Chatbot Development Cost (2026): Support Bots, RAG Assistants and What They Cost to Run

Chatbot development costs about $4,200 to $48,000 for a simple AI bot that answers from a small set of documents on one channel, and about $13,300 to $171,200 for a production support chatbot with retrieval over your knowledge base, a web widget, handoff to human agents and proper testing, depending on the team’s rates. Bots that take actions in your systems across several channels cost more, and every AI chatbot adds a monthly model bill.

We build AI assistants and connect them to the systems behind them: helpdesks, CRMs, ERPs and in-product data. The figures below are planning ranges from that engineering work combined with market hourly rates. They are not a quote and not what any client paid. This guide covers customer-facing and internal chatbots specifically; for AI features and AI products in general, see our guide to AI development cost.

Cost at a glance

Costs are hours multiplied by market rate bands. 2026 outsourcing rate surveys put Eastern Europe at about $35 to $80 per hour, and North America at $80 to $180 or more; we use $80 to $160 for the US band. The low end assumes the low hours at the low rate, the high end the high hours at the high rate, and calendar time assumes about 30 productive hours per engineer per week.

What drives the cost

1. What the bot is allowed to do. A bot that only answers questions is the cheapest kind. A bot that looks up an order, reschedules an appointment or updates a CRM record needs authenticated calls to your systems, confirmation steps, error handling and audit logs. Each action is a small integration project of its own, which is why “actions in your systems” is 60 to 200 hours in the estimator.

2. The state of your knowledge base. Retrieval-augmented generation (RAG) means the bot searches your content and answers from what it finds. If the help center has three outdated versions of the refund policy, the bot will find all three. Cleanup, filling gaps for the top questions and setting up automatic sync usually take longer than wiring up the model.

3. Handoff to humans. A support bot must know when to stop: when the user asks for a person, when the question is out of scope, or when it cannot find a grounded answer. Good handoff opens a ticket or live chat in your helpdesk with the full transcript, so the agent does not ask the customer to repeat everything.

4. Channels. A website widget is the default. WhatsApp, Slack, Microsoft Teams, Telegram or a mobile app each bring their own message formats, limits, approval processes and, in WhatsApp’s case, per-message fees. Plan 40 to 120 hours for extra channels.

5. Permissions for internal bots. An internal assistant over wikis, shared drives and tickets must not show a junior employee the board’s salary spreadsheet. Permission-aware retrieval with SSO means indexing access rights along with content and filtering on every query.

6. Evaluation and guardrails. Before launch you need a set of real customer questions with expected answers, scored automatically, plus tests for jailbreak attempts and off-topic requests. Teams that skip this find out about wrong answers from customers.

7. Languages and volume. Each extra language needs tested sources and evaluation cases. High volume makes model choice and caching worth engineering time, because every request is billed.

Turn components on or off below to size your bot, and pick a rate band.

The default selection is a typical customer support chatbot: discovery, knowledge base cleanup, the chat backend, retrieval, a web widget, helpdesk handoff, evaluation, analytics and a pilot. That adds up to 380 to 1,070 hours, the “support chatbot with RAG” row above. Adding actions, extra channels and an admin panel moves you toward the transactional row.

Cost by component

Running costs after launch

Model usage is the cost people worry about most, and for a typical support bot it is often the smallest line. List prices per million tokens on the official pricing pages in November 2026: Claude Haiku 5.5 at $0.10 input and $0.50 output (for prompts up to 100,000 tokens), Claude Sonnet 5.5 at $2 and $10, Gemini 3.8 Flash at $0.75 and $3.75 through December 31, 2026 and $1.50 and $7.50 from January 1, 2027, and OpenAI’s gpt-6-luna at $0.10 and $0.50.

A RAG conversation sends the system prompt, retrieved passages and chat history with every turn. Assume a four-turn conversation with about 3,000 input and 300 output tokens per turn: roughly 12,000 input and 1,200 output tokens in total. Then 10,000 conversations a month cost about $18 on Claude Haiku 5.5, $135 on Gemini 3.8 Flash and $360 on Claude Sonnet 5.5. Anthropic’s pricing page also lists much cheaper cache hits for repeated prompt prefixes, which suits chatbots because the system prompt rarely changes.

Embedding your documents for search is cheap: OpenAI lists text-embedding-3-small at $0.02 per million tokens, so indexing 10 million tokens of content costs about $0.20.

Other running costs:

• Channel fees. Meta’s WhatsApp Business Platform pricing page says that from October 1, 2026 each business phone number gets 1,000 free service messages a month, with a charge per message after that. Web chat has no per-message fee.
• Hosting: the backend, a database, a vector index and logs. For a single support bot this is usually a modest cloud bill.
• Content upkeep: someone has to review unanswered questions weekly and update the knowledge base. This is the habit that makes bots improve.
• Maintenance: model retirements and price changes, helpdesk API updates and evaluation reruns. Plan a few days of engineering per quarter at minimum.

Build, buy or combine

Buy when your answers already live in your helpdesk’s help center and the bot does not need to act in other systems. Helpdesk vendors sell AI agents for exactly this. Intercom, for example, lists its Fin AI agent at $0.99 per outcome, on top of seat prices from $19 per seat per month on annual billing for the Essential plan. That is fast to launch and needs no engineers. It also scales with volume: 5,000 outcomes a month is $4,950, every month.

Combine when the vendor bot is fine for answers but you need one or two custom actions or a private data source. A small integration that exposes order status or account data to the existing bot is often cheaper than a full build.

Build when the bot must work with your own systems and rules, needs private data with per-user permissions, must run on several channels, or when per-outcome fees at your volume exceed what it costs to run your own. A custom bot also lets you choose and switch models freely. If your bot should plan and execute multi-step tasks on its own rather than answer and hand off, it is closer to an agent; see our breakdown of AI agent development cost.

How to get an accurate estimate

1. Export your top 100 questions from the helpdesk, chat logs or internal channels, with the answers your best agents give.
2. List the sources the bot should answer from and rate how current they are.
3. Decide what the bot may do beyond answering, and which systems it needs for that.
4. Choose channels and languages for launch, and which can wait.
5. Define handoff rules: when the bot must pass to a person, and where the conversation goes.

With these five inputs, an engineering team can give you a component-level estimate and an expected monthly model bill. Without them, any fixed price is a guess.

From our work

We have connected AI services to existing CRM, ERP and marketing tools so that automation runs inside the workflow people already use. Part of that work is chatbots and virtual assistants that answer inquiries, book appointments and send replies around the clock, alongside sentiment analysis of feedback and fraud alerts. Customer response times improved by 40%. Read the AI business automation case.

For Lokad, a supply chain software company, we built an assistant inside its platform that answers questions about the product, writes scripts in Lokad’s own Envision language and suggests the right Lokad specialist to contact. It keeps context in 99% of conversations, suggests a relevant contact in about 78% of cases, and users spend about 60% less time searching documentation or looking for the right person. Generated scripts are tested against the real Envision runtime before users rely on them. Read the Lokad case, and our post on validating LLM-generated code for how that testing works.

Where OSKI fits

OSKI is an engineering company with a team mostly in Ukraine, registered in Estonia, working in .NET, React and AI. We build chatbots end to end: knowledge base preparation, retrieval, helpdesk and CRM integrations, evaluation and monitoring. See our AI development services, integrations and maintenance and support. Send us your top questions and the systems the bot should touch, and we will return a component-level estimate with an expected monthly running cost.