Case studies — AB-620

Contoso is a multinational manufacturer with 18,000 employees. HR wants an employee service agent in Microsoft Teams that answers policy questions, retrieves each employee's remaining leave, and submits leave requests.

6 chained questions

Fabrikam operates a logistics marketplace used by 600 partner companies. It wants an authenticated external agent embedded in its partner portal. A partner may ask about its own shipments, request a delivery reschedule, and analyze monthly delivery performance.

6 chained questions

A. Datum manufactures industrial sensors and is building a Copilot Studio agent for authenticated support engineers. The agent must diagnose incidents using product manuals and live device telemetry, then create an escalation when automated guidance is insufficient.

8 chained questions

Northwind Retail has 140 stores and is introducing a Copilot Studio agent for store associates in Microsoft Teams. Associates sign in with Microsoft Entra ID. The agent answers operating-policy questions, retrieves live inventory, requests stock transfers, and helps recover a legacy point-of-sale terminal.

10 chained questions

The City of Alpine is building a Copilot Studio agent for its public website. Anonymous visitors may ask questions from published service guides. Authenticated residents may check permit status and submit address-change requests. The website uses an approved citizen identity provider and must never rely on personal identifiers typed into chat as proof of identity.

10 chained questions

Fabrikam Bank is building a Copilot Studio assistant for authenticated branch advisors in Microsoft Teams. It summarizes approved product policies, retrieves a customer's current application status, drafts suitability notes, and requests exceptions.

10 chained questions

Contoso Health is building an authenticated Copilot Studio agent for care coordinators in Microsoft Teams. It answers questions from approved clinical operations procedures, retrieves appointment availability, drafts non-clinical follow-up tasks, and submits transport requests. The agent must never diagnose patients. Procedures are indexed in Azure AI Search with facility, effective-date, and access-group metadata. Appointment availability is exposed through an MCP server using delegated OAuth scopes. Transport requests are irreversible after dispatch and therefore require coordinator confirmation; high-cost requests additionally require supervisor approval through an agent flow. A specialist Foundry agent classifies non-clinical follow-up notes into a fixed schema, but its output must be validated before storage. Application Insights must correlate retrieval, MCP calls, specialist delegation, and flows without recording patient text or tokens. The solution moves through development, validation, and production using managed solutions, environment variables, connection references, and pipeline gates. A versioned multilingual test set covers authorization, groundedness, refusals, tool failures, and approvals.

8 chained questions

Adatum Retail is building an authenticated Copilot Studio agent for store associates in Microsoft Teams. It answers questions from approved return policies, checks inventory, creates return authorizations, and delegates delivery rescheduling to a logistics agent through A2A. Policies are indexed in Azure AI Search with country, brand, effective-date, and access metadata. Inventory and return operations use a governed custom connector. Read operations may retry transient failures, but return creation uses an idempotency key and status lookup because a timeout can leave an unknown outcome. Refunds above a threshold require manager approval before submission. Application Insights correlates the conversation, retrieval, connector calls, A2A task, and approval without storing customer payloads or tokens. The solution moves through development, test, and production with managed solutions, environment variables, connection references, and pipeline approvals. A bilingual versioned test set covers grounding, access, tool failures, duplicate prevention, delegation, and approvals.

8 chained questions

Northwind Council is building a Copilot Studio agent for residents and permit officers. Anonymous visitors may read public guidance, but application status and document submission require resident authentication. Officers use Microsoft Teams and role-based access. Current regulations are retrieved from Azure AI Search using permit type, district, effective date, and access metadata. A custom connector exposes application status and submission operations; submission supports an idempotency key and status lookup. Complex zoning questions delegate to a specialist Foundry agent, whose response must cite approved sources and remain advisory. Fee waivers require an officer decision, and high-value waivers additionally require supervisor approval before the permit API is called. Telemetry uses correlation, sanitized outcomes, duration, and dependency data without storing documents or tokens. Managed solutions, variables, connection references, and pipeline gates promote releases. A bilingual test set covers anonymous and authenticated boundaries, grounding, duplicate prevention, specialist delegation, approvals, and failures.

8 chained questions