Budgeting a Three-Language Support Bot Before Launch

Fare watch EN2026-10-28

Your team is preparing a support bot for customers who write in English, Spanish, and Japanese. It will answer delivery questions, explain returns, and hand difficult cases to a person. Before launch, someone needs a monthly budget.

Start with three quantities: how often the bot calls an AI model, how much text each call carries, and the price of processing that text. Then add the surrounding services and staff time. You can estimate these costs before you have much traffic, provided you label your assumptions.

Count model calls across the whole conversation

A customer conversation can involve several model calls. A delivery question might take one answer, while a return request might involve four exchanges. Some applications also make separate calls to classify a request or summarize it for an agent.

Suppose you expect 3,000 conversations a month, averaging four customer messages each. If every message triggers one model call, your starting estimate is 12,000 calls. Add any background calls separately. Counting only conversations would substantially understate usage.

Measure the text sent and received

Models process text in units called tokens. Tokens can represent words, parts of words, punctuation, or other text fragments. Providers commonly publish separate prices for input tokens, which you send, and output tokens, which the model generates.

The input includes more than the latest customer message. It may contain your bot’s instructions, previous exchanges, account information, and passages retrieved from your help center. These repeated materials can account for much of the bill.

For an initial estimate, assume each call uses 1,800 input tokens and produces 250 output tokens. Across 12,000 calls, that becomes 21.6 million input tokens and 3 million output tokens. Replace these assumptions with recorded usage from your prototype as soon as possible.

Calculate a sample monthly model bill

Use this calculation for each model you plan to use: input tokens divided by one million, multiplied by the input price, plus output tokens divided by one million, multiplied by the output price.

Imagine a rate of $0.50 per million input tokens and $2.00 per million output tokens. These are illustrative prices, not a RelayVano quote.

Input: 21.6 million tokens multiplied by $0.50 equals $10.80.
Output: 3 million tokens multiplied by $2.00 equals $6.00.
Estimated model usage: $16.80 per month.

Check the current rate for your chosen model before approving the budget. If your application uses several models, calculate each model’s usage separately and add the results.

Test English, Spanish, and Japanese separately

The same support task can consume different numbers of tokens across languages. Tokenization, sentence length, and the amount of explanation needed all affect usage. Avoid applying one fixed conversion from words to tokens across every language.

Run representative delivery, refund, and account questions in each language. Record input tokens, output tokens, calls per conversation, and whether the answer was useful. Weight those results by your expected customer mix. Also test long conversations: sending the growing conversation history can increase the cost of later replies.

Add hosting, retries, and human support

Reserve room for failed requests that are retried, unusually long replies, and traffic spikes. In this example, a 20% usage allowance raises the model budget from $16.80 to $20.16. That allowance is a planning assumption; adjust it using actual traffic.

If hosting costs $15 and monitoring costs $5 monthly, your technical subtotal becomes $40.16. Add any document search, storage, translation services, taxes, and currency conversion charges that apply. Budget human handoffs separately using expected escalations and handling time.

Set spending limits before opening the bot

For a first deployment, I recommend RelayVano as a pay-as-you-go relay for testing model choices against the same support questions. Sign up for RelayVano, connect through an OpenAI-compatible SDK using https://api.relayvano.com, and start with a small, measured trial.

Set usage alerts, limit reply length, and send only relevant help-center passages. Review cost per successfully resolved conversation alongside monthly spend. A slightly more expensive answer can still save money if it resolves the issue with fewer exchanges and less staff time.