Generative AI & LLMs
How Much Does It Cost to Build an AI Chatbot? Build and Running Costs Explained
Written by Tehreem FatimaReviewed by Umaid Asim
Published 10 min read
There is no single price for an AI chatbot, because you pay for two things. Building it is a one-off cost, set mostly by people’s time and scope. Running it is a monthly cost that grows with every message, plus keeping answers correct. In the worked example below, one reply costs between $0.0008 and $0.015, depending on the model.
This article answers how much it costs to make a chatbot for your business by showing how to estimate both parts: what drives the build, how to turn a scope into a number, how platform pricing differs from paying for the model directly, and what a month of running costs looks like, using prices from the providers’ own pricing pages. If you are still deciding how the chatbot should work, start with our guide to AI chatbot development.
The Two Costs of an AI Chatbot
AI chatbot development cost has two parts. The build cost is a one-off cost: scoping the job, preparing the content it answers from, connecting it to your systems, designing the handoff to a person, and testing. The running cost continues every month: fees for each message the model processes, search and hosting fees, and staff time to keep answers accurate.
| Cost | When you pay | What it covers | What makes it go up |
|---|---|---|---|
| Build | Once, plus later changes | Scoping, content preparation, integrations, conversation and handoff design, security, accessibility, testing | More systems to connect, messy documents, strict access rules, actions instead of answers |
| Model usage | Every reply | Text sent to the model and text it writes back, billed per token | Long instructions and history, many retrieved passages, a larger model, long replies |
| Other running costs | Monthly | Search or tool fees, storage, hosting, monitoring | More conversations, hosted search tools |
| Upkeep | Ongoing | Updating content, reviewing conversations, fixing gaps | Policies, prices or products that change often |
Table 1. What you pay for when you build and run an AI chatbot.
Many published price ranges for chatbots combine these four costs without showing the assumptions behind them. A useful estimate keeps them apart, because each one responds to different decisions: the build to scope, model usage to design and traffic, and upkeep to how often your information changes.
What Drives the Cost to Build an AI Chatbot
The biggest driver of build cost is which kind of chatbot the job needs. Our chatbot guide describes three kinds of business chatbot: one that answers a fixed set of questions, one that answers from your documents, and one that also completes tasks in your systems. Each step up adds work.
| Kind of chatbot | Main build work | Build cost rises when |
|---|---|---|
| FAQ chatbot | Writing and checking a fixed set of answers, setting it up on a platform | The list of questions keeps growing, or several teams must approve the answers |
| Knowledge chatbot | Preparing documents, setting up search, showing sources, checking who may see what | Documents are outdated, scattered or contradictory, or different staff may see different content |
| Action chatbot | Connecting to systems such as order or ticketing tools, approval rules, handoff, records of every action | Each new system, each action that changes data, and each approval step |
Table 2. How the kind of chatbot shapes the build.
Within each kind, a few more choices move the number: how much testing the use case needs before launch, how many languages and channels it runs in, and how strict the security and privacy requirements are. A chatbot that answers staff questions about a holiday policy needs less testing than one that tells customers what their insurance covers.
How to Estimate the Build Cost
To estimate the cost to build or develop an AI chatbot, list the work packages, estimate the days each one needs, and multiply by the day rate of the people doing the work. A ballpark figure without this breakdown cannot show which part of your scope is expensive, or what to cut to fit a budget.
| Work package | Example days |
|---|---|
| Scope the job and decide what counts as a resolved conversation | 5 |
| Gather, clean and organise the documents | 10 |
| Build search, answers and links to sources | 12 |
| Connect one system, such as the help desk for handoffs | 6 |
| Design the conversation and the handoff to a person | 4 |
| Security, privacy and accessibility checks | 5 |
| Test with real questions and fix what fails | 8 |
| Launch to a small group and monitor | 5 |
| Total | 55 |
Table 3. An illustrative build estimate for a knowledge chatbot with one connected system. The days are assumptions to show the method, not a typical project or a SensViz quote.
To turn days into money, use the rates you will actually pay. As a public reference point, the U.S. Bureau of Labor Statistics puts the 2025 median pay for software developers, quality assurance analysts and testers at $64.44 an hourSource [1]. At that rate, the 55 days in Table 3, at eight hours a day, come to about $28,350 in wages alone. Agency and freelance rates differ from this figure, because they also cover management, overheads and margin, and an in-house team has benefits and other employment costs on top of pay.
The same method works at any size. A small FAQ chatbot on a platform might need a handful of the packages above. An action chatbot connected to several systems repeats the integration, security and testing rows for each one.
Platform or Custom Build: Two Cost Shapes
Chatbot platforms and custom builds put the money in different places. A platform usually reduces the build, because much of the chat window, hosting and handoff already exists, and charges for usage through its own plans. Microsoft Copilot Studio, for example, sells “Copilot Credit capacity packs of 25,000 Copilot Credits each, priced at $200.00/pack/month”Source [2]. Credits are used “whenever Copilot or an agent performs a task or generates a response”, and the number billed varies with what the agent doesSource [2], so measure credit use in a pilot before you budget a full year.
A custom build usually costs more to set up, but you pay the model provider directly for what you use and keep control of the logic, data and integrations. Some providers also offer a free tier for experiments. Google describes its Gemini API free tier as “For developers and small projects getting started with the Gemini API”Source [3], which is useful for testing an idea before you commit to a build.
What It Costs to Run an AI Chatbot: A Worked Example
Model providers charge per token, a small piece of text. For English text, OpenAI’s rule of thumb is that “100 tokens are approximately 75 words”Source [4]; other languages can use more or fewer tokens per word. Every reply is billed for the text sent to the model (input) and the text it writes (output). The model does not keep the conversation itself. OpenAI’s documentation treats each request as separateSource [5], so the chatbot sends its instructions, any retrieved passages and the conversation so far with every message.
The example below uses a knowledge chatbot with these assumptions:
| What is sent or written | Tokens per reply |
|---|---|
| Instructions and rules (the same on every reply) | 2,000 |
| Passages retrieved from your documents | 3,000 |
| Conversation so far (average across a conversation) | 1,000 |
| The customer’s question | 100 |
| Input total | 6,100 |
| The chatbot’s reply (output, about 225 words) | 300 |
Table 4. Token assumptions for the worked example. Volume: 5,000 conversations a month with four replies each, or 20,000 replies.
What One Reply Costs
Both providers in this example offer two similar price levels. At the time of writing, Anthropic’s Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, and Claude Haiku 5.5 costs $0.10 and $0.50 for prompts up to 100,000 tokensSource [6]. OpenAI lists the same two price points for its gpt-6.1-sol and gpt-6-luna modelsSource [7].
At the higher price, one reply costs $0.0122 for input plus $0.003 for output, or about $0.015. Across 20,000 replies, that is about $304 a month. At the lower price, one reply costs about $0.0008, or about $15 a month.
Output tokens cost five times more than input tokens at both price levels, yet input makes up 80% of the bill in this example. In this example, the chatbot sends about 20 tokens for every token it writes, so the instructions, retrieved passages and history are where most of the money goes. In this example, halving the retrieved passages saves as much as the entire reply costs.
What Caching Changes
Text that repeats on every reply, such as the instructions, can be cached and billed at a much lower rate. For Claude Sonnet 5.5, a cache read costs 0.05 times the base input priceSource [6], and OpenAI lists cached input for gpt-6.1-sol at $0.10 per million tokensSource [7]. If the 2,000 instruction tokens are read from the cache on most replies, the higher-priced reply falls to about $0.0114, or $228 a month, a 25% saving. At the lower price, the monthly bill falls from about $15 to about $12.
Caching has a cost of its own: saving text to the cache is billed above the normal input price. Anthropic charges 1.25 times the base input price to write a cache that stays valid for five minutesSource [6], and OpenAI lists cache writes for gpt-6.1-sol at $2.50 per million tokensSource [7]. A chatbot with long quiet periods between messages rebuilds its cache more often and saves less.
Costs Beyond the Model
The model bill is not the whole running cost, and at low prices it may not be the largest part.
- Search. If the chatbot uses a hosted search tool to find passages, that tool has its own fees. OpenAI’s file search costs $2.50 per 1,000 tool calls, plus $0.10 per GB per day for storage after the first GBSource [7]. With one search per reply, the example’s 20,000 replies add $50 a month, more than the lower-priced model itself.
- Hosting and monitoring. The chat window, logs, dashboards and alerts run somewhere, and their cost depends on your setup.
- Upkeep. Someone has to update documents when policies change, review conversations the chatbot handled badly, and fix the gaps. This is people’s time, every month. Fossilite’s guide to software maintenance cost after launch breaks down what to budget for once software is live.
Prices also change, sometimes on a published schedule. Google lists Gemini 3.8 Flash input at “$0.75 through December 31, 2026” and “$1.50 starting January 1, 2027”Source [3], which doubles that part of the bill overnight. Check the pricing pages again before you fix a budget.
Why Action Chatbots Cost More to Run
An action chatbot does more work per conversation. To check an order or update a ticket, the model decides which tool to call, reads the result, and then writes the reply, and each of those steps sends the context to the model again. Anthropic reports from its own systems that “agents typically use about 4× more tokens than chat interactions”Source [8].
For action chatbots, budget per completed task rather than per reply, and measure it in a pilot. A task that saves a member of staff ten minutes can justify several cents of model cost; the same cost spent on a question your website already answers may not.
How to Build Your Monthly Budget
A simple formula covers most cases:
Monthly running cost = conversations × replies per conversation × cost per reply + search and tool fees + hosting and monitoring + upkeep time
- Run a pilot with real users. Record the actual number of conversations, replies per conversation, and tokens per reply. Assumptions like the ones in Table 4 are a starting point, not a forecast.
- Price it from the current pricing page of the provider you will use, not from an old estimate.
- Add the costs beyond the model, including the hours someone will spend on upkeep.
- Compare the total with what handling those conversations costs today.
If the model bill is the problem, the levers are a smaller model for simple questions, fewer retrieved passages, caching and shorter history. Our chatbot guide explains what drives AI chatbot development cost and how each lever works. Before moving questions to a cheaper model, check that the answers still hold up; our post on LLM testing covers how.
How SensViz Can Help
SensViz builds AI chatbots as part of our generative AI development work, from internal knowledge assistants to customer support assistants that prepare replies for agents to review. Discovery confirms the use case, data readiness and integrations before we give a delivery estimate or quote, and our pricing page explains what shapes a project’s cost. If a platform will do the job, we will say so.
Frequently Asked Questions
These answer three questions that often come up once a chatbot budget is on the table.
Do You Pay for an AI Chatbot When Nobody Is Using It?
With pay-per-token pricing, model fees follow usage, so a quiet month costs little. Platforms vary: Copilot Studio sells monthly capacity packs and also offers a pay-as-you-go meterSource [2]. Stored files can still be billed dailySource [7], and hosting, monitoring and content upkeep continue whether or not people use the chatbot.
How Long Does It Take to Build an AI Chatbot?
The work-package estimate that gives you the build cost also gives you the timeline. The 55 days in Table 3 would take one person about 11 working weeks, and less if several people work in parallel. Waiting for documents, access to systems or approvals can add calendar time that a count of working days does not show.
Can You Cap What an AI Chatbot Spends Each Month?
Yes, at the model provider. Anthropic, for example, lets an organization “set your own spend limit below your tier’s cap to control costs”Source [9]. Once usage reaches that limit, requests are refused until access resumes or the limit is raised, so the chatbot needs a fallback, such as handing the conversation to a person.
Sources
- [1] U.S. Bureau of Labor Statistics, “Software Developers, Quality Assurance Analysts, and Testers (opens in a new tab),” Occupational Outlook Handbook, accessed 16 October 2026.
- [2] Microsoft, “Microsoft Copilot Studio pricing (opens in a new tab),” accessed 16 October 2026.
- [3] Google, “Gemini Developer API pricing (opens in a new tab),” Gemini API documentation, accessed 16 October 2026.
- [4] OpenAI Help Center, “Understanding and counting tokens (opens in a new tab),” accessed 16 October 2026.
- [5] OpenAI, “Conversation state (opens in a new tab),” OpenAI API documentation, accessed 16 October 2026.
- [6] Anthropic, “Pricing (opens in a new tab),” Claude Platform documentation, accessed 16 October 2026.
- [7] OpenAI, “Pricing (opens in a new tab),” OpenAI API documentation, accessed 16 October 2026.
- [8] Anthropic, “How we built our multi-agent research system (opens in a new tab),” 13 June 2025.
- [9] Anthropic, “Rate limits (opens in a new tab),” Claude Platform documentation, accessed 16 October 2026.
