Comparing only upfront prices leads to poor decisions. A software subscription can look cheaper than a custom build until per-user licenses, integration work and workarounds are added up over several years, and a custom build can look cheaper until maintenance and hosting are included.
For AI systems, total cost of ownership also includes model usage, monitoring, evals and the time people spend reviewing exceptions. For example, a document workflow's running costs grow with volume, so a sound estimate uses expected volumes, not just the pilot. The misconception is that TCO is a finance exercise alone. Operations and IT hold much of the information needed to estimate it honestly.