The prototype is not the product
A working demo takes a weekend and a few hundred dollars of API credits. Production is a different animal: evaluation harnesses, guardrails, fallbacks, logging, human-review queues and integration with systems that were never designed for machine users. Teams consistently report that the demo represents about 10 percent of total effort.
The real cost structure
Expect three buckets. Engineering: one to three engineers for three to six months for a serious single-workflow agent, the dominant cost. Inference: from a few cents to a few dollars per completed task depending on model tier and retry rates; at scale this becomes a unit-economics line, not a rounding error. Operations: monitoring, evaluation and prompt maintenance run 15 to 25 percent of the build cost annually, forever.
Where teams overspend
Using frontier models for steps a small model handles, retrying failures instead of fixing root causes, and building platform infrastructure before the first workflow earns its keep. Start narrow: one workflow, one success metric, boring reliability.
FAQ
Build or buy?
Buy the horizontal plumbing (orchestration, evals, monitoring), build the workflow logic that touches your domain data. That boundary shifts yearly; revisit it.
What should an agent task cost?
Price the human alternative. If a task costs a person 20 minutes, an agent completing it for under a tenth of that loaded cost has room to be a business.
Biggest hidden cost?
Evaluation. Knowing whether the agent did the job correctly, at scale, costs more than most teams budget and matters more than the model choice.