route.
- Model router. The outline mentions implementing a model router that sends each request to the most suitable model. The idea is to use a cheaper or faster model for simple requests and a stronger one only when needed.
Example questions
Question 1. A finance team asks for an agent to answer policy questions. Pilot answers are often wrong, and review shows the knowledge source includes three superseded versions of the same policy. What should you recommend first?
- A. Switch to a larger language model
- B. Train a custom model on the policy library
- C. Clean the knowledge source so only current policy versions remain
- D. Add a prompt telling the agent to be careful with policies
Show answer
Answer: C
The failure comes from grounding data quality, specifically timeliness and cleanliness, so removing or retiring superseded versions fixes the cause. A larger model still retrieves wrong content, a custom model is a heavy answer to a content problem, and a stricter prompt cannot tell which document version is current.
Want more questions like this? Full AB-100 practice tests →
Question 2. Users already work all day in Microsoft 365 Copilot. They need it to use data and perform actions from the company's order system. A standalone portal would be a new tool for them to learn. Which approach fits best?
- A. Extend Microsoft 365 Copilot with an agent that reaches the order system
- B. Build a separate web portal with its own agent
- C. Train a custom model on the order data
- D. Wait for a prebuilt agent that covers order management
Show answer
Answer: A
Extending Microsoft 365 Copilot puts the capability where users already are, which minimises adoption effort. A new standalone agent adds a place to go, building a custom model solves a problem nobody described, and waiting for a prebuilt agent ignores the stated need for company-specific data and actions.
Want more questions like this? Full AB-100 practice tests →
Question 3. An agent handles a mix of simple FAQ-style requests and a smaller number of complex reasoning requests. Costs are higher than forecast because every request uses the most capable model. What should you add?
- A. A larger prompt library
- B. An even more capable model for every request
- C. Removal of the knowledge source
- D. A model router that selects a model per request
Show answer
Answer: D
A model router sends each request to the most suitable model, so simple requests use a cheaper model and only complex ones use the strongest. Prompt libraries improve consistency, not cost per request. A stronger model would raise costs. Removing the knowledge source would hurt accuracy without addressing model choice.
Want more questions like this? Full AB-100 practice tests →
How to revise this domain
Practise short written justifications. For any scenario, write one sentence for the choice and one for the rejected alternative. The exam rewards seeing why the obvious answer, such as a custom model, is wrong when a cheaper step is available. Then continue to Design explained.