AB-100 Practice Test

Sample AB-100 Practice Questions

10 real questions from the free AB-100bank. Pick an answer in your head, then open “Show answer” for the explanation.

1. 1. A customer-service agent built on Copilot Studio is grounded against a SharePoint knowledge base containing both a retired 2022 returns policy and the current 2025 policy. Agents have begun citing the retired document about 8 percent of the time. Which mitigation most directly addresses the accuracy risk before relying on the agent in production?

   • A. Increase the temperature of the underlying model so it generates more varied responses and is less likely to repeat the wrong document each time.
   • B. Train a custom small language model on the current policy and skip grounding entirely so the model itself becomes the source of policy answers.
   • C. Remove or archive the retired 2022 policy from the indexed source and add a freshness filter on the knowledge source so only the current version is retrieved.
   • D. Add a topic in Copilot Studio that tells the agent to prefer the latest information via a system prompt instruction on each turn of the conversation under the team's current assumption.
   Show answer

   Correct answer: C

   Accuracy issues in grounded agents are usually data-source issues: stale content in the index produces stale citations. Removing the retired document and adding freshness filtering at the knowledge-source level is the correct architectural fix. Higher temperature (A) worsens accuracy. Custom SLMs (B) require retraining for each policy change. System prompt instructions (D) are unreliable when conflicting evidence is in the index.

2. 2. A regulated lender wants an agent to recommend credit-line increases up to $50,000 based on customer payment history. Regulators require that every decision is reproducible, auditable, and explainable to a supervisor. Which agent strategy aligns with these constraints?

   • A. A fully autonomous generative agent that issues credit decisions, with explanations produced from the model's output for each case raised.
   • B. A task agent that applies a documented, deterministic scoring rule, augmented by a generative summary explaining the decision to the customer.
   • C. A custom engine agent in Microsoft 365 Copilot that asks the underwriter conversational questions and decides based on free-text input each time.
   • D. A prompt-and-response agent that lets each underwriter improvise their own credit logic with AI assistance each time they evaluate a case.
   Show answer

   Correct answer: B

   Regulated decisions need deterministic, auditable rules - a task agent applying documented scoring is reproducible. The generative layer only summarises the decision rather than produces it. Option A makes the LLM the decision-maker, breaking reproducibility. Options C and D introduce variability that fails audit and explainability tests required by regulators.

3. 3. A Dynamics 365 Sales customer wants sellers to get account briefings, opportunity summaries, and email assistance directly inside Outlook and Teams without building a custom agent. Which Microsoft option meets the need with minimal custom build?

   • A. Build a custom Copilot Studio agent from scratch and integrate it with Outlook through a custom add-in for sellers.
   • B. Adopt Microsoft 365 Copilot for Sales (the prebuilt sales agent experience) and configure it for the Dynamics 365 environment.
   • C. Build a Foundry agent with custom code and embed it in Outlook through a manifest-based extension package.
   • D. Skip agents and write a Power Automate flow that emails sellers each morning with their account briefings instead.
   Show answer

   Correct answer: B

   Microsoft 365 Copilot for Sales is the prebuilt sales agent surfaced in Outlook and Teams, grounded on Dynamics 365 or Salesforce. It delivers the requested capabilities without custom build. Custom Copilot Studio (A) or Foundry (C) reinvent existing prebuilt functionality. Email flows (D) ignore the conversational requirement.

4. 4. An architect is organizing business solution data so that one customer-service agent today and several future agents (returns, warranty, billing) can reuse the same product, order, and customer records. Which organizing pattern best supports reuse across multiple AI systems?

   • A. Duplicate the product, order, and customer tables into each agent's private knowledge source so each agent owns and controls its own copy of the data.
   • B. Have each agent call the source operational system through bespoke point-to-point connectors with no shared metadata layer between any of them.
   • C. Store everything in a single Foundry vector store that the agents query directly, ignoring the operational systems of record where the data originated.
   • D. Centralize the canonical records in Microsoft Dataverse and Dynamics 365. Then expose them to each agent as shared knowledge sources or actions.
   Show answer

   Correct answer: D

   Reuse across AI systems requires a shared canonical layer with governed access. Microsoft positions Dataverse and Dynamics 365 as the operational data platform that agents ground against. Option A drifts and duplicates ownership. Option B creates per-agent connector sprawl with no shared semantics. Option C makes a vector store the system of record, which it is not.

5. 5. A retailer's AI strategy targets two outcomes: cut contact-centre handle time and increase store-manager productivity on inventory tasks. The architect proposes one massive autonomous agent that does everything for both audiences. Why is this the wrong agent strategy?

   • A. Autonomous agents in Copilot Studio cannot be deployed in retail organisations at the present time.
   • B. Microsoft does not support autonomous agents that span multiple audiences inside the same tenant configuration.
   • C. Copilot Studio cannot host two agents in one tenant, so a single agent is required to satisfy both audiences.
   • D. Different audiences, data sources, and outcomes are better served by purpose-scoped agents (one service, one store-manager) so each can be tuned, governed, and measured against its outcome.
   Show answer

   Correct answer: D

   Agent strategy aligns scope to outcome and audience. A monolithic agent obscures measurement, dilutes tuning, and complicates governance. Purpose-scoped agents per outcome (service, store ops) match audiences and metrics. The other options state platform limits that do not exist.