$10 a Month Is a Budget, Not a Chatbot Guarantee

01 / Split a $10 chatbot budget before writing code

A support chatbot can burn a tiny budget by answering questions it should never have received. Start with a narrow job: explain shipping rules, locate setup instructions, and point customers toward the right support form. Leave refunds, account changes, and unusual disputes to a person.

For a first-month experiment, allocate $7 to model usage and keep $3 uncommitted for unexpected usage. This is a planning budget, not a quoted service package. It also assumes your existing website can host the chat interface and a small backend; new hosting or paid search infrastructure would need separate funding.

02 / Calculate the cost of 10,000 replies

Use this formula: input tokens multiplied by the input rate, plus output tokens multiplied by the output rate. Divide token counts by one million when prices are quoted per million tokens.

For arithmetic only, suppose input costs $0.20 per million tokens and output costs $1 per million. These are hypothetical rates, not CheapRelay prices. A reply using 1,200 input tokens and 200 output tokens would cost $0.00044. Ten thousand such replies would cost $4.40.

That estimate excludes retries, extra model calls, cache charges, and longer conversations. Ten thousand replies also means model responses, not necessarily ten thousand resolved tickets. Check the current rate for your chosen model and replace both example prices before making a budget commitment.

03 / Connect one backend route, not a browser-held key

Your minimum architecture is a chat box, a server endpoint, a small collection of support articles, and one model call per answer. Keep your API key on the server in an environment variable. Never ship it in browser JavaScript.

CheapRelay provides a pay-as-you-go API relay with no monthly fee or subscription. Its OpenAI-compatible base URL is https://api.cheapapirelay.com/v1. Use an OpenAI SDK on your backend, set that base URL, and copy the current model identifier from the service rather than guessing it.

Sign up for CheapRelay when you are ready to test a small batch of real support questions. Start by evaluating one model, such as Gemini 2.5 Flash, against your questions and current pricing; do not assume its name guarantees either adequate quality or the lowest cost.

04 / Send three relevant passages, not your entire help center

Begin with 20 to 30 clean FAQ entries stored in a file or database. Match questions using keywords and aliases: “delivery,” “shipping,” and “parcel,” for example. Send at most three relevant passages with the customer’s question. You can add more sophisticated retrieval when simple matching demonstrably fails.

Tell the model to answer only from those passages, avoid inventing policies, and offer a support handoff when information is missing. Treat customer text and retrieved documents as information, not instructions that can override these rules.

Keep only the last two conversation turns initially. Ask for short answers and set an output-token limit. Your 1,200-input-token estimate must include instructions, passages, history, and the latest question—not just what the customer typed.

05 / Block waste before it reaches the model

• ■ Answer exact, approved FAQ matches without a model call.
• ■ Limit message length and apply per-user and global request limits.
• ■ Allow at most one retry for a temporary failure; count it as another potentially billable attempt.
• ■ Reuse public FAQ answers only when the question and policy version match. Never share personalized answers between customers.
• ■ Keep account actions outside this first version and route sensitive cases to human support.

CheapRelay bills real tokens, including cache reads and writes. Do not treat provider-side caching as free. Application-side FAQ reuse avoids a model request entirely, but still needs expiry when policies change.

06 / Pause at $7 instead of hoping the bill stays small

Record token usage, estimated cost, retries, and handoffs for each request without logging unnecessary personal data. Compare your ledger with billed usage. Test 30 representative questions before launch, including missing-policy questions and attempts to override instructions.

At $5, review spending. At $7, disable new model requests and show static FAQs plus the support form. Reserve estimated cost before dispatching concurrent requests, then reconcile actual usage afterward. A local counter is not a guaranteed billing cap, so use provider-side limits too if available. Expand only when measured costs and useful answers justify it.