Why does each turn cost more than the last?

LLMs are stateless, so every turn re-sends the entire conversation history as input. A 6-turn conversation sends turn 1 six times, turn 2 five times, and so on — input grows roughly quadratically, which this calculator approximates.

What is the escalation rate for?

Not every ticket can be resolved by AI. The escalation rate is the percentage of tickets that still get handed to a human, so the blended cost is: AI cost for all tickets + human cost for the escalated fraction.

Are the human cost numbers realistic?

The human comparison uses your hourly rate times an average handle time. Real fully-loaded agent cost includes benefits, tooling, and management overhead, so treat the human figure as a conservative floor.

Does the AI replace agents entirely?

Usually no. The realistic model is deflection — AI resolves the simple majority and escalates the rest. Tune the escalation rate to your deflection target to see the blended cost.

AI Customer Support Cost Calculator

What does AI customer support actually cost?

“Pennies per conversation” is the pitch, but a real support deployment re-sends conversation history every turn, injects retrieved knowledge-base context, and still escalates a fraction of tickets to humans. This calculator models all three so you get a credible cost per ticket and monthly spend, side by side with the human-agent equivalent.

How the cost is built up

Because LLMs are stateless, every turn re-sends the whole conversation so far. For a conversation of n turns where each turn adds roughly the same number of tokens, total input tokens scale with the triangular number of turns:

turns_input = tokens_per_turn × (n × (n + 1) / 2)
ai_cost_per_ticket = (turns_input/1e6 × in_price)
                   + (output_tokens/1e6 × out_price)
blended = ai_cost_per_ticket
        + escalation_rate × human_cost_per_ticket

That quadratic growth is why long conversations get expensive fast — and why trimming history or summarising it matters.

Tips to lower cost per ticket

Cap conversation length, summarise old turns instead of re-sending them, and route trivial intents to a cheap model (GPT-4o mini, Claude Haiku, Gemini Flash) while reserving a frontier model for hard cases. Most of the bill hides in re-sent context, not in the final answer.