Production readiness is not a level of model accuracy. It is the point at which a workflow can run in the business without an engineer nearby, and the people who depend on it know what to do when it behaves unexpectedly.
Data and access
- The workflow reads and writes real records with scoped credentials
- Access follows the same permission model as the rest of the business
- It is documented what data leaves your environment, where it goes and how long it is retained
Behaviour under uncertainty
Every automated decision needs a defined path for low confidence, for missing inputs and for a dependency being unavailable. The safest fallback is usually the previous manual process, made explicit rather than assumed.
Observability
- Each run is traceable: inputs, sources used, output and who reviewed it
- Failures raise an alert that reaches a named person
- Volume, escalation rate and correction rate are visible over time
Change control
Prompts, thresholds and rules are configuration, and configuration changes behaviour. Version them, review them like code, and keep a record of what changed when the output pattern shifts.
Operability
The final test is organisational rather than technical: can the operations team pause the workflow, understand a specific output, correct it, and explain it to a customer or auditor? If the honest answer requires an engineer, the workflow is not ready to be handed over.
