A demonstration proves that a model can produce an interesting output under controlled conditions. An operational system must work with real users, data, permissions, suppliers, exceptions and consequences.
Before deployment, the team should be able to answer
- What is the intended purpose?
- Which use is explicitly excluded?
- Which source or system is authoritative?
- What data can enter the workflow?
- Which external actions are permitted?
- Who reviews consequential or uncertain outcomes?
- How is quality evaluated?
- What happens when a provider, model or prompt changes?
- How are incidents reported and contained?
- Which evidence will support approval and future reassessment?
Evaluation must test the workflow rather than only the model response. That includes retrieval, tool use, permission handling, latency, failure states, escalation and the usefulness of the output to the intended user.
The handover should name the owner, users, support boundary, monitoring cadence, change process, known limitations and stop conditions. Without those elements, the organisation has a continuing experiment rather than an operating system.
DDAI view: production is an evidence state, not a label applied after a successful demonstration.