Deploy AI-Powered Business Solutions
Deployment is a lifecycle, not a publish button
Deployment questions test ALM discipline and operational readiness.
- Environment strategy: development → test → production, each with its own Dataverse instance. Author as unmanaged; ship managed.
- Solution hygiene: no hardcoded values - use environment variables; no embedded credentials - use connection references so each environment supplies its own connection.
- Pipelines: Power Platform pipelines or Azure DevOps to automate export and import, with solutions stored in source control.
- Channel configuration: publish to Teams, Microsoft 365, web, custom app, or voice. Channel enablement is per environment, and authentication (SSO, OAuth) must be configured before users can sign in successfully.
- Evaluate before go-live: assemble a test set of real questions with expected answers, score groundedness and completeness, red-team for prompt injection and out-of-scope responses, then run a pilot cohort.
- Operate: monitor analytics, session transcripts, and errors; track message and capacity consumption against the licensing plan; feed failures back into knowledge and instructions.
- Govern at scale: managed environments, environment groups, DLP policies, and an agent inventory so unmanaged agents are visible.
Traps: testing only in the authoring environment and assuming parity; deploying the unmanaged solution; forgetting that connection references must be re-bound after import; assuming publishing propagates instantly to every channel; ignoring consumption costs until the invoice arrives; having no rollback - keep the previous managed solution version.
Rule of thumb: promote, don't rebuild. If a fix requires hand-editing production, the ALM process is broken.
Worked example: an agent goes live in Teams. In week two, 12% of sessions end with "I don't know." Transcripts show the SharePoint site was reorganized. Fix: update the knowledge source in the test environment, re-run the evaluation set, promote the managed solution through the pipeline, then republish the Teams channel. No production hand-editing.