Kodakod Team
There's no single answer to “How much does a chatbot cost?”, but the factors that drive the price are clear. From setup to monthly usage, here's how to plan your chatbot budget the right way.
In short: The cost of an AI chatbot has two parts: a one-off setup fee and a monthly running cost that changes with usage. Setup is driven by scope, the state of your knowledge base, the number of channels and integrations; the monthly cost is driven by model choice, conversation volume, channel fees and maintenance. Starting with a single scenario and calculating up front how many requests it will resolve each month is the most reliable way to make the budget predictable.
Asking “How much does an AI chatbot cost?” is a lot like asking “How much does a house cost?”: the answer depends on what you build. An assistant that answers the FAQs on your website and a system that tracks orders on WhatsApp, creates records in your CRM and hands tricky cases to a human agent are not built on the same budget.
The good news is that the factors behind the price are well known. In this article we break down what makes up the cost of a chatbot project, how the monthly running cost is calculated and how to plan your budget the right way.
The cost has two parts: setup and running
When people discuss a chatbot budget, they usually think only of the setup fee. But once an AI assistant goes live, it generates a small but steady cost for every day it runs:
| Item | When is it paid? | What does it cover? |
|---|---|---|
| Setup | Once, at the start of the project | Scenario design, knowledge base preparation, integrations, testing and launch |
| Running | Monthly, based on usage | Model usage (API) fees, hosting, channel fees and maintenance |
The right question isn't “How much does it cost to build a chatbot?” but “How many requests will this assistant resolve each month, and what will that cost us?”
The 7 factors that set the price
1. Use case and scope
Will the assistant only provide information, or will it also get things done? Answering frequently asked questions is the narrowest scope. Scenarios that take action, such as checking an order, booking an appointment or preparing a quote, each mean a separate flow and a separate round of testing. Cost grows with scope, which is why the healthiest starting point is a single scenario with clearly measurable results.
2. The state of your knowledge base
Today's good chatbots run on a RAG (Retrieval-Augmented Generation) architecture: before generating an answer, they look up your documents, product catalog and FAQs. If your information is up to date and well organized, setup moves quickly. If it is scattered across PDFs, email threads and people's memories, it first has to be gathered and cleaned up. This is the most underestimated item in most projects.
3. Number of channels
An assistant that runs only on your website is a different job from one that runs on your website, WhatsApp, Instagram and an internal dashboard at the same time. Every channel has its own rules, message formats and approval processes, and some channels add their own fees. Since 1 July 2025, Meta has charged per message on the WhatsApp Business Platform: only messages from a business to a user are charged, and only when they are delivered; messages users send you are free. The rate depends on the message category (for example marketing, utility, authentication and service messages) and on the recipient's country, and Turkey is listed as a separate market on the rate card. Each business phone number gets 1,000 free service messages a month (Meta: WhatsApp Business Platform pricing). Meta updates this table from time to time, so calculate your budget with the rates in force on the day you write it. We covered the setup steps in our WhatsApp chatbot setup guide.
4. Integrations
The assistant's real value shows up once it is connected to your systems: pulling order status from your e-commerce platform, saving a new lead in your CRM or reading stock levels from your ERP. The cost of each integration depends on how good an API the other system offers. Working with a platform that has ready-made connectors is easy; connecting to a legacy or closed system takes extra development.
5. Model choice and conversation volume
AI models are usually billed by the amount of text they process (tokens), priced per million tokens. The text sent to the model (input) and the text the model generates (output) are priced separately, and output costs more; on Anthropic's price list, an output token costs five times as much as an input token (Anthropic: Claude pricing). Within a single provider's list, the input price gap between the most economical model and the most powerful one can exceed 100 times (OpenAI: API pricing). That is why you don't need the most powerful model for every task: simple questions can be handled by a fast, economical model and complex requests by a stronger one. Using prompt caching for the fixed instructions and information resent in every conversation also lowers the cost; at Anthropic, input read from the cache is billed at 10 percent of the standard price or less. Batch jobs that don't need an instant answer, such as summarising conversation logs overnight, can run through the Batch API at a 50 percent discount. We compared which model stands out for which job in our article on the most powerful AI models of 2026.
6. Security, compliance and hosting
If the assistant processes personal data such as names, phone numbers or order details, data protection obligations kick in; in Turkey that means KVKK, the national data protection law. These include privacy notices, data minimization, access controls and record keeping. The AI recommendations published by Turkey's Personal Data Protection Authority advise building systems around data protection by design, preferring anonymisation where the same result can be reached without processing personal data, carrying out a privacy impact assessment where high risk is foreseen, and defining at the start of the project which parties act as data controller or data processor (KVKK: Recommendations on the Protection of Personal Data in the Field of Artificial Intelligence, in Turkish). If the model provider is abroad, the rules on transferring personal data out of the country also have to be taken into account. If the data has to stay in the country or on your own servers, hosting costs go up. Skipping this item is cheap in the short term and expensive in the long run; we covered the framework in detail in our article on responsible AI.
7. Maintenance and continuous improvement
A chatbot is not a “set it and forget it” product. As your products, prices and campaigns change, the knowledge base has to change too. As real conversations are reviewed and unanswered questions are added, the assistant gets more accurate every month. Building this improvement loop into the budget from the start keeps the assistant from becoming useless a few months after launch.
How do you calculate the monthly running cost?
You don't need a complicated tool to estimate the model fee. Four numbers are enough:
- Monthly conversations: the number of requests you receive today by phone, email and messaging is a good starting point.
- Average messages per conversation: an order query ends in a few messages; product advice takes longer.
- Input and output tokens per message: input includes not just the user's question but also the system instructions, the conversation history and the document passages retrieved through RAG. Input is usually much larger than output.
- The chosen model's unit price: taken from the provider's current pricing page.
The model cost is roughly conversations × messages × (input tokens × input price + output tokens × output price). On top of that come channel fees (WhatsApp messages, for example), hosting and maintenance. Run the calculation for two scenarios: expected volume and twice that. That way a spike during a campaign won't catch the budget off guard.
Off-the-shelf chatbot tool or custom development?
There are also off-the-shelf chatbot tools available on a monthly subscription. Which one is right depends on your needs:
| Off-the-shelf tool | Custom development | |
|---|---|---|
| Upfront cost | Low | Higher |
| Customization | As far as the tool allows | As far as you need |
| Integration | Limited to built-in connectors | Any system with an API |
| Data control | With the vendor | With you |
| Long-term cost | Rises with users and volume | Predictable, based on usage |
For a small business answering a limited set of questions, an off-the-shelf tool may be enough. If the assistant will connect to your systems, touch sales or handle sensitive data, custom development is usually more economical in the long run. We covered a similar decision for software in general in our article on off-the-shelf versus custom software.
5 questions to ask when planning your budget
- How many requests will the assistant handle each month? This is the basis of your running cost.
- Which requests will it fully resolve, and which will it hand to a human? This clarifies the scope.
- Which systems will it connect to? This sets the integration cost.
- What personal data will it process? This surfaces your security and compliance requirements.
- How will we measure success? This lets you track the return on the investment.
Questions to ask your vendor
- Is the model fee included in the quote, or billed separately? If included, up to what volume?
- Which model is used, and can it be changed? Being tied to a single provider leaves you exposed when prices change.
- Where are conversation logs stored, for how long, and who can access them?
- What does the assistant do when it doesn't know the answer? It should hand off to an agent rather than make something up.
- Who will update the knowledge base, and how often?
- Is there reporting? Can you see the resolution rate and the hand-off rate?
Common mistakes
- Comparing only the setup fee. A low setup fee can hide a high and unpredictable monthly cost.
- Trying to automate everything from day one. Building ten scenarios at once multiplies both cost and testing time; starting with one delivers results faster.
- Starting before the knowledge base is ready. An assistant built on scattered information gives wrong answers and erodes trust.
- Not designing the hand-off path. If a request the assistant can't resolve hits a dead end, the customers lost can outweigh what the chatbot brings in.
How do you measure the return?
A chatbot's cost means little on its own; it has to be read alongside what it brings in. After launch, track these metrics:
- Automatic resolution rate: conversations closed without reaching an agent.
- First response time: how long a customer waits for an answer.
- Hand-off rate: shows where the assistant falls short.
- Conversion: the share of visitors who buy or fill in a form after talking to the assistant.
We explained how these metrics improve in customer service in detail in our article on AI-powered customer service.
With the right scenario, a chatbot budget isn't a cost but a measurable investment. At Kodakod, we review your conversation volume, data and systems together in a discovery call and lay out your setup and monthly running costs transparently from the start. You can explore our packages and their scope on our AI integration and chatbot development page.
Frequently asked questions
Mainly four things: the number of conversations per month, how long they are, the token price of the model used and channel fees. Hosting and maintenance come on top. As volume grows, so does the model cost, which is why routing simple questions to an economical model matters.
Yes, on top of the model fee there are Meta's message fees. Meta charges only for delivered messages sent from a business to a user; the rate depends on the message category and the recipient's country, and each phone number gets 1,000 free service messages a month. Check the current table on Meta's pricing page.
No. Simple tasks such as FAQs and order queries are answered well by an economical model. When the powerful model handles only complex requests, quality is preserved and cost drops noticeably.
If personal data is processed, you need a privacy notice, data minimization, access controls, a retention period and record keeping. If the model provider is abroad, the rules on cross-border transfers also have to be assessed. These questions should be settled with your legal adviser at the start of the project.
For an assistant that answers a limited, fixed set of questions, won't connect to your systems and won't handle sensitive data, an off-the-shelf tool is often enough. If you need integrations, a sales flow or data control, custom development is more economical in the long run.
