AB-100: Practice Architecture Decisions

AB-100 is an architecture exam, but architecture becomes easier to defend when its assumptions are tested. A diagram showing a Copilot Studio agent, Dataverse and a Foundry model does not prove the agent chooses an authorized source, that a workflow can recover from a retry or that the organization can detect a bad deployment. Small, reversible experiments can make those properties visible without requiring a production rollout.

The current Microsoft AB-100 guide evaluates planning, designing and deploying AI-powered business solutions. These original exercises support that judgment through evidence, not reproduction of live exam questions. Use only authorized sandbox environments, fictional records, scoped identities and a known cleanup procedure. The output of each exercise is a short architecture decision that a reviewer can challenge—not a large unmaintainable demonstration environment.

Set a single reference process and review contract

Use a fictional chain of public sports centers that books facilities, responds to membership queries and approves refunds when an event is cancelled. Staff currently work in a Dynamics or Dataverse-backed case system and share published policy documents. A proposed assistant may explain membership rules, draft a change request and help a supervisor find the right information. It is not authorized to issue refunds or see another venue’s restricted contracts without appropriate business approval.

For every lab record the business request, applicable source, authorized user, permitted action, expected output and negative test. Keep a small evaluation sheet with whether the recommendation was supported, whether authority was enforced, which component failed and what a safe correction would require. Avoid claiming that one successful sample establishes enterprise readiness.

Plan for synthetic dates and records to be clearly labeled as test material. Real customers, payment instruments and confidential contracts are unnecessary for demonstrating the intended architecture properties and should not be copied into prompts or debugging logs.

Exercise 1: replace a vague request with an architecture requirement

The business sponsor asks for “an AI agent that handles cancellations.” Rewrite the request into specific responsibilities: answer policy questions from the current document, validate which venue owns the booking, prepare a refund-review draft and require a supervisor to authorize an actual financial adjustment.

Draw the business flow and identify which decisions remain deterministic. A validated booking ID and the customer’s entitlement to view it should be checked by the application. The model may summarize the policy and draft an explanation. The source policy’s effective date must be preserved. The flow should show what happens when no valid policy exists or a customer asks for a refund outside the authorized process.

Produce a decision record with business outcomes, service owner, data owners, risk, cost assumptions and a measurable completion criterion. An architecture that cannot state what should count as success is not ready for a tooling decision.

Exercise 2: compare prebuilt, extended and custom implementations

Sketch three plausible solutions: a supported Microsoft 365 Copilot experience for policy search, a Copilot Studio agent with an approved venue-data connector and a more specialized Foundry-backed component. For each, state what the application must integrate, which identity it will use, the control over prompts or models, and what the team will need to support after rollout.

Change one constraint: the policy must be combined with a real-time venue booking status before drafting a response. Which option still meets the requirement with a bounded scope? Then change it again: the assistant must interpret an unusual scanned inspection form. Does that justify an additional specialized extraction capability without replacing the rest of the approved business process?

The output is not “custom is better” or “prebuilt is safer.” It is a trade-off table with the decisive requirement, plausible alternatives, expected operational cost, unproven assumptions and a test that could reverse the recommendation.

Exercise 3: make grounding failure visible

Build three synthetic policy files with source identifiers: an outdated venue rule, a current rule and a rule for a different location. The assistant should return the applicable source with an appropriate version reference. Test whether a search query retrieves the right document and whether the generated answer accurately represents the rule.

Distinguish two failure conditions. The correct source can be absent from the search results even when it exists in the data store. Alternatively, the source can be present while the model invents an extra exception. These require different repairs: ingestion and filtering for the first, model instructions or evaluation for the second.

Grounding also needs access control. A policy restricted to facility managers should not reach a public booking assistant merely because it has similar keywords. The data and AI relationship matters when source authority, freshness and permissions determine whether a fluent answer can be used.

End with an evidence note showing the retrieved source IDs, applicability conditions and what the assistant did when the current source was missing or conflicting.

Exercise 4: define an agent tool with real authority

Expose a harmless read-only booking lookup in the sandbox. Require a validated venue and booking identifier and use the requesting person’s actual scope to decide which data may be returned. Test one approved record and an attempt to read a different venue’s private record. The denied request should not become authorized merely because the model generates it in correct JSON.

Next add a draft-refund-review action that creates a synthetic record. Simulate a tool timeout after record creation. If retrying creates a second record, identify the missing idempotency safeguard and specify a stable operation identifier. An apparent agent success after a duplicate write is not acceptable evidence that the workflow behaved correctly.

Finally, require an actual authenticated supervisor action before any financial execution step. A conversational prompt instructing an agent to “confirm first” is not equivalent to a permission check at the destination. Apply zero-trust principles when a service or tool crosses a data or authority boundary.

Exercise 5: test whether another agent improves the system

Create two designs for the same fictional refund case. One uses a fixed sequence of policy retrieval, booking lookup and structured draft. The other allows a bounded coordinator to select from a small number of specialist sources. Keep the user request and expected result constant so the comparison reveals differences caused by orchestration rather than by changed input.

Record quality, number of tool calls, latency, cost and the completeness of source references. A specialist agent that produces a polished summary without retaining its source ID can degrade the final result. A coordinator must not treat a specialist recommendation as authority to issue a refund or to access another venue’s record.

If both approaches meet the business need but the fixed workflow is easier to test and operate, document why additional autonomy is not justified. If the path genuinely depends on information discovered during execution, identify what the bounded agent gains and what controls become necessary. The task is to defend an architecture decision with evidence.

Exercise 6: make business application state enforce the approval

Build a simple Power Apps or comparable authorized sandbox review interface showing the booking, source policy, proposed refund and supervisor decision. The assistant can draft text, but it cannot silently change the status of a denied or unreviewed record.

Test three cases: approved record, pending record and unauthorized employee. Each must show the permitted actions appropriate to its actual state. If a workflow notification fails or is repeated, the application should preserve a consistent record and avoid confusing “email sent” with “refund authorized.”

Write a short state-transition diagram and identify where Dataverse security roles or other application permissions enforce the transition. A convincing agent response is not proof that a business process was correctly executed.

Exercise 7: design a regression gate for multiple asset types

Change the policy document, prompt, model configuration and connected flow one at a time in a training environment. Record which version of each asset was used for a test case. A release may change the assistant’s answer without any new application code, so the architecture’s lifecycle process must cover data and low-code components as well as ordinary software.

Define a small acceptance corpus with a standard refund, an expired booking, a wrong venue, missing evidence, a repeated tool call and an unauthorized execution attempt. Run it before and after the change. Mechanical tests should check schema and prohibited actions; domain reviewers should assess policy interpretation and business appropriateness.

Use appropriate CI/CD and ALM controls to capture approvals, promotion and rollback. A green deployment status does not establish that an assistant remains grounded or that a connector’s permissions were not widened. The project should explain how the team responds when a candidate release fails a high-risk case.

Exercise 8: defend the operating model and economics

Prepare an architecture review for a hypothetical expansion from one venue to twenty. Estimate the drivers of model calls, connector usage, data preparation, human correction and support effort. Which failures become more expensive at scale? What telemetry would reveal wrong policy versions, excessive tool retries or a rising escalation rate?

Describe the monitoring contract: technical request health, source and model versions, tool outcomes, approval events and task-specific quality evaluations. Minimize sensitive record copies in logs. Give each control an owner and a response procedure rather than placing a generic “observability” box on the diagram.

Then present a concise recommendation: proposed system, two rejected alternatives, business value, access controls, limitations, cost assumptions, test evidence and rollback strategy. Another reviewer should be able to challenge the assumptions without rebuilding the entire prototype.

Judge the portfolio by the decisions it can support

An AB-100 practice portfolio is useful if the candidate can explain what a source proves, what a tool may do, why a model or agent pattern was chosen and how a changed requirement affects the recommendation. It is not necessary to create a large production-like environment for every exam objective. Controlled experiments with synthetic data often reveal the key architectural mistake more clearly.

Review the current October 14, 2026 Microsoft guide when deciding whether the portfolio covers planning, design and deployment sufficiently. The documented November 16 revision has a different future weighting, so do not silently use it for a current sitting. The strongest result is a set of small, testable architecture arguments that connect business responsibility with enforceable technical behavior.