Microsoft AB-100: Connecting the Exam Objectives

The AB-100 blueprint separates planning, designing and deploying AI-powered business solutions, but a production system cannot keep those decisions isolated. A business requirement determines which data can be used; source authority constrains the agent’s answer; tool permissions define what the assistant may do; and the way it is tested determines which design risks are visible after release. The same choice can create obligations in all three exam areas.

This article traces those connections through a fictional industrial parts supplier using Dynamics 365, Copilot Studio, Power Platform and Foundry. It is an interpretation of the dependency between objective groups, not a separate study calendar or another list of abstract architecture principles. Microsoft AB-100 currently uses the English guide effective October 14, 2026; the published November 16 update remains future-dated at this checkpoint. The objective groups must be understood through the version relevant to the actual examination.

One customer request becomes an enterprise architecture problem

An equipment distributor sells replacement components to factories. Customer service staff handle installation guidance, warranty eligibility and urgent replacement requests. Different divisions use Dynamics 365 business processes, store signed warranties in a controlled document collection and exchange order updates through an existing ERP integration. The business wants agents to help staff work faster while preserving approval controls for costly replacements.

Define a customer-facing outcome before selecting technology. A service representative should be able to ask whether a particular equipment failure might be covered, review applicable evidence and create a draft return authorization. The system should not expose other customers’ contracts, authorize a refund or create multiple replacement orders because the assistant retries a tool.

These statements become acceptance criteria. An authorized request must identify the right customer and policy, cite source evidence and leave the draft in the correct business state. An unauthorized request must be denied. An unavailable dependency must produce a clear incomplete or review-needed result. The model’s ability to generate persuasive text is useful but does not substitute for these constraints.

Planning identifies the system that owns each fact

The distributor’s warranty policy and customer status may live in different systems. Planning asks where the authoritative record is, what happens when it changes and who may use it. A policy document may be semantically similar to another yet apply to a different equipment model or date. The assistant needs version, applicability and customer-scope information in addition to keyword relevance.

Create a data contract: customer ID, asset or serial number, warranty version, effective period, source ID, review owner and requested business action. Decide which records should be indexed for search and which should be read through an authenticated business application tool at request time.

If a contract contains private pricing, the retrieval system must enforce permitted access. Telling the agent not to mention confidential prices cannot secure an index that makes them broadly retrievable. The broader data and AI relationship appears in this planning decision: clean, timely and authorized source information is a prerequisite for useful AI behavior.

Planning should also assign ownership. Who approves a new warranty revision? Who corrects a customer ID mismatch? Who must respond when a retrieval index is stale? These are not operational afterthoughts; they determine whether the architecture has a reliable source-of-truth process.

Design selects an agent pattern that fits the decision

The simplest workflow could validate the customer, retrieve the relevant policy, draft a recommendation and send it to a reviewer. If that meets the requirement, a fixed process may be easier to test than a general autonomous agent. Model-directed tool selection becomes more valuable if the next legitimate action depends on evidence discovered during the request.

Consider three possible tools: read warranty eligibility, inspect shipment history and create a draft review. They need distinct input contracts and errors. The application, not the generated text, must check the user’s authority for the requested customer. A read-only tool may be permitted in a broader set of circumstances than a tool that changes inventory or billing state.

A multi-agent design might delegate warranty policy research and logistics analysis, but the coordinator needs source references, incomplete-state handling and a way to resolve contradictory findings. If two specialists disagree, a fluent final summary is not an acceptable substitute for a review process that can identify and investigate the difference.

Choosing between Copilot Studio and Foundry is similarly tied to the actual requirement. Copilot Studio may provide a managed business-facing agent integrated with existing apps; Foundry may support specialized model or retrieval behavior where justified. Azure AI architecture provides useful implementation concepts without requiring every business process to become a fully custom application.

Business applications enforce the state that the agent may propose

The draft return authorization belongs in the application owning the order process. A Power Platform or Dynamics 365 workflow can represent pending, reviewed, approved and completed states with appropriate identity checks and business rules. The agent may populate draft content, but it should not manufacture an “approved” flag merely because it predicts that the customer probably qualifies.

Define one deterministic authority boundary for the eventual financial or inventory effect. Before creating an approved replacement order, validate the customer, approver, asset, amount and recorded decision. Include an idempotency or deduplication mechanism so retries cannot produce several orders for one authorized case.

A model can ask to invoke a tool and provide well-formed arguments while still requesting an unauthorized action. Business application logic and data permissions must decide whether execution is allowed. A useful Power Apps interface may make those states inspectable by supervisors, but the authorization must exist at the backend that performs the action.

Designing an agent and designing its business application are thus related but distinct. One produces recommendations and orchestrates information; the other preserves authoritative records and approved transitions.

Testing depends on what planning promised to protect

When the architecture says it will use only authorized, current warranty material, evaluation must include a restricted customer and an outdated document. When it promises one draft per request, a test must simulate a tool timeout and retry. When it promises human approval before a costly replacement, a test must attempt the action without approval and observe a denial.

Create a small corpus of synthetic cases: ordinary warranty, expired warranty, missing serial number, conflicting policy versions, unauthorized customer, unavailable shipment service and duplicate tool call. Evaluate schema compliance and actual content correctness separately. A valid structured response may still contain the wrong rule or source citation.

Subject-matter reviewers should judge nuanced eligibility recommendations, while deterministic checks can enforce operation scope and required fields. A model that refuses every request could be safe against one kind of error while failing to deliver the agreed business value. The acceptance contract should measure both useful task completion and protected boundaries.

Record exact source and configuration versions for each test so a later prompt or model update can be compared fairly. An improvement in average response quality should not conceal a single newly permitted unauthorized billing action.

Deployment is the continuation of the design, not its end

A change to a Copilot Studio action, Dataverse rule, policy source, model deployment or search index can alter the entire workflow. Some assets can be deployed through platform solution processes and CI/CD practices, but the release plan must identify stateful data and external effects that may not be reversible by restoring a prior prompt.

Before promotion, record the deployed versions, approved identities and integration endpoints. Validate a representative customer request from the actual intended environment. A connector that works under a developer’s broad test account may fail or expose too much information under production identity settings.

Plan rollback with the business owner. If a new agent version produces unsupported warranty decisions, the team may need to require manual review, revert a knowledge index or disable a write-capable tool. If a wrong order already exists, an authorized reconciliation process must correct that business record. Merely restoring the old model version does not undo an external action.

Lifecycle ownership belongs in the architecture decision: who approves configuration changes, who evaluates model quality, who can disable an agent, and who maintains the policy evidence the assistant is expected to trust.

Telemetry connects operating outcomes to redesign

The service’s API can be healthy while representatives receive outdated warranty answers. Infrastructure latency and error-rate metrics cannot by themselves establish groundedness. The architecture should connect each request to the customer scope, retrieved source version, tool result, model configuration, review decision and final business effect, with sensitive detail controlled.

An unexpected increase in review rejections could indicate a stale index, a new model behavior, confusing prompts or a changed customer population. Those require different investigations. Monitoring should distinguish the component that failed rather than label every adverse AI outcome a hallucination.

Performance and cost also create feedback. Excessive sequential agent calls may slow the case workflow, while overly aggressive shortcuts may omit a required warranty lookup. Evaluate quality, latency, cost and compliance together. The correct fix might be retrieval optimization, bounded tool use, a different model for a narrow task or an adjustment to when human review occurs.

The operating model is part of the design because it determines whether the organization can identify and repair failures after the original development team is no longer watching every request.

Return operational findings to the next planning decision

Suppose monitoring shows that the assistant often recommends the wrong policy when a customer owns multiple equipment models. The immediate investigation identifies a missing model identifier in retrieval. The architecture correction may require changing source metadata, the query contract and evaluation cases; a general prompt asking for accuracy would not be a reliable substitute.

Now repeat the full loop. Planning must ensure asset identifiers are authoritative. Design must pass that identifier through the correct connected tool. Testing must include customers with multiple assets. Deployment must preserve the changed source fields and prompt/tool versions. Monitoring must identify regressions if the identifier is omitted again.

This feedback loop is why the October 2026 AB-100 categories are not three unrelated chapters. Requirements create design constraints; design establishes the operational controls; and actual use supplies the evidence for revising requirements and implementation. An architect succeeds when those connections are explicit enough that a future team can challenge, operate and safely change the solution.