How much does it cost to build an AI chatbot?

AI chatbot development cost in October 2026 runs about $500 to $6,000 to set up a SaaS AI agent, $2,000 to $18,000 for a no-code bot on your documents, $12,500 to $90,000 for a custom chatbot, and $30,000 to $225,000 for an agent that acts in your systems. Model usage adds about $6 to $630 a month at 2,000 conversations.

The spread is wide because "chatbot" covers everything from a help-center widget switched on in an afternoon to an assistant that books jobs and updates a CRM.

On this page

• AI chatbot development cost at a glance
• What you pay for in a chatbot build
• Rule-based, LLM or RAG chatbot: which type you need
• Should you buy a SaaS AI agent or build a custom chatbot?
• When does building your own chatbot pay off?
• How much does the AI model cost per conversation?
• How caching, routing and price changes move the bill
• How hours and hourly rates set the build price
• Which features raise chatbot development cost the most?
• Chatbot cost by industry
• What voice chatbots and regulated industries add
• How much does an AI chatbot cost per month after launch?
• Worked example: first-year cost for a small service business
• How to lower the cost of building an AI chatbot
• Checklist before you ask for chatbot quotes

AI chatbot development cost at a glance

Every build figure in this guide comes from one formula: hours of work multiplied by an hourly rate, plus the running costs vendors publish. Two reference rates appear throughout, $50 an hour for an offshore agency team and $150 an hour for a US agency. Swap in the rates from your own quotes and the ranges move with them.

What you pay for in a chatbot build

Most of the jump between bands is integration and testing work. A bot that only answers questions can be wrong in ways a person catches on the next message. A bot that changes a booking or issues a refund needs permission checks, logging and a way to undo mistakes. Those safeguards are where the hours go.

A chatbot quote bundles five kinds of work. Ask any vendor to split its number across these lines, because a single lump sum hides where a cheaper quote cut corners.

• Discovery and content work collects the questions customers ask, cleans the documents the bot will answer from and sets the rules for what it must never say.
• Conversation and interface design covers the widget, the greeting, the handoff to a person and the fallback reply when the bot does not know.
• Build and retrieval is the code that searches your content, sends the right passages to the model and returns an answer with its source attached.
• Integrations connect the bot to your CRM, booking tool, order system or helpdesk, and each one is a small project of its own.
• Testing and launch means a test set of real questions, scored answers, load testing and a soft launch. Budget at least 15 percent of the hours here.

Running costs start on launch day: model usage, a search database, hosting and the hours someone spends reading transcripts and correcting answers. Those lines decide whether a cheap build stays cheap.

Rule-based, LLM or RAG chatbot: which type you need

Rule-based bots follow buttons and decision trees. They cost the least and never invent an answer, but they break the moment a customer types something the tree did not expect. They still fit narrow jobs such as a five-button order-status menu or an appointment confirmation.

LLM chatbots send each message to a large language model such as GPT, Claude or Gemini and let it write the reply. Without access to your own content, the model answers from general knowledge, which is how a bot ends up quoting a price you never charged.

RAG chatbots (retrieval-augmented generation) search your documents first, pass the relevant passages to the model and instruct it to answer only from them. This is the right default for almost every business bot in 2026. Answers stay current the moment you update a document, and each reply can show the page it came from.

AI agents add tools: the model can call your booking system, look up an order or write a CRM record. The guidance for agents is the same as for staff. Give them the narrowest access that does the job, and require a person to approve anything involving money, such as a refund over $50.

Fine-tuning a model on your data is the line item to question. For a support or sales bot, retrieval gives current, citable answers that update the moment a document changes. Fine-tuning changes tone and format, costs training runs and must be redone whenever prices or policies change. Pay for it only after retrieval has been tested and falls short on a specific task.

Should you buy a SaaS AI agent or build a custom chatbot?

Buying is the cheaper start for most small businesses that already run a helpdesk. Intercom's pricing page lists its Fin AI agent from $0.99 per outcome, with no seats required. Seat plans run $19 to $132 a month, and the $19 rate is a new-customer price. Zendesk's Suite pricing starts at $55 per agent a month on annual billing and includes 5 to 10 automated resolutions per agent each month. Extra resolutions cost $1.50 committed or $2.00 pay as you go.

Per-resolution pricing stays simple until volume grows. Say a store has 1,600 conversations a month and the bot resolves 1,000 of them. Fin costs about $990 a month and Zendesk's pay-as-you-go rate about $2,000, before seats. The same 1,600 conversations on a custom bot running gpt-6.1-sol cost about $101 in model fees. The SaaS fee also buys the interface, analytics, hosting and updates, so the fair comparison is that monthly fee against a build cost plus upkeep.

When does building your own chatbot pay off?

Break-even arrives sooner than most owners expect. A store resolving 2,000 conversations a month through Fin pays $23,760 a year in outcome fees. A custom bot costing $30,000 to build and $700 a month to run saves about $15,360 a year against that. It pays back in roughly two years. At 500 resolutions a month, Fin costs $5,940 a year, which is less than the custom bot's running cost alone.

How much does the AI model cost per conversation?

Model providers bill per token, a short chunk of text that can be a whole word or part of one, priced per million tokens. Input tokens are what you send: instructions, retrieved documents and the chat history. Output tokens are what the model writes back. Output costs about five times more per token, but a support bot sends far more than it receives, so input usually dominates the bill.

The table uses one standard support conversation of six exchanges. Each exchange sends about 4,000 input tokens and gets back about 250 output tokens, so the total is 24,000 input and 1,500 output tokens. Prices are the standard rates on OpenAI's API pricing page, Anthropic's Claude pricing page and Google's Gemini API pricing page as of October 2026.

How caching, routing and price changes move the bill

Three details change these numbers in practice. Prompt caching bills repeated instructions at a fraction of the rate. Cached input on gpt-6.1-sol is $0.10 per million tokens against $2.00, and cache hits on Claude Sonnet 5.5 cost 5 percent of its base input price. Anthropic also notes that Claude 4.7 and later models count about 30 percent more tokens for the same text than earlier Claude models.

Published prices also move. Google lists Gemini 4 Argon input doubling from $2 to $4 per million tokens on December 1, 2026. Gemini 3.8 Flash input doubles to $1.50 on January 1, 2027.

The model fee is rarely the budget problem. A mid-tier model at 2,000 conversations a month costs less than one hour of US developer time. The expensive mistake is sending every message to the top model. Route 80 percent of messages to gpt-6-luna and 20 percent to gpt-6.1-sol, and the 2,000-conversation bill falls from $126 to about $30.

How hours and hourly rates set the build price

Build cost is hours times rate, and the rate gap is wider than the hours gap. The Bureau of Labor Statistics puts the median annual wage for US software developers at $135,980 in May 2025, about $65 an hour. That is salary alone, before benefits, office, sales and profit, which is why US agency rates sit far above it. Offshore teams in India and Eastern Europe bill a fraction of US rates for the same scope.

A 400-hour custom RAG chatbot splits into 40 hours of discovery and content cleanup and 60 of conversation and interface design. Retrieval, backend and model work take 160, one or two integrations take 60, and testing, evaluation and launch take 80. A quote far below 250 hours for that scope usually means testing was cut, and the bot will answer confidently from the wrong document.

Fixed-price quotes are safer for a first chatbot because the scope is easy to pin down: these questions, these documents, these two integrations. Time-and-materials billing suits agents whose list of actions grows after launch. Either way, ask for the hour estimate per line. Confirm in writing that you own the code, the prompts and the cloud accounts the bot runs in.

Which features raise chatbot development cost the most?

Each estimate below is for one feature added to a custom RAG chatbot, at the two reference rates.

The CRM line is where most small-business bots earn their keep. A chatbot that captures a lead but leaves it in a spreadsheet saves nobody any time. The guide to what a CRM is and what it should track covers the fields a bot should write to. For low-volume connections, a webhook into an n8n or Zapier workflow can replace a custom integration line in the quote.

Chatbot cost by industry

Industry changes cost through the integrations and controls it demands. The table maps common US small-business uses to the build type and hour range they usually need.

Professional-services firms sit across a wide range. Many only need an intake bot, while answers drawn from case files or client records move a firm into the custom band. If the chatbot ships with a site rebuild, budget both together; the guide on small business website cost covers the website side.

What voice chatbots and regulated industries add

Voice bots are priced per minute as well as per token. OpenAI lists a GPT-Live session at $0.05 per minute, billed per second. Its gpt-realtime-2.1-mini model costs $10 per million audio input tokens and $20 per million audio output tokens. Google lists Gemini 3.8 Live audio at about $0.005 a minute for input and $0.018 a minute for output. A 4-minute call with 2 minutes of bot speech costs about $0.20 on GPT-Live and about $0.06 on Gemini Live, before your phone carrier's charges.

Voice also adds build hours for telephony setup, interruption handling and testing with accents and background noise. Plan 150 to 400 hours on top of the chat build for a phone line that answers, books and transfers calls.

Healthcare, finance and insurance bots cost more because of contracts and controls. Under HIPAA, covered entities and business associates generally must sign written contracts, according to the HHS guidance on business associate contracts. Those contracts limit how protected health information is used and require safeguards. A clinic chatbot that sees patient details needs that agreement with each vendor in the chain, including the model provider and the host.

Add audit logs, access controls and a security review on top, and a regulated build needs 20 to 40 percent more hours. This is general information, not legal advice, so confirm requirements with your compliance counsel.

How much does an AI chatbot cost per month after launch?

The four lines beside model usage add up over a year.

• Search database. Pinecone's pricing page shows a free Starter plan with up to 2 GB of storage and a Standard plan with a $50 monthly minimum. OpenAI's built-in file search charges $0.10 per GB per day after the first free GB, plus $2.50 per 1,000 tool calls.
• Web search, if the bot looks things up. OpenAI and Anthropic each charge $10 per 1,000 searches plus tokens. Google gives 5,000 grounded searches a month free across Gemini 3 and newer models, then charges $14 per 1,000.
• Hosting and monitoring. A small cloud server or serverless functions plus error logging, usually the smallest line on the bill.
• Upkeep hours. Reading a sample of transcripts each week, fixing wrong answers, adding new documents and re-testing after model updates. Plan 6 to 20 hours a month for a business bot.

Model retirement is the cost nobody quotes. Anthropic's pricing page already marks Claude Sonnet 4.5 as deprecated on its cloud platform listings, and OpenAI's flagship lineup has moved to the GPT-6 family. Each forced model switch means re-running your test set and adjusting prompts, which takes 10 to 30 hours. Ask for a build that can swap models through configuration, so a retirement notice becomes a test run instead of a rebuild. Re-run your cost math whenever a vendor publishes new rates.

Worked example: first-year cost for a small service business

Say a 12-person HVAC company in Ohio wants a website chatbot. It should answer questions about service areas, maintenance plans and the diagnostic fee, capture leads into the CRM and offer booking slots. It expects about 1,500 conversations a month, and the scope fits a custom RAG chatbot with two integrations.

Year two falls to the running lines only: about $7,000 at offshore rates or $16,600 at US rates. Switching to Gemini 3.8 Flash would cut model usage from $1,134 to about $425 a year at the 2026 rate. After Google's January 2027 increase, that becomes about $850. Either figure is small next to the upkeep hours, which makes staff time the largest controllable cost after launch.

Set that against the work the bot absorbs. At 4 minutes per chat, 1,500 chats a month is 100 hours of staff time. If the bot resolves half of them, the dispatcher gets back about 50 hours a month for calls that book work. At an assumed $25 hourly wage, that time is worth $15,000 a year.

How to lower the cost of building an AI chatbot

• Launch with one job. Handling the top 30 questions and handing off the rest keeps a build near 250 hours instead of 600. It also gives you real transcripts to plan phase two.
• Clean your content first. Outdated PDFs and conflicting policy pages create wrong answers that take hours to trace. A week of internal cleanup costs less than paid debugging.
• Route by difficulty. Sending 80 percent of messages to a small model cut the 2,000-conversation bill in the example above from $126 to about $30.
• Cache the fixed part of the prompt. Instructions and policies repeat on every request, and cached input costs 5 percent of the normal rate on gpt-6.1-sol and Claude Sonnet 5.5.
• Batch the background work. Nightly transcript summaries and answer scoring can run through batch APIs, which Anthropic and Google price at 50 percent off.
• Use workflow tools for simple integrations, since a webhook into n8n or Zapier can replace a custom CRM connector for low-volume lead capture.

Cutting testing is the false saving. One wrong price quoted to a customer costs more in refunds and trust than the 40 hours of evaluation that would have caught it.

Checklist before you ask for chatbot quotes

1. Pull the last 90 days of emails, chats and call notes, and list the 30 questions customers ask most.
2. Decide what the bot may do: answer only, capture leads, or take actions such as booking.
3. Name every system it must connect to and who holds the admin login for each.
4. Estimate monthly conversations from your current chat, email and phone volume.
5. Set the handoff rule: which topics always go to a person.
6. Flag any regulated data, such as health, payment card or financial details.
7. Ask each vendor for hours per line, the model proposed, a monthly running estimate and the ownership terms.
8. Run the model math from the table above at your own volume before comparing quotes.

Start with step 1 this week, because the question list decides which band your project falls into. For a second opinion on scope, Techno Mantra builds chatbots on OpenAI, Claude and Gemini through its AI development service and can price your project by hours and running cost. You own the code, prompts and accounts. Send your question list and expected volume through the contact page or to info@technomantraa.com, and every inquiry gets a reply within one business day.