How Much Does It Cost to Build an AI Chatbot in 2026?

Most answers to "how much does an AI chatbot cost?" give a price range with no working. This guide shows the working instead. We split the cost into three parts: building it, running it and keeping it accurate.

We define four common chatbot types, from a scripted FAQ bot to an agent that takes actions. For each one, we give a planning estimate of hours and multiply it by published 2026 rates. Model prices come from the providers' own pricing pages, checked on 12 October 2026.

We are Webify.AI, and we build chatbots and AI agents, so we have an interest here. That is why every number below is either sourced or clearly labelled as our own estimate.

How much does it cost to build an AI chatbot in 2026?

A first version takes roughly 40 to 1,600 hours of work, by our planning estimates, depending on what the chatbot does. At published 2026 rates, that is about $1,000 to $384,000. Simple FAQ bots sit at the bottom and action-taking agents at the top. Monthly running and maintenance costs come on top.

Each range multiplies our planning hours by the low and high published rates. Later sections show every calculation, with GBP figures for UK readers.

The low end of each range assumes clean content, one channel and few integrations. The high end assumes messy data, several systems and a stricter accuracy bar. Our rates sit below the Asia band for type 1, inside it for type 2 and above it for types 3 and 4. We are not the cheapest option on every line, and the table shows that.

How do you work out what a chatbot will cost?

Use one formula: build hours multiplied by hourly rate, plus monthly running costs, plus maintenance. In our examples, the build is the biggest single number. Running costs depend on traffic and model choice. Maintenance covers content updates, testing and fixes after launch, and it never falls to zero.

• Build cost: hours of discovery, design, development, testing and launch, times each person's rate.
• Running costs: model usage billed per token, plus hosting, search and monitoring tools.
• Maintenance: monthly people time to review conversations, update content and keep up with model changes.

We use three published rate levels. Accelerance's survey of 60 software development partners puts 2026 senior developer rates at $31-41 an hour in Asia (Accelerance). Its European band is $64-76, with the strongest downward pressure in Central and Eastern Europe.

KORE1, a US staffing firm, lists 2026 agency bill rates of $72-115 an hour for software developers (KORE1). Senior software engineers bill at $108-160, and AI/ML engineers at $140-240.

For UK readers, IT Jobs Watch puts the median contract day rate for software developers at £525 (IT Jobs Watch). For AI engineers it is £600 a day (IT Jobs Watch). Both cover the six months to 12 October 2026. We assume an 8-hour day.

Two caveats matter. Accelerance's figures cover senior developers in general, not AI specialists, so treat them as a floor for AI work. KORE1 states openly that its own margin is included in its bill rates.

Where a chatbot type needs two roles, we assume each person does half the hours and we average the two rates. The hours are our planning estimates for a typical first version. They are not quotes, and your scope will move them.

What are the main types of AI chatbot?

Most business chatbots fall into four types. They are a scripted FAQ bot, a knowledge-base assistant using RAG, an assistant connected to business systems and an agent that takes actions. Each step up adds data work, integration work and testing. So each type costs more to build and more to run.

RAG stands for retrieval-augmented generation. Anthropic describes it as combining information retrieval with language model generation, to ground answers in evidence (Anthropic). In plain terms, the bot looks up relevant passages first, then writes its answer from them.

How much does a scripted FAQ chatbot cost?

A scripted FAQ bot is the cheapest type. Our planning estimate is 40-120 hours for a first version. That is $1,000-$3,000 at Webify.AI's mid-level developer rate, $1,240-$4,920 at Asia senior rates and $2,880-$13,800 at US staffing-agency rates. Running costs stay low because most replies are pre-written.

This bot follows a script or a decision tree. It may use a small language model only to match each question to the right approved answer. It suits opening hours, delivery times, returns and other questions with one correct answer.

Typical team: one developer, plus someone on your side who writes and approves the answers. The work covers conversation design, the chat widget, a handoff to a human and basic reporting.

This type is enough when most questions have one approved answer that rarely changes. When customers ask open questions about long documents, you need type 2.

How much does a RAG chatbot cost?

A knowledge-base assistant using RAG answers questions from your own documents. Our planning estimate is 240-480 hours. That is $8,400-$16,800 at Webify.AI's blended rate, $7,440-$19,680 at Asia senior rates and $25,440-$85,200 at US staffing-agency rates. Monthly model costs depend mostly on traffic and model choice.

This assistant answers from your own content: help centre articles, policies, manuals or product data. For each question, it retrieves relevant passages and writes an answer from them, ideally with links to the source. It should say when it does not know and hand over to a person.

Typical team: an AI/LLM engineer for retrieval, prompts and evaluation, and a developer for the document pipeline, chat interface and deployment. We assume a 50/50 split of hours.

Blended Webify.AI rate: ($45 + $25) ÷ 2 = $35 an hour, or (£36 + £20) ÷ 2 = £28. Blended US agency rate: KORE1's AI/ML band ($140-240) averaged with its software developer band ($72-115) gives $106-177.50. Blended UK day rate: (£600 + £525) ÷ 2 = £562.50.

What moves the number: messy documents, many file formats, complex access rules and extra languages. Scanned PDFs and spreadsheets take longer than clean web pages.

Our RAG development page explains how we approach this work.

How much does an AI assistant connected to business systems cost?

An assistant that reads live data from your CRM, order system or helpdesk costs more, because every integration needs building, securing and testing. Our planning estimate is 480-960 hours. That is $22,800-$45,600 at Webify.AI's blended rate, $14,880-$39,360 at Asia senior rates and $59,520-$192,000 at US agency rates.

This assistant answers questions about one customer or case, such as "Where is my order?" or "What is left on my contract?". It needs sign-in, permission checks and careful logging. It must never show one customer's data to another.

Typical team: a senior AI engineer for the model, retrieval and guardrails, and a senior developer for the integrations and authentication.

Blended Webify.AI rate: ($60 + $35) ÷ 2 = $47.50 an hour, or (£48 + £28) ÷ 2 = £38. Blended US agency rate: KORE1's AI/ML band ($140-240) averaged with its senior software engineer band ($108-160) gives $124-200. The UK blend stays at £562.50 a day.

What moves the number: how many systems are involved, whether they have usable APIs and how strict your access rules are.

To add engineers to your own team instead, see hire LLM engineers.

How much does an AI agent that takes actions cost?

An agent that changes things in your systems is the most expensive type to build and test. Our planning estimate is 800-1,600 hours for a first version. That is $42,000-$84,000 at Webify.AI's blended rate, $24,800-$65,600 at Asia senior rates and $112,000-$384,000 at US staffing-agency rates.

This agent does more than answer. It might issue a refund within a limit, book an appointment, update a record or raise a ticket. Each action needs clear limits, an approval step for risky cases and an audit log.

OWASP lists "Excessive Agency" among its top 10 risks for LLM applications (OWASP). In practice, that means giving the agent only the permissions it needs. It also means testing what happens when the agent is tricked or simply wrong.

Typical team: a senior AI engineer and an AI/LLM engineer, sharing the integration work. Blended Webify.AI rate: ($60 + $45) ÷ 2 = $52.50 an hour, or (£48 + £36) ÷ 2 = £42. For the US, we use KORE1's AI/ML engineer band of $140-240. For the UK, we use the £600 AI engineer median.

Agents also cost more to run. One user request can trigger several model calls and several tool calls. See our AI agents page for the kinds of agents we build.

How much does an AI chatbot cost to run each month?

Running costs are mostly model usage, billed per token, plus hosting, search and maintenance time. In our worked example, a knowledge-base assistant handling 5,000 conversations a month costs about $13 in model fees on a small model. On a mid-priced model it costs $260. Traffic, context length and model choice drive the number.

How does token pricing work?

Language models read and write text in tokens. A token can be a word, part of a word or a single character. Anthropic says a token is about 3.5 English characters on its earlier models (Anthropic). Its newer tokenizer produces about 30% more tokens for the same text.

Providers charge separately for input tokens, which is everything you send, and output tokens, which is everything the model writes. Output costs more per token. Many models also reason before answering, and OpenAI states that reasoning tokens are billed as output tokens (OpenAI).

Every call resends the context the model needs. That includes your instructions, the retrieved passages and the conversation so far. So the input bill grows with each turn of a long conversation.

What do the models cost per token?

These are list prices per million tokens, checked on 12 October 2026 on the providers' official pages. Both price in US dollars, and prices change often.

Sources: Anthropic pricing and OpenAI pricing. The OpenAI figures are its standard, short-context prices. Both providers also charge for writing content into the cache, at a higher rate than normal input.

A worked example

Assume a knowledge-base assistant with 5,000 conversations a month and four questions per conversation. That is 20,000 model calls. Assume each call sends 4,000 input tokens and returns 500 output tokens, including any reasoning. These are illustrative assumptions, not measurements.

Monthly volume: 20,000 × 4,000 = 80 million input tokens, and 20,000 × 500 = 10 million output tokens.

The most expensive row costs 100 times the cheapest. Doubling conversations doubles the bill, and so does doubling the context sent with each call. That is why model choice and context length matter so much.

Caching can cut this further. Suppose 1,500 of the 4,000 input tokens are fixed instructions. On Sonnet 5.5, those 30 million tokens would cost 30 × $0.10 = $3 as cache hits. Without caching, they cost 30 × $2 = $60. Cache-write charges come on top.

What else do you pay for each month?

Model tokens are not the whole bill. You also pay for retrieval, tools, hosting and people.

• Embeddings: turning documents into searchable vectors is cheap. OpenAI lists text-embedding-3-small at $0.02 per million tokens, so embedding 2 million tokens of documents costs $0.04 (OpenAI).
• Managed search: OpenAI's hosted file search charges $0.10 per GB per day after 1 GB free, plus $2.50 per 1,000 tool calls. At 20,000 searches a month, the call fees alone are $50.
• Tool use: tool definitions and tool results count as input tokens. Anthropic also adds a 286-token tool-use system prompt to each call with tools on Claude Opus 5.5 and Sonnet 5.5 (Anthropic).
• Web search: if the bot searches the web through Anthropic's tool, Anthropic charges $10 per 1,000 searches on top of tokens.
• Data residency: Anthropic charges 1.1 times standard prices for US-only inference on Claude 4.6 and later models.
• Hosting and logs: the server, database and log storage. Get these quoted for your traffic and how long you keep logs.

What does maintenance cost?

Maintenance is people time, so it follows the same hours-times-rate method. Our planning estimate for the first months after launch is 8-16 hours a month for a type 1 or type 2 bot. For a type 3 assistant or type 4 agent, we plan 20-40 hours a month.

At Webify.AI's blended rates, a RAG assistant comes to 8 × $35 = $280 to 16 × $35 = $560 a month. An agent comes to 20 × $52.50 = $1,050 to 40 × $52.50 = $2,100 a month. These are planning estimates, not fixed fees.

What hidden costs do chatbot budgets miss?

Most hidden costs are people costs, not model costs. Preparing data, building a test set, reviewing security, checking data protection, monitoring conversations and handling change requests all take time. None of them shows up in a token calculator. Budget for them from the start, or the launch date slips.

• Data preparation: removing out-of-date documents, fixing contradictions and deciding who may see what. A bot repeats whatever you give it.
• Evaluation: a set of real questions with agreed answers, rerun after every change. OpenAI notes that LLM output is non-deterministic and that behaviour changes between model versions (OpenAI).
• Security review: testing for the OWASP Top 10 for LLM applications, including prompt injection, sensitive information disclosure and system prompt leakage (OWASP).
• Data protection: in the UK, the ICO says most AI uses of personal data legally require a data protection impact assessment (ICO).
• Monitoring: someone must read samples of conversations, watch costs and correct wrong answers. Logs alone fix nothing.
• Change requests: new products, policies, channels and integrations. Each one needs building and retesting.
• Model changes: providers retire older models. Anthropic's pricing page marks several earlier Claude models as retired, so plan time for migration and retesting (Anthropic).

Our effort ranges above include basic testing and a launch review. They do not include a formal external security test, a data protection impact assessment or a large clean-up of your content.

Wrong answers carry real costs. In February 2024, a Canadian tribunal ordered Air Canada to pay a passenger C$812 over its chatbot's wrong advice on bereavement fares (CBC News). The tribunal rejected the airline's argument that the chatbot was responsible for its own actions.

Our security page explains how we approach these reviews on client projects.

How can you cut the cost of an AI chatbot?

Start with the simplest bot that solves the problem, use retrieval before fine-tuning, and pick the smallest model that passes your tests. Then cut tokens: cache fixed instructions, send fewer but better passages and cap usage per user. These steps lower both the build bill and the monthly bill.

• Start small: launch a type 1 or type 2 bot on your most common questions. Add integrations once you see real usage.
• Use RAG before fine-tuning: RAG puts the right content into each prompt, which OpenAI's guide recommends. OpenAI says prompt engineering "may be all you need" (OpenAI). It is also winding down its fine-tuning platform, which new users can no longer access. Anthropic says the Claude API does not currently offer fine-tuning (Anthropic).
• Use smaller models where they pass: at list prices, Claude Haiku 5.5 input costs $0.10 per million tokens against $2 for Sonnet 5.5. That is a 20-fold difference, so test the small model first.
• Cache fixed content: Anthropic prices cache hits at 0.1 times the base input price on most models (Anthropic). On Opus 5.5 and Sonnet 5.5, it is 0.05 times.
• Know the caching rules: caching needs a repeated prompt prefix above a minimum length. On OpenAI's GPT-5.6 and later models, that minimum is 1,024 tokens (OpenAI).
• Trim context: retrieve fewer, better passages and summarise long conversation histories instead of resending them in full.
• Batch offline work: Anthropic's Batch API gives a 50% discount on input and output tokens for work that can wait. OpenAI's batch prices for its flagship models are also half its standard prices.
• Cap usage: set per-user limits so a script or an attacker cannot run up your bill. OWASP calls this risk "Unbounded Consumption".
• Use private models only when needed: open-weight models on your own servers can suit strict data rules or very high volume. But you pay for those servers whether or not anyone is chatting.

If your data cannot leave your network, see our offline LLM development service for private and on-premise models.

Should you build or buy a chatbot?

Buy a SaaS chatbot if your needs match what a platform already does and your volume is modest. Build a custom one if you need deep integration, control over where data goes, or lower cost per conversation at high volume. You can also start on a platform and move later.

SaaS platforms usually charge per outcome or per credit, on top of seats. Intercom lists its Fin AI Agent at $0.99 per outcome, with helpdesk seats from $29 per seat a month, paid annually (Intercom). It counts an outcome when a customer confirms resolution, asks for no more help, or Fin completes a workflow.

Microsoft sells Copilot Studio in packs of 25,000 Copilot Credits for $200 a month (Microsoft). It says each agent response or action uses a varying number of credits, depending on the usage.

Compare like with like. In our worked example, 5,000 conversations cost $260 a month in tokens on a mid-priced model. If a platform resolved 2,500 of them at $0.99, the fee would be 2,500 × $0.99 = $2,475.

That gap does not make custom cheaper by itself. The platform fee includes software, hosting, updates and support that you would otherwise pay people to build and maintain.

How does Webify.AI price chatbot development?

We price by time, not by a fixed chatbot price. You hire our developers and AI engineers by the hour or by the month. We agree the scope and an effort estimate with you before work starts. Our hourly rates run from $25 for a mid-level developer to $60 for a senior AI engineer.

• Hourly rates (USD): $25 mid-level developer, $35 senior developer, $45 AI/LLM engineer and $60 senior AI engineer.
• Hourly rates (GBP): £20, £28, £36 and £48 for the same four roles.
• Monthly (160 hours): $3,400, $4,750, $6,100 and $8,150 for the same four roles. A senior AI engineer then works out at about $50.94 an hour.
• Trial: two weeks, with no charge if you stop within the first five working days.
• Ownership: you own all the code we write for you.
• Platforms: we deploy on IBM watsonx and on open-weight models, and we build offline and on-premise LLM systems.

We are an AI and software company headquartered in Ahmedabad, India, and an IBM Partner. Some honest limits: we do not hold security certifications. If your procurement process requires one, tell us at the start.

The hour ranges here are planning estimates. Before work begins, we give you a scoped estimate using the same hours-times-rate method.

Full rates are on our pricing page. To start a chatbot or agent project, see hire AI agent developers. Country pages cover AI engineers for UK companies and AI engineers for US companies.

What should you prepare before asking for a quote?

Bring the questions or tasks the bot should handle, samples of the documents it will use and the systems it must connect to. Add your expected monthly conversations and your rules on accuracy and data. With these, any supplier can give you an hours estimate that you can check against the tables above.

• Top questions: 50-100 real customer questions, with the correct answers.
• Content: where your documents live, how current they are and who owns them.
• Systems: the CRM, helpdesk or order system involved, and whether each one has an API.
• Volume: conversations per month now, and what you expect in a year.
• Rules: where data must stay, who may see what and what the bot must never do.
• Success measure: for example, the share of test questions answered correctly.

When you have these, talk to our AI agent developers. We will send back an hours estimate, a team and a monthly running-cost estimate, with the arithmetic shown.

Sources

• Accelerance: 2026 outsourcing rate trends in Asia, Europe and Latin America
• KORE1: Tech contractor hourly rates 2026
• IT Jobs Watch: Software Developer contract rates, UK
• IT Jobs Watch: AI Engineer contract rates, UK
• Anthropic: Claude API pricing
• Anthropic: Glossary
• OpenAI: API pricing
• OpenAI: Reasoning models
• OpenAI: Model optimization
• OpenAI: Prompt caching
• OWASP: Top 10 for LLM Applications 2025
• ICO: Accountability and governance implications of AI
• CBC News: Air Canada found liable for chatbot's bad advice
• Intercom: Pricing
• Microsoft: Copilot Studio pricing