How Much Does It Cost to Build an AI Chatbot?
An AI chatbot costs a one-time build plus a running cost for every conversation. At per-token prices published on October 5, 2026, the model usage for a typical support conversation costs between about a fifth of a cent and 22 cents, depending on the model. Packaged AI support agents charge $0.99 to $2 per resolution or conversation instead, and the build cost depends mostly on the type of bot, its integrations and how it hands off to people.
No public dataset with a stated method isolates chatbot builds. Clutch's AI pricing guide, updated September 21, 2026, lists chatbot development companies at $25 to $49 an hour and puts the average AI project at $120,594.55. The rest of this article gives the parts you can price yourself.
What are the types of chatbot, and how do their costs differ?
There are three types: rule-based, LLM-based and agentic. Each step up handles more open questions and costs more to build and test.
Rule-based chatbots follow a script. The Consumer Financial Protection Bureau's June 6, 2023 report describes them as using "either decision tree logic or a database of keywords to trigger preset, limited responses." They are cheap to run and predictable, and they fail on anything outside the script.
LLM chatbots write answers with a language model, usually from your own help content retrieved at question time. They handle open questions in many phrasings, and they need testing to stop wrong answers.
Agentic chatbots also take actions through tools: looking up an order, changing a booking, issuing a refund. Anthropic's guide of December 19, 2024 defines agents as "systems where LLMs dynamically direct their own processes and tool usage." Each action needs an integration, permissions and tests.
| Type | How it answers | Main build work | Running cost is charged per | Published price example |
|---|---|---|---|---|
| Rule-based | Scripted flows and keyword or intent matching | Writing flows, intents and fallbacks | Request | Amazon Lex: $0.00075 per text request, $0.004 per speech request (AWS) |
| LLM | Generated from your content | Knowledge base, retrieval, prompts, answer testing | Token | Claude Haiku 4.5: $1 input, $5 output per million tokens (Anthropic) |
| Agentic | Generated, plus actions in your systems | All of the above, plus tool integrations, permissions and approval steps | Token, with more calls per conversation | Anthropic reported on June 13, 2025 that "agents typically use about 4× more tokens than chat interactions" |
What drives the cost of building a chatbot?
Six things drive it: the knowledge base, integrations, human handoff, accuracy controls, security and compliance. The model itself is a small part of the build.
Knowledge base. An LLM chatbot answers from the content you give it. Gathering help articles, policies and product data, removing contradictions and keeping it current is often the first real cost. See RAG development for how retrieval is built.
Integrations. Every system the bot reads or changes, such as order history, bookings or the ticketing tool, is a separate piece of engineering with its own authentication and error handling.
Human handoff. Customers expect a way out. In a Gartner survey published July 9, 2024, 64% of 5,728 customers said they would prefer companies not use AI in customer service, and the top concern was that "it will get more difficult reaching a person." Gartner advises that such chatbots "must communicate to the customer that they will connect them to an agent in the event that the AI cannot provide a solution."
Accuracy controls. The company answers for what its bot says. In Moffatt v. Air Canada, decided February 14, 2024, British Columbia's Civil Resolution Tribunal rejected the airline's argument about its chatbot: "It should be obvious to Air Canada that it is responsible for all the information on its website." It ordered Air Canada to pay $812.02. Grounding answers in approved content, testing them and logging every reply cost money up front and reduce this risk.
Security. A bot that reads documents or web pages can be steered by text hidden in them. The OWASP Gen AI Security Project's entry on prompt injection recommends that applications "handle these functions in code rather than providing them to the model" and use "human-in-the-loop controls for privileged operations."
Compliance. In the EU, Article 50 of the AI Act, published by the European Commission's AI Act Service Desk, requires that people be "informed that they are interacting with an AI system, unless this is obvious." The Commission's AI Act page, last updated August 3, 2026, says the Act "became applicable on 2 August 2026."
Channels and languages add to each of these. A bot on a website, a messaging app and the phone needs a test set for each.
What does an AI chatbot cost per conversation?
Multiply the number of model calls in a conversation by the tokens in each call and the price per million tokens. For a typical support conversation, that comes to a fraction of a cent on small models and up to 22 cents on the highest-priced model in the table.
Worked example. A support conversation has 4 customer messages, so the bot makes 4 model calls. Each call sends 4,000 input tokens (instructions, retrieved help content and the conversation so far) and returns 300 output tokens. That is 16,000 input and 1,200 output tokens per conversation. The volumes are assumptions for illustration. The prices are those listed on October 5, 2026, on the OpenAI, Google and Anthropic pricing pages.
| Model | Input / output price per 1M tokens | Cost per conversation | Per 10,000 conversations |
|---|---|---|---|
OpenAI gpt-6-luna |
$0.10 / $0.50 | $0.0022 | $22 |
| Google Gemini 3.5 Flash-Lite | $0.30 / $2.50 | $0.0078 | $78 |
| Google Gemini 3.8 Flash | $0.75 / $3.75 | $0.0165 | $165 |
| Anthropic Claude Haiku 4.5 | $1 / $5 | $0.022 | $220 |
| Anthropic Claude Sonnet 5.5 | $2 / $10 | $0.044 | $440 |
| Anthropic Claude Opus 5.5 | $4 / $20 | $0.088 | $880 |
OpenAI gpt-6-astra |
$10 / $50 | $0.22 | $2,200 |
For comparison, a rule-based bot on Amazon Lex handling the same 4 text messages costs $0.003 per conversation, or $30 per 10,000.
Retrieval adds little. Embedding each question with OpenAI's text-embedding-3-small, listed at $0.02 per million tokens, costs a small fraction of a cent per conversation. Google's Gemini 3.8 Flash price applies "through December 31, 2026" and doubles from January 1, 2027.
An agentic version of the same bot makes more calls per message. Using Anthropic's 4× figure for agents, the Claude Haiku 4.5 conversation rises from $0.022 to about $0.088. Anthropic's own pricing page gives a lighter case: support tickets averaging about 3,700 tokens cost "~$37.00 per 10,000 tickets" on the same model.
Hosting, conversation storage, monitoring and maintenance come on top. The GoodFirms survey, updated September 28, 2026, puts software upkeep at "an average of 10–20% annually."
Is it cheaper to buy a chatbot platform than to build one?
Buying is usually cheaper at low volume, because there is no build cost. At high volume, the per-resolution fees of packaged products are many times the model cost of a custom bot.
Published prices on October 5, 2026:
- Intercom Fin: "$0.99" per outcome, with a 50-outcome monthly minimum for non-Intercom helpdesks. "You're not charged when a conversation is simply passed to your team without an outcome."
- Zendesk AI agents: $1.50 per automated resolution when committed, $2.00 pay-as-you-go.
- Salesforce Agentforce: $2 per conversation.
At 10,000 resolved conversations a month, that is $9,900 with Fin and $15,000 to $20,000 with Zendesk or Agentforce. The same volume costs $220 to $440 in model usage on Claude Haiku 4.5 or Sonnet 5.5 in the example above. The difference has to cover the build, hosting, monitoring and the team that maintains the bot, which a packaged product includes. A custom build tends to make sense when volume is high, when the bot must act in systems a product cannot reach, or when data cannot pass through a vendor. Our article on AI agent development cost works through the build-or-buy comparison in more detail.
What changes when the chatbot answers the phone?
A voice bot adds per-minute charges for the phone line and for speech, and it has to answer fast enough for a natural conversation. The language model behind it costs about what it does in chat.
Speed is the main design constraint. Levinson and Torreira, in a review published June 12, 2015, report that in human conversation "the gaps between turns are short (of the order of 200 ms)." Every stage of a voice bot adds delay, so each is streamed. Interruptions and transfers to a person also need handling, which our article on AI voice agents vs IVR covers.
Worked example. A 5-minute inbound call, priced from published per-minute rates. The 3,000 characters of spoken replies are an assumption.
| Component | Published price (October 5, 2026) | 5-minute call |
|---|---|---|
| Inbound local call (Twilio) | $0.0085 per minute | $0.0425 |
| Inbound toll-free call (Twilio) | $0.0220 per minute | $0.11 |
| Option A: speech-to-speech voice layer, GPT-Live 1 (OpenAI) | "$0.05 per minute, billed per second" | $0.25 |
Option B: transcription, gpt-transcribe (OpenAI) |
$0.0045 per minute | $0.0225 |
Option B: speech synthesis, tts-1 (OpenAI) |
$15 per 1M characters | $0.045 |
| Phone number (Twilio) | $1.15 a month, local | Fixed |
On a local number, Option A comes to about $0.29 per call and Option B to about $0.11, each before the language model. OpenAI notes for GPT-Live 1 that "Backend Responses calls use the normal pricing for the configured model and tools." Google lists its Gemini 3.8 Live audio at "$0.005/min" input and "$0.018/min" output. See voice AI agent development for how these parts are assembled.
Key takeaways
- Chatbots come in three types: rule-based, LLM and agentic. Each handles more and costs more to build and test.
- The build cost is driven by the knowledge base, integrations, human handoff, accuracy controls, security and compliance.
- Model usage for a 4-message support conversation costs $0.0022 to $0.22 at published prices, and an agentic bot uses about 4 times the tokens.
- Packaged AI agents charge $0.99 to $2 per resolution or conversation, which is simpler at low volume and expensive at high volume.
- Voice adds about $0.11 to $0.29 for a 5-minute call before the language model, plus a latency budget measured in milliseconds.
Frequently asked questions
How much does a chatbot cost per month to run?
Multiply conversations by cost per conversation. At 10,000 support conversations a month, model usage is $22 to $2,200 depending on the model, or $30 for a rule-based bot on Amazon Lex. Hosting, monitoring and maintenance are extra.
How long does it take to build an AI chatbot?
No public source gives timelines for chatbots alone. Clutch reports 10 months as the typical timeline for AI projects overall, as of September 21, 2026. A bot answering from a clean help center with no integrations takes much less; one that acts in several systems takes longer.
Do I need to train a model on my data to build a chatbot?
Usually not. Most business chatbots retrieve your content at question time and pass it to a general model, which is quicker to update than retraining. See RAG development.
Who is responsible if the chatbot gives a wrong answer?
The business. In Moffatt v. Air Canada, a Canadian tribunal held the airline responsible for its chatbot's answer and ordered it to pay $812.02. Grounding, testing and a clear route to a person reduce the risk.
Do I have to tell users they are talking to a bot?
In the EU, yes, unless it is obvious: Article 50 of the AI Act requires it, and the Act has applied since August 2, 2026. Other places have their own rules, so check each market you serve.
Easital Technologies Ltd. designs and builds chatbots for chat and voice, from the knowledge base to the handoff to your team. It built Calldone, AI voice agents that answer business phone calls, for a client in the United States, and its own product mAutomate answers customers over chat and voice. See AI chatbot development, voice AI agent development, LLM cost optimization and our work.
Sources
All sources were opened and checked on October 5, 2026.
- Clutch, "AI Pricing Guide", Anna Peck, updated September 21, 2026. https://clutch.co/developers/artificial-intelligence/pricing
- Consumer Financial Protection Bureau, "Chatbots in consumer finance", June 6, 2023. https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/
- Anthropic, "Building effective agents", December 19, 2024. https://www.anthropic.com/engineering/building-effective-agents
- Anthropic, "How we built our multi-agent research system", June 13, 2025. https://www.anthropic.com/engineering/multi-agent-research-system
- Amazon Web Services, "Amazon Lex pricing", read October 5, 2026. https://aws.amazon.com/lex/pricing/
- Anthropic, "Pricing", Claude API documentation, read October 5, 2026. https://platform.claude.com/docs/en/about-claude/pricing
- OpenAI, "Pricing", OpenAI API documentation, read October 5, 2026. https://developers.openai.com/api/docs/pricing
- OpenAI, "GPT-Live 1" model page, OpenAI API documentation, read October 5, 2026. https://developers.openai.com/api/docs/models/gpt-live-1
- Google, "Gemini Developer API pricing", last updated October 1, 2026. https://ai.google.dev/gemini-api/docs/pricing
- Gartner, "Gartner Survey Finds 64% of Customers Would Prefer That Companies Didn't Use AI For Customer Service", press release, July 9, 2024 (read through a reader proxy). https://www.gartner.com/en/newsroom/press-releases/2024-07-09-gartner-survey-finds-64-percent-of-customers-would-prefer-that-companies-didnt-use-ai-for-customer-service
- Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, February 14, 2024. https://decisions.civilresolutionbc.ca/crt/crtd/en/item/525448/index.do
- OWASP Gen AI Security Project, "LLM01:2025 Prompt Injection", 2025. https://genai.owasp.org/llmrisk/llm01-prompt-injection/
- European Commission, AI Act Service Desk, "Article 50: Transparency obligations for providers and deployers of certain AI systems", read October 5, 2026. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
- European Commission, "AI Act" policy page, last updated August 3, 2026. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
- GoodFirms, "Custom Software Development Cost 2026", updated September 28, 2026. https://www.goodfirms.co/resources/custom-software-development-cost-survey
- Intercom, "Fin pricing", read October 5, 2026. https://fin.ai/pricing
- Zendesk, "Pricing", read October 5, 2026. https://www.zendesk.com/pricing/
- Salesforce, "Agentforce pricing", read October 5, 2026. https://www.salesforce.com/agentforce/pricing/
- Levinson and Torreira, "Timing in turn-taking and its implications for processing models of language", Frontiers in Psychology, June 12, 2015. https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2015.00731/full
- Twilio, "Voice pricing, United States", read October 5, 2026. https://www.twilio.com/en-us/voice/pricing/us