· 4 min
After the demo: production work
A demo proves that a path can work once with chosen inputs. Production work proves that the organization can use, support and change the path under ordinary conditions. The gap is often less about choosing a different model and more about connecting the feature to real data, real permissions and a team that owns its behavior.
Walk the path from request to support
- Confirm the job. Name the user, decision or action the feature supports, and the cases where it should decline or hand off.
- Validate the data path. Check source quality, access, retention, logging and any provider or subprocessors that receive information.
- Connect systems safely. Define service identities, field mappings, write permissions, duplicate handling and rollback for downstream updates.
- Test behavior over time. Keep representative examples and review changes to prompts, models, retrieval data or integrations before they reach users.
- Plan operations. Assign an owner for incidents, user questions, budgets, monitoring, dependency updates and routine content changes.
- Prepare the handover. Deliver runbooks, known limitations, access inventory, recovery steps and a walkthrough with the people who will operate the feature.
Hypothetical example: a demo answers employee policy questions from a handful of documents. Before launch, the organization must verify that employees only retrieve material they are allowed to see, set a process for updating superseded policies, show which sources informed an answer, and define where unanswered questions go. A polished response in a demo does not establish any of those controls. The example does not claim that one configuration is universally compliant.
Release in a way the organization can support
Begin with a limited user group and a clear fallback to the current process. Give users a way to report a harmful or incomplete result. Monitor meaningful signals—such as completion, escalation, latency, cost and review effort—without logging more personal content than is needed. Decide who can pause the feature and how the team will restore a known-good version.
Use the review to decide whether the workflow is ready for broader use, needs another iteration or should stop. Record outstanding risks and owners instead of hiding them in a handover deck. Production readiness is an operating agreement between product, engineering, security and the business owner.
For reusable operation artifacts, see AI runbooks and handover packages. Review Custom AI Services support or discuss the demo and its launch constraints.