Running an AI system: four cost lines and why estimates often miss the mark

The article breaks down post-launch AI costs into cloud and model usage, maintenance hours, ownership and coordination, and retraining where applicable. It notes that infrastructure is frequently overestimated while maintenance is underestimated, leading to budget overruns. It also advises what to include in support agreements to make recurring costs visible upfront.

The article highlights a recurring mismatch between projected and actual AI operating expenses, noting that infrastructure estimates often run two to three times higher than reality while maintenance is frequently omitted from budgets entirely. This discrepancy stems from architectural choices—event-driven systems that activate only upon triggers consume far fewer cloud resources than continuously running services.

A third hidden cost category emerges at the handover point between build and run phases. Hosting credentials, repository access, and similar transition items typically fall outside both build and maintenance contracts, leaving them unowned until an urgent need arises. The author suggests support agreements should explicitly itemize these recurring costs to prevent post-launch surprises.