AI engineering.
Choosing an AI delivery partner, what the work looks like once the demo is over, and why prompt engineering stopped being a job title of its own.
What AI work looks like once the demo is over.
The gap between a working prototype and a feature customers depend on is where most AI roadmaps stall, and it is not usually a modelling problem. It is evaluation, retrieval quality, cost per request, latency under real traffic, and what the system does with the inputs nobody thought to test.
That work does not finish, which is why it is poorly served by a scoped project and why AI engineering services are assigned continuously instead. A model swapped in eight months from now should be a configuration change and a re-run of the evaluation set, not a rebuild.
The operational half of it — deployment, monitoring, retraining, drift — is a distinct discipline, and where models are already in production it is usually the binding constraint rather than the modelling. MLOps engineering covers that side.
The posts below are written for the person choosing a partner or deciding how to staff the work, not for the person building it.