An open-data portal team explains AI data flow as collect → clean → train/evaluate → monitor in production. What literacy point are they teaching?

Open-data and municipal AI teams teach a basic lifecycle literacy point when they explain AI data flow as collect → clean → train/evaluate → monitor in production. That sequence emphasizes that civic projects need deliberate intake, quality work, model fitting with evaluation, and ongoing production monitoring rather than a single leap from raw files to live decisions. Each stage catches different failure modes, from bad source extracts to silent drift after deployment. Saying effective municipal AI follows a basic flow of collect data, label it, train and evaluate, then monitor in production omits the cleaning step; labeling is only one possible preparation task, while cleaning is the required quality work that must happen before models are fit and evaluated. Saying effective municipal AI follows a basic flow of collect data, clean it, fine-tune and deploy, then monitor in production skips evaluation; train and evaluate is the required fitting-with-testing stage before deployment, while fine-tuning and deploying alone provide no evaluation gate before production monitoring. Saying effective municipal AI follows a basic flow of collect data, clean it, train and evaluate, then archive in production replaces ongoing oversight with storage; production monitoring comes last as active oversight for drift and failures, not archiving. Exam caveat: stages iterate; monitoring findings often drive new collection or cleaning, so the flow is a loop in practice even when taught linearly at essentials level. Operational check: for each municipal AI pilot, name owners for collection, cleaning, train/evaluate, and production monitoring, and refuse go-live if the monitoring stage has no metrics, owners, or rollback path.