6.1.3. ALM for AI in Dynamics 365 Apps
💡 First Principle: AI features in D365 apps are managed differently from custom agents — they're platform features that Microsoft updates, not custom code that you version. The architect's ALM responsibility is managing the configuration of these features, not the features themselves.
What the Architect Manages vs. What Microsoft Manages:
| Element | Who Manages | ALM Approach |
|---|---|---|
| AI model (Copilot, forecasting, etc.) | Microsoft | Feature toggles; test after platform updates |
| Feature configuration | Architect/Admin | Export as part of D365 solution; version in source control |
| Business terms | Architect/Admin | Managed through D365 admin; export/import per environment |
| Knowledge sources | Architect/Admin | Separate data management; sync schedules per environment |
| Custom prompts and plugins | Development team | Solution-managed; standard code ALM applies |
D365 Platform Update Testing:
When Microsoft releases platform updates that change AI behavior, the architect must ensure regression testing covers AI-dependent business processes. A platform update that improves Copilot's general intelligence might change how it interprets specific business terms — testing must verify these scenarios.
F&O-Specific ALM Considerations:
Dynamics 365 Finance and Supply Chain Management apps use a different deployment model (cloud-hosted environments with scheduled updates) than CX apps (Dataverse-based). AI feature configuration in F&O may involve feature management flags, parameter configurations, and data entity exports — all of which must be tracked in the ALM process.
⚠️ Exam Trap: A scenario describes a company whose D365 Customer Service Copilot behavior changed after a platform update. A distractor recommends "rolling back the AI model." The correct answer is to test the updated AI behavior against business term configurations and knowledge sources — the issue is likely a compatibility gap between the platform update and the existing configuration, not the model itself.
Troubleshooting Scenario: A company enables Copilot features in D365 Finance across 12 countries simultaneously. Within a week, the French subsidiary reports that Copilot's account reconciliation suggestions conflict with local GAAP requirements. The fix required disabling the feature in France, reconfiguring, and re-enabling — disrupting all users. A phased rollout would have caught this: deploy to one region first, validate AI outputs against local requirements, then expand. D365 AI features may behave differently across regions because financial regulations, date formats, currency conventions, and language nuances all affect AI interpretation.
The principle for D365 AI ALM: never assume that an AI feature validated in one regulatory context works in another. Each jurisdiction is effectively a separate validation scope, even when the underlying platform is identical.
Reflection Question: A multinational company uses D365 Finance with AI-powered account reconciliation across 8 legal entities. Each entity has different reconciliation rules. Design the ALM process for promoting reconciliation rule changes across environments while ensuring entity-specific configurations are preserved.
