The build fee is the part everyone asks about. The running cost is the part that decides whether the thing survives its first busy month, and it is made of four bills, not one.
The four bills
Every voice agent is paying for all of these simultaneously:
- Telephony — the number and the minutes. Cheap, predictable, the one everyone budgets for.
- Speech to text — converting the caller's words. Priced per minute of audio.
- The language model — deciding what to say. Priced per token, and it grows with how much context you hand it on every turn.
- Text to speech — saying it back. Priced per character, and the good voices cost noticeably more than the acceptable ones.
Where the surprise usually comes from
Not from volume. From context length.
A naive implementation sends the entire conversation history to the model on every single turn. A six-minute call with thirty turns re-sends a growing transcript thirty times. The per-token price was never the problem; the multiplication was. Summarising older turns instead of resending them is the single change that most often halves a bill.
The second surprise is the voice. Teams pick the most natural-sounding option during a demo nobody is paying for, then discover it is several times the cost of one that callers would not have noticed.
What this means for the brief
Decide early what a call is worth to you. A business where a booked job is worth ₹15,000 can afford a generous per-minute cost and a premium voice. A business triaging support queries at low value per call needs the cheap path, shorter prompts and a quick hand-off.
That number should be in the scope before anything is built, because it changes the architecture — not just the invoice.
The one that is free
Hanging up quickly. An agent that identifies what the caller needs in forty seconds and books it costs a fraction of one that explores. Most of the tuning work on a voice agent is making it shorter, which is also what the caller wanted.
