The most common question at the end of a first conversation is whether any of this pays for itself. What people usually mean is a comparison between a monthly API invoice and a one-off server purchase. The calculation is not that simple, but it is also not so complicated that you cannot do it in advance.
The cost blocks people think of first
Hardware. The most visible item, and the easiest one to overestimate. What matters is not which model is theoretically best, but which one is sufficient for the task at hand. Classifying incoming documents makes entirely different demands than an assistant that has to write free-form answers.
Running it. Power, rack space, network. Usually modest for a single server in your own rack, but not zero.
The cost blocks people miss
Setup and integration. A model on its own does nothing. The effort sits in connecting it to the systems where the work actually happens: document management, ERP, mailboxes, ticketing. This is almost always the single largest item, and it applies whether the model runs locally or in the cloud.
Maintenance. Models get replaced, libraries change, requirements grow. If you run it yourself, you take on that work, either in-house or bought in. It belongs in the calculation, not in a footnote.
Review and sign-off. For sensitive processes, the time spent with data protection officers, the works council and internal audit is a real cost. That time is usually considerably higher for an external API than for a system that never leaves your own infrastructure. This is the point at which local solutions often pay off, before hardware even enters the discussion.
What can be settled in advance
Before anything is purchased, three numbers should be on the table: how many transactions occur per day, the longest response time that is still acceptable, and how much staff time the process consumes today. Those three figures are enough to establish the order of magnitude of the hardware required and to estimate the benefit, with no test rig and no purchase.
If the numbers are not known, that is not an obstacle. Establishing them is usually the first sensible step, and it is worth doing even if no AI system is ever bought.