The unglamorous engineering work of running AI systems in production reliably, as opposed to building a model that works once in a demonstration. Version control, monitoring, testing, deployment, cost management and rollback. Most AI projects that fail do so here rather than at the modelling stage.
Technical view
Adds concerns to standard software practice that arise from behaviour being learned rather than specified: data and model versioning, reproducibility, and monitoring for drift, where input distributions shift away from training conditions and performance degrades silently. For language model applications specifically, sometimes called LLMOps, the emphasis shifts towards prompt versioning, evaluation pipelines, output monitoring, cost and latency tracking, and managing changes to underlying models supplied by third parties.
Related
Part of AI Terminology Index, section: Testing, Safeguarding and Running Models in Practice.