MICROSOFT / TECHNICAL EXPLAINER

Microsoft AB-100: Dynamics 365 Finance and Supply Chain Agent Knowledge

Architect ERP-grounded agents for Dynamics 365 Finance and Supply Chain Management, including in-app help, Dataverse virtual entities, controls and approvals.

In this article

ERP agents must respect the system of record

Dynamics 365 Finance and Supply Chain Management hold operational records whose accuracy can affect inventory, supplier payment, fulfillment and financial reporting. A Microsoft AB-100 architect should not turn these into unverified chatbot knowledge. Define which business entity owns a fact, which roles may retrieve it, whether it can be changed, and what approval the transaction needs. A generated explanation is distinct from a posted ledger entry.

Imagine a purchasing manager asking about a delayed supplier shipment. The assistant could retrieve an authorized purchase order status and summarize a documented procedure. It should not modify supplier banking information, release a blocked invoice or misstate an available quantity because a stale document contained a different number.

Choose between static help and operational data

In-app help and guidance can use curated documentation explaining procedures. Questions about current inventory, account credit limits or order processing require authoritative operational records with the correct access scope. Combining the two can be valuable, but define when to retrieve a policy article versus a live table lookup and how to handle contradictions.

A user’s role may permit reading one legal entity’s accounts but not another’s. An agent that uses a privileged service connection without preserving data authorization can bypass the normal ERP boundary. Plan delegated or supported service access explicitly, and validate actual record-level security rather than assuming a chat interface inherits every permission automatically.

Understand Dataverse virtual entity integration

Microsoft describes finance and operations virtual entities that expose live operational data through Dataverse representations under supported configurations. The F&O agent knowledge tutorial illustrates adding an entity such as customer information as a Copilot Studio knowledge source. The exact table names, prerequisites and behaviors must be checked for the tenant and environment used.

A virtual entity is not simply a duplicate of all ERP data into a free-form document index. Treat the live source, query behavior, permissions, freshness and transactional API as separate decisions. A supported read from a Dataverse representation does not authorize every change that the underlying operational application might permit to another user.

Design finance-side data questions carefully

For a credit-risk question, define the customer and legal entity, the official credit-limit fields, permitted user role, currency, timestamp and explanation needed. If an agent asks for ‘all customers above a threshold,’ ensure that the underlying query is supported and scoped, and that numeric results are validated from records rather than model estimates.

Test one user allowed to see the customer’s limit and one restricted user. Capture the difference without copying sensitive real customer data into test prompts or logs. If a query returns no records, distinguish “not found” from “no permission” where security policy allows disclosure of that distinction.

Design supply chain exceptions as controlled workflows

A supply chain agent can summarize supplier delays, explain replenishment policy or suggest follow-up actions. A purchase order amendment, inventory adjustment or supplier onboarding approval has transactional side effects. Define user authority, validation and idempotency in the underlying business application, and preserve the pre-change record and audit metadata.

When a shipment status is late, the model may recommend alternative suppliers or expedited shipping. The actual procurement action should pass a deterministic approval and budget check. The approval architecture explains consequences and control; ERP implementation must enforce those controls independently.

Connect in-app help and external knowledge safely

Approved help documents can explain finance or supply chain procedures, but not every website or PDF is authoritative. Curate current source versions and indicate effective dates when regulations, fiscal rules or process changes matter. An agent responding from an obsolete product manual should disclose uncertainty or route the user to current support.

When extending in-app guidance with Copilot Studio, define a source inventory, accepted document formats, refresh owner, access policy and validation test for each knowledge integration. A helpful answer with a reference link should still be checked for relevance and permissions; source provenance is not the same as user authorization.

Design multi-application requests with explicit ownership

An employee may ask one question involving CRM opportunity, Sales order, supplier status and finance invoice. Decide which application is authoritative for each field before aggregating it into an answer. Use stable business identifiers and avoid matching two unrelated customers by display name alone. When data cannot be reconciled, present the conflict and escalate rather than inventing a merged status.

Orchestration patterns provide the state and handoff boundaries needed for reliable workflows. In an ERP scenario, tool ordering and retry rules must also prevent accidental duplicate financial transactions. A timeout after committing an invoice requires checking transaction status before retrying.

Observe operational outcomes beyond response fluency

Track correct record retrieval, data freshness, authorization denials, unsupported operations, human corrections, latency and business outcome. A generated narrative saying a payment was approved does not establish that a posted transaction exists. Audit the target ERP state and include appropriate correlation data while protecting confidential financial fields.

Write scenario tests for invalid legal entity, withdrawn supplier, outdated help article, unavailable Dataverse connection and approval rejection. Each should produce an understandable safe result and a route to the responsible team. Not all scenarios require a model: structured validation may be a better deterministic control.

Deploy only with tenant-specific feature verification

Capabilities can depend on Dynamics version, license, Copilot configuration, regional support and preview/GA status. Check current Microsoft Learn documentation and the actual tenant before promising a specific in-app location or autonomous function. The AB-100 exam focuses on architecture decisions rather than a fixed visual navigation path.

Produce a design record with data owners, entities, permissions, knowledge sources, allowed tool operations, approval requirements, tests, telemetry and rollback. The official AB-100 outline covers Finance and Supply Chain agent functionality and knowledge extension; the response should remain grounded in actual ERP authority.