Vendor pricing pages quote per-token or per-seat costs. Those are real, and they are rarely the number that decides whether an agent project succeeds. Here is the fuller picture for a mid-sized Indian marketing team.
The costs vendors quote
Model usage for a narrow internal task, running a few hundred times a month, generally lands low enough to disappear inside an existing software line. Orchestration platforms charge per seat or per run and vary widely. Neither is usually the binding constraint.
The costs nobody quotes
Example collection comes first. You need fifty to a hundred examples of the task done correctly to know whether the agent is any good. Someone senior enough to judge quality has to produce those, and it is a real chunk of a week.
Review time comes second and lasts longest. For the first month somebody checks every output. Budget a few hours a week for a task running daily. Teams that skip this do not save the time, they spend it later cleaning up.
Parallel running is third. You keep doing the task manually while the agent runs, because you cannot trust it yet. For a month you are paying twice. Nobody puts this in the business case and it is often the largest single cost.
Data cleanup is fourth and the most variable. If the agent reads your CRM and your CRM is messy, cleanup happens before anything works. This can dwarf everything else on the list.
How to think about payback
The honest framing is that the first agent rarely pays back on the first task. What you buy is the setup: an evaluation habit, a review process, a cleaned data source, and a team that has done it once. The second and third agents are much cheaper because that groundwork carries over.
So judge the first project on whether it built reusable capability, not on whether it saved hours in quarter one. Teams that demand immediate ROI pick tasks that are too big, skip the example set, and get nothing.
When not to bother
If the task runs fewer than about twenty times a month, an agent is not worth the setup. If output goes straight to customers with no review, the risk is not worth it yet. And if the underlying data is bad, fix the data first. An agent on bad data produces confident, well-written, wrong output, which is worse than no output.