An agent does not have one cost. It has a cost per successful outcome — and the gap between the two is operational design. Gartner found that agentic workflows consume 5–30× the tokens of a standard chatbot query, which means the architecture choices between "call the model" and "get a useful result" are where most of the money actually goes.
The token price is the visible fraction
Model pricing is easy to compare, so it dominates early architecture discussions. But an agentic workflow may plan, retrieve, call tools, critique its own answer and retry before producing a usable result. Research from Microsoft and the Stanford Digital Economy Lab found that agentic tasks can consume roughly 1,000× more tokens than a standard chat interaction — which means the per-token price you negotiated is not the cost you will actually pay.
The relevant unit is not cost per call. It is cost per completed task at the quality and latency the business actually needs.
Reliability changes the denominator
A cheaper model that triggers more retries or more human correction ends up costing more in practice. The same is true of broad agents that burn context and tool calls deciding something a narrower workflow could handle deterministically.
Evaluation should connect quality to economics: success rate on representative tasks, steps per completion, failure recovery and human effort — not aggregate token consumption.
Not every step needs a model
Different steps deserve different levels of intelligence. Classification, extraction and routing often belong in smaller models or deterministic logic; ambiguous synthesis may justify a more capable model. Caching and pre-computation remove repeated work entirely.
Rebuilding one multi-agent platform along these lines — moving deterministic steps out of the model, caching aggressively and narrowing agent scope — cut cost per task roughly fifteenfold between v1 and v2, with no loss in output quality. The lesson was not about choosing a cheaper model but about deciding what should never hit a model at all.
This is where orchestration earns its complexity: by matching each step to the cheapest reliable mechanism, rather than by adding more agents.
Start with a value envelope
Before optimising the stack, define what a successful outcome is worth. That creates a cost ceiling and exposes the cases where volume or consequence never justified an agentic approach in the first place. Gartner expects more than 40% of agentic AI projects to be cancelled by 2027 — and in most of those cases the issue will not be that the agent could not do the job, but that nobody defined what "doing the job" was worth before building it.
The best economics come from product discipline: a bounded job, a clear service level and evidence that the workflow improves a decision someone cares about.