Chatbot budget guide
AI Chatbot Cost Calculator Guide
A chatbot budget depends on more than one prompt and one answer. Estimate each conversation turn, the chat history resent with it, retrieved context, retries, and the number of conversations your product handles each month.
Start with one conversation
Collect several real or realistic chat transcripts. For each turn, separate the user message, system instructions, conversation history, retrieved help articles, tool results, and assistant answer. The billable input often grows during a conversation because previous turns are sent again for context.
If the chatbot answers support questions, estimate a short case, a normal case, and an escalation case. A pricing estimate based only on the first message will usually miss long troubleshooting threads, repeated clarifying questions, and large knowledge-base snippets.
The chatbot cost formula
Monthly planning formula
Input per conversation = system prompt + user messages + resent chat history + retrieved context + tool results
Output per conversation = assistant answers + summaries + tool-call arguments
Monthly cost = conversations per month x turns per conversation x model token prices, adjusted for retries and escalations
Use the calculator in text mode for a sample turn, then multiply by average turns per conversation. For products with a document or help-center search step, combine this guide with the RAG cost guide so retrieved context is not treated as free.
Inputs to estimate before launch
- System prompt size: policies, role instructions, examples, formatting rules, and tool descriptions.
- Average user message size: short questions, pasted errors, order details, or long support descriptions.
- History window: how many prior turns are included before the app summarizes or trims the thread.
- Retrieved context: help articles, account data, logs, search results, or product records sent to the model.
- Answer length: concise answers are cheaper than long step-by-step replies for every turn.
- Retry and escalation rate: validation failures, tool errors, policy checks, and human handoff summaries.
Example support chatbot estimate
A support chatbot uses a 1,500-token system prompt, receives 300 input tokens per customer turn, retrieves 1,800 tokens from the help center, and answers with 500 output tokens. If a normal conversation has five turns, the cost estimate should include the growing chat history across those turns, not five copies of the first turn.
For a launch forecast, multiply the per-conversation estimate by expected daily conversations and by 30 days. Add a separate high-end estimate for users who paste logs, attach long text, or ask follow-up questions until the issue is solved.
Ways to reduce chatbot token spend
Summarize older chat history, cap retrieved article length, remove unused tool fields, keep answers concise by default, route simple FAQ answers to a lower-cost model, and cache stable instructions when the provider supports it. If the same prompt prefix is reused often, compare the result with the prompt caching cost guide.
Before launch, also check token rate-limit capacity. A chatbot can look affordable on monthly volume but still fail during a support spike if too many long conversations happen in the same minute.
Measure after launch
Log input tokens, output tokens, cached tokens when available, retrieved-context size, turns per conversation, retries, escalations, and cost per resolved issue. Replace estimates with provider usage fields as soon as the chatbot has enough real conversations to represent normal traffic.