Estimating your AI provider bill for a business chatbot
Your AI provider bill depends on tokens used per conversation multiplied by your monthly conversation volume, at your chosen model's per-token rate. Estimate by counting a typical conversation's length, checking your provider's current published pricing, and multiplying by expected volume — then set a spend limit in your provider account as a safety net.
- AI providers bill by tokens (roughly word-fragments), not by conversation — always check the current rate on the provider's own pricing page.
- Estimate tokens per conversation from a real test conversation, including your system prompt and knowledge base context, not just the visible reply.
- A smaller/cheaper model tier is often enough for FAQ-style bots; save the larger model for complex reasoning or sales conversations.
- Prompt caching can reduce cost when the same system prompt and knowledge base context repeat across many conversations.
- Set a monthly spend limit directly in your OpenAI, Anthropic or Google account — this is the real safety net against surprise bills.
How AI providers actually bill you
OpenAI, Anthropic (Claude) and Google (Gemini) all bill by tokens, small chunks of text, roughly three-quarters of a word each, split into input tokens (what you send: system prompt, conversation history, knowledge base context, the user's message) and output tokens (what the model generates back). Rates differ by model tier and change over time, so this article deliberately does not quote a number. Check the current rate directly on OpenAI's API pricing page or Google's Gemini API pricing page before estimating, and do the same on Anthropic's own pricing page if Claude is your chosen provider.
What goes into one conversation's token count
- System prompt: your instructions to the AI agent, sent with every message in the conversation.
- Knowledge base context: if the agent uses RAG (retrieval-augmented generation) to answer from your documents, relevant chunks are added to each request.
- Conversation history: earlier turns in the same conversation are often re-sent so the model has context, which means longer conversations cost more per additional message, not less.
- The reply itself: output tokens, usually a smaller share of the total than people expect.
A short FAQ-style exchange might use a few hundred tokens; a longer sales or support conversation with several back-and-forth turns and a large knowledge base context can use several thousand.
A simple monthly estimate formula
Monthly cost ≈ Conversations per month × Average tokens per conversation × Your model's per-token rate (from the provider's current pricing page).
| Step | How to get it |
|---|---|
| Conversations per month | Your own WhatsApp/chat volume from analytics or Business Manager. |
| Average tokens per conversation | Run 5–10 real test conversations through your agent and check the token usage your provider's dashboard reports. |
| Per-token rate | Copy directly from the provider's live pricing page for the exact model you selected. |
Run this for a low-volume month and a high-volume month (e.g. a festive campaign) to see the range you should budget for.
Model tiers: match the model to the task
All three major providers offer a range of models — smaller, faster and cheaper versus larger and more capable. A bot answering repetitive FAQs from a knowledge base (store hours, delivery policy, product specs) rarely needs the most powerful model tier. A bot doing complex sales qualification, nuanced troubleshooting, or reasoning over ambiguous requests benefits more from a stronger model. Since Convo360 lets you choose the model per agent, you can run your FAQ bot on a lighter model and your sales agent on a stronger one, rather than paying the same rate for every conversation.
Scaling the formula to your monthly volume
Once you have a per-conversation token estimate, monthly volume is where the real budget number comes from. Work through it in this order, filling in your own figures:
| Step | Formula shape | Your number |
|---|---|---|
| 1. Tokens per turn | From a real test conversation in your provider's dashboard | ___ |
| 2. Turns per conversation | Average back-and-forth exchanges before resolution | ___ |
| 3. Tokens per conversation | Tokens per turn × turns per conversation | ___ |
| 4. Conversations per month | Your own WhatsApp/chat volume | ___ |
| 5. Monthly tokens | Row 3 × Row 4 | ___ |
| 6. Monthly cost | Row 5 × your provider's current published rate for the chosen model | ₹___ |
The reason this is broken into six steps rather than one line is that each step is something you can sanity-check independently — if your final number looks wrong, you can trace it back to whichever row is off, rather than re-guessing the whole estimate.
Reducing cost without reducing quality
- Prompt caching. Some providers offer reduced rates when the same system prompt and context repeat across requests — check whether your provider and model support this, since it can meaningfully cut cost for high-volume bots with a stable system prompt.
- Trim your knowledge base context. Only surface the most relevant document chunks per query rather than dumping your entire knowledge base into every request.
- Cap conversation history length. Very long-running conversations can be summarised periodically instead of resending every earlier turn in full.
- Set a spend limit. Every major provider lets you set a monthly spend cap in your account dashboard — this is your real safety net against a runaway bill, not just estimation.
- Review usage weekly in month one. Cross-check your provider's dashboard against your own volume estimate early on, so a miscalibrated prompt or an unexpected loop is caught before it becomes a full month's bill.
How this works on Convo360
Convo360 is BYOK — you connect your own OpenAI, Anthropic Claude or Google Gemini key, choose the model per agent, and the platform adds zero markup on top of your provider's own token rate. This means the estimate above is your actual cost, with nothing hidden in a platform fee layered on usage. Your provider's own dashboard remains the authoritative source for exact spend, and the Provider Health monitor on Convo360 alerts you if your key runs into an error (such as hitting a spend limit), so a billing issue does not silently stop your bot from responding. If you run several brands or departments as separate workspaces, each can hold its own AI provider key, which makes it straightforward to see which workspace's usage is driving the bill when you check your provider's dashboard.
Example conversation
Connect your own AI key on Convo360
Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.
Frequently asked questions
How much does an AI chatbot cost per conversation?
It depends on your model choice, conversation length and each provider's current token rate, which changes over time. Check OpenAI, Anthropic or Google's own pricing page and run a real test conversation through your agent to estimate your specific cost.
Does a longer conversation cost more?
Yes. Because conversation history is often re-sent with each new message for context, a conversation with many back-and-forth turns uses more cumulative tokens than a short one-question exchange.
Can I limit how much I spend on AI tokens?
Yes — OpenAI, Anthropic and Google all let you set a monthly spend limit directly in your provider account. This is the most reliable way to avoid an unexpected bill.
Does Convo360 add a markup on top of AI provider costs?
No. Convo360 is BYOK — you connect your own AI provider key and pay that provider directly at their published rate, with zero markup added by the platform.
Which AI model is cheapest for a chatbot?
Rates and model line-ups change over time across providers, so check each provider's current pricing page directly rather than relying on a fixed comparison. Smaller models in any provider's range are typically cheaper than their largest models.
Related reading
Available across India — see WhatsApp & AI calling automation by state and city.