Agents are cheap per conversation and expensive per surprise. The token bill for a chat conversation is usually pennies; a voice call adds telephony and speech costs measured in cents per minute; the surprises come from long conversations, retries, verbose prompts and models chosen by default rather than by need. Budgeting means understanding the drivers, not the headline price.

Chat

Cost is input tokens (your prompt plus the knowledge you attach plus the conversation so far) times the number of turns, plus output tokens. A bloated system prompt multiplies across every turn of every conversation; retrieval that attaches only the relevant document instead of the whole knowledge base cuts cost sharply. A capable mid-tier model handles most support and qualification; reserve the largest models for the escalation classifier or the hardest cases.

Voice

Add speech-to-text and text-to-speech per minute, telephony per minute, and latency-optimised models that cost more per token. A booking call of three minutes is inexpensive; a rambling ten-minute call is not, and design should keep calls short — confirm, book, send the details by message, end.

The hidden lines

Orchestration and logging infrastructure, evaluation runs after each prompt change, the human review time, and the occasional escalation that costs a person twenty minutes. Include them; they are real and they are still small next to a salary.

Comparing with a person

Compare per outcome, not per conversation: cost per booked appointment, per resolved ticket, per qualified lead — including the conversations at 2 am that a person would not have had at all. The agent usually wins on volume and coverage and loses on the hard conversations, which is why the hand-off design matters more to the economics than the model price.