ALM for D365 AI Features
Manage AI feature rollouts across Dynamics 365 Finance, Supply Chain, Customer Service, and Sales — from feature flags and configuration export to knowledge source management.
D365 AI is a different ALM beast
Copilot Studio ALM = you package and ship the agent. Foundry ALM = you package and ship the model. D365 AI ALM = Microsoft ships the model, you manage the configuration.
Most D365 AI features are built into the platform. Microsoft trains and updates the models. Your job as an architect is to manage the feature enablement, configuration, and customisation lifecycle — not the model itself.
Think of it like a car: Microsoft builds and updates the engine (the AI model). You configure the dashboard, select the driving mode, and choose which roads to drive on (the business configuration).
Configurable vs custom AI in D365
| Feature | Configurable AI (Microsoft-Managed) | Custom AI (Customer-Built) | ALM Implication |
|---|---|---|---|
| Model ownership | Microsoft trains and updates the model | Customer trains or fine-tunes the model | Configurable AI: manage config. Custom AI: manage the full model lifecycle. |
| Examples | Copilot in D365, demand forecasting, invoice capture, sentiment analysis | Custom Foundry models, custom Copilot Studio agents integrated with D365 | Know which features are configurable vs custom for the exam. |
| What you manage | Feature flags, knowledge sources, business rules, Copilot customisation | Model training, evaluation, deployment, prompt engineering | Configurable AI has simpler ALM — toggle and configure, not build and deploy. |
| Environment strategy | Enable in sandbox first, validate, then enable in production | Full Dev-Test-Prod pipeline with model registry | Both need environment promotion — but the artefacts differ. |
| Update control | Microsoft controls model updates. You control feature enablement timing. | You control everything — model version, deployment timing, rollback | Configurable AI: plan for Microsoft-initiated model changes. |
ALM for D365 Finance and Supply Chain AI
| Feature | ALM Consideration | Environment Strategy |
|---|---|---|
| Copilot in Finance | Enable via feature management. Customise with knowledge sources and business terms. | Enable in sandbox → validate with finance team → promote to production |
| Demand forecasting | Configure forecast models, parameters, and data sources. Export configuration via data entities. | Test forecast accuracy in sandbox with historical data before enabling in production |
| Supplier Communications Agent | Configure agent topics, knowledge sources, and escalation rules. | Deploy customisations as solution components through D365 environment management |
| Invoice capture | Configure recognition models, validation rules, and approval workflows. | Train recognition on sample invoices in sandbox → validate accuracy → enable in production |
| Warehouse management Copilot | Configure task assistance settings, safety rules, and integration points. | Enable in test warehouse environment → pilot with one warehouse → full rollout |
ALM for D365 Customer Experience and Service AI
| Feature | ALM Consideration | Environment Strategy |
|---|---|---|
| Copilot in Customer Service | Configure knowledge sources, response generation settings, business terms glossary. | Export business terms and knowledge config → import to production after UAT |
| Contact Center agent config | Configure agent skills, routing rules, escalation policies, real-time suggestions. | Test routing and escalation in sandbox → validate with pilot group → full deployment |
| Customer Insights data unification | Configure entity mapping, match rules, enrichment settings. | Export configuration → test with anonymised data in sandbox → import to production |
| D365 Sales Copilot | Configure CRM field mapping, email assistance settings, meeting preparation rules. | Enable per-security role in sandbox → validate with sales pilot group → broader rollout |
| Sentiment analysis | Configure sensitivity thresholds, alert rules, and escalation triggers. | Calibrate thresholds in sandbox using historical case data → deploy to production |
Feature flags vs solution deployment
A critical distinction for the exam: many D365 AI capabilities are controlled by feature flags in Feature Management, not solution deployment.
- Feature flags — enable or disable a D365 capability at the environment level. No packaging, no import/export. Just toggle on or off.
- Solution deployment — package customisations (custom agents, flows, business rules) and promote through environments.
- Configuration export — export configuration settings (business terms, knowledge sources, model parameters) as data entities and import into the target environment.
Not every AI rollout requires a solution deployment. Sometimes it is just a feature flag toggle and a configuration import.
Scenario: Kai manages AI feature rollouts at Apex Industries
Kai Mercer manages Apex Industries’ D365 Supply Chain Management environment. The team is rolling out three AI features simultaneously:
1. Demand Forecasting (configurable AI):
- Kai enables the feature in the sandbox via Feature Management
- Lin Chen’s team configures forecast parameters (seasonality, lead times, data sources)
- Tomasz validates forecast accuracy against the last 12 months of actual demand
- After approval, Kai enables the feature in production and imports the configuration
2. Supplier Communications Agent (custom agent):
- Kai’s team customises the agent topics and connects it to Apex’s supplier portal API
- The customisations are packaged as a D365 solution component
- Deployed through the standard Dev → Test → Production pipeline
- Environment variables handle the different supplier portal URLs per environment
3. Copilot in Finance (configurable AI):
- Kai enables Copilot in the sandbox
- Priya Sharma configures the knowledge sources (internal policy documents, accounting standards)
- Finance team validates Copilot responses for 2 weeks in sandbox
- Kai enables in production and imports the knowledge source configuration
Key lesson: Each feature uses a different ALM pattern. Kai does not deploy a solution for demand forecasting — that is a feature flag plus configuration. The Supplier Communications Agent customisation does require solution deployment. Copilot is a feature flag plus knowledge source configuration.
Exam tip: feature flags are not the same as solution deployment
When the exam describes a D365 AI scenario:
- Enabling a built-in D365 AI feature = Feature Management toggle. No solution packaging required.
- Customising a D365 AI feature (custom topics, business rules, flows) = Solution deployment through the standard ALM pipeline.
- Configuring a D365 AI feature (knowledge sources, parameters, business terms) = Configuration export/import as data entities.
- Microsoft updates the AI model = You do not control this. Plan for it by testing in sandbox after Microsoft updates before allowing the update in production (where available).
If the exam asks “how should the architect deploy demand forecasting to production?” — the answer involves Feature Management and configuration export, NOT a managed solution.
Flashcards
Knowledge check
Kai needs to deploy demand forecasting to Apex's production D365 SCM environment. The feature is already enabled and configured in the sandbox. What is the correct ALM approach?
After Microsoft releases an update that improves the Copilot model in D365 Customer Service, what should the architect do before enabling it in production?
Which of the following D365 AI scenarios requires a solution deployment (not just a feature flag toggle)?
🎬 Video coming soon
Next up: Agent Security — designing security architectures for agent authentication, data access, model protection, and content safety.