Saltar al contenido principal
    Skip to main content

    How much does an AI chatbot cost for your business? A real breakdown for 2026

    Build, model usage, hosting, channel and maintenance: a breakdown with official October 2026 prices and a worked example, step by step.

    By Filip Balik, co-founder · Updated on 17 October 2026

    An AI chatbot has five costs: building it, model usage, hosting, the channel and maintenance. For a custom bot, the big item is the build, which you pay once. Model usage is usually the small one: in the example below, at official October 2026 prices, 10,000 replies a month cost between $5.25 and $105 depending on the model. WhatsApp has its own charges from Meta. Maintenance is not optional.

    The five cost items

    Item When you pay What drives it
    Build Once Integrations, actions the bot performs, channels, human review
    Model usage Monthly, by usage The model chosen, conversation length, volume
    Hosting Monthly A cloud platform or your own server
    Channel Monthly, if the channel charges Website, email or WhatsApp
    Maintenance Monthly or by the hour Model changes, content updates, reviewing conversations

    1. Building it

    This is the heaviest cost at the start. These are our published ranges, ex VAT:

    Type of project Range Fits when
    One automation workflow €1,500 to €6,000 The bot answers and does one thing, such as creating a ticket
    AI MVP €4,000 to €12,000 It answers from your documents and looks things up in one or two systems
    Several workflows €5,000 to €20,000 It reads and writes in several systems, across several channels

    At Quantum Nova we give a fixed price after a free 2 to 5 day assessment, and getting it running usually takes 2 to 4 weeks. We do not subcontract.

    What pushes the build cost up or down:

    1. Integrations. Answering from a document is simple. Checking an order's status in your ERP or booking a slot in a calendar means connecting and testing each system.
    2. Actions. A bot that only gives information is cheaper than one that creates, changes or cancels things. Every action needs controls.
    3. Your knowledge base. If your policies, prices and FAQs are organised and up to date, the work goes down. If they have to be gathered from scratch, it goes up.
    4. Channels. Website, email and WhatsApp each have their own formats and rules.
    5. Handover to a person. Deciding when it escalates, to whom and with what context.
    6. Languages. Each language needs its own testing.
    7. Data requirements. Where it is hosted, what is stored and for how long.

    2. Model usage, with a worked example

    Providers charge per million tokens, and input (what you send to the model) is priced separately from output (what it replies). A token is a piece of a word. With every reply you send the model its instructions, the relevant extracts from your documents and the conversation history.

    Assumption: 2,000 conversations a month, 5 customer messages per conversation. That is 10,000 calls to the model. Each call sends about 4,000 input tokens and gets back about 250 output tokens. Per month: 40 million input tokens and 2.5 million output tokens.

    Prices in US dollars per million tokens, as published by Anthropic, OpenAI and Google on 17 October 2026:

    Provider Model Input Output Monthly cost of the example
    Anthropic Claude Haiku 5.5 $0.10 $0.50 $5.25
    Anthropic Claude Sonnet 5.5 $2 $10 $105
    OpenAI gpt-6-luna $0.10 $0.50 $5.25
    OpenAI gpt-6.1-sol $2 $10 $105
    Google gemini-3.8-flash $0.75 $3.75 $39.38

    The sum is always the same. For Claude Haiku 5.5: 40 × $0.10 = $4 for input, plus 2.5 × $0.50 = $1.25 for output. Total: $5.25.

    Four details that change the number:

    1. Prices change. Google states that gemini-3.8-flash moves to $1.50 input and $7.50 output from 1 January 2027. The same example would then cost $78.75 a month.
    2. Tiers. The Claude Haiku 5.5 price in the table applies to requests of up to 100,000 tokens. Above that it rises to $0.50 and $2.50.
    3. The same text does not take up the same space in every model. Anthropic warns that its models from Claude 4.7 onwards use a tokenizer that produces roughly 30% more tokens for the same text.
    4. Reasoning is paid for too. OpenAI and Google include reasoning or "thinking" tokens in the output price, even if you never see them (OpenAI, Google).

    What brings usage costs down:

    • Prompt caching. If the instructions and the catalogue repeat on every call, they can be read from the cache. On Claude Sonnet 5.5, a cache read costs $0.10 per million, 5% of the base price. If, in the example, 3,000 of the 4,000 tokens always came from the cache, Sonnet 5.5's input cost would drop from $80 to $23 (30 × $0.10 + 10 × $2), not counting cache writes, which cost $2.50 per million on the 5-minute cache.
    • Batch processing. Anthropic gives a 50% discount on its Batch API. It suits tasks that are not urgent, such as summarising the day. Not a live chat.
    • The right-sized model. OpenAI's guide to agents recommends measuring first with the most capable model, then trying smaller models where they give the same result.

    What pushes it up: long conversations, sending too many documents with every question, and requiring everything to be processed in a specific region. Anthropic charges 1.1 times the price for US-only inference. OpenAI adds 10% on its regional processing endpoints for models released from 5 March 2026.

    3. Hosting

    The bot needs somewhere for its logic to live: receiving the message, searching your documents, calling the model, keeping the log. If it is built on n8n Cloud, the plans published as of October 2026 are Starter, at €20 a month billed annually with 2,500 executions, and Pro, at €50 a month billed annually with 10,000 executions. n8n counts one execution each time a workflow runs to completion. If every message in the example triggers the workflow once, that is 10,000 executions: exactly the limit of the Pro plan. According to the same page, n8n Cloud stores data on servers in Frankfurt.

    The other option is to install it on your own server in the EU. The n8n Community edition is free, but you pay for the server and its upkeep. We explain it in what is n8n and when to use it.

    4. The channel

    • Chat on your website. You add it to your site. No third party charges per message, unless you use a paid chat tool.
    • Email. It uses the mailbox you already pay for.
    • WhatsApp. Meta has charged per delivered message since 1 July 2025, and since 1 October 2026 it also charges for service messages (replies inside the customer service window) beyond the first 1,000 per number each month, and for utility templates sent inside that window. We break it down, with the rates for Spain, in WhatsApp Business API pricing, and our WhatsApp Business API cost calculator gives you Meta's monthly bill for your volumes. The rate depends on the message category and the recipient's country code. Check the current tables on Meta's page. Meta also has separate rules for AI providers offering general-purpose assistants, which according to Meta do not change what other businesses pay.

    5. Maintenance

    You cannot leave a chatbot to run on its own. Three reasons:

    1. Models get retired. Anthropic gives at least 60 days' notice, and from the retirement date requests to that model fail. For example, Claude Sonnet 4.5 is retired on 30 November 2026. Switching models means testing again.
    2. Your content changes. Prices, opening hours, terms. If the bot is not updated, it answers with old information.
    3. Someone has to read the conversations. Models make mistakes: Anthropic admits that even the most advanced can generate incorrect text. Reviewing samples every month is how you catch it.

    Always ask, in writing, what maintenance includes and what it costs each month.

    What makes it cheaper and what makes it dearer

    Cheaper Dearer
    Answering only from your documentation Taking actions in several systems
    One channel Website, email and WhatsApp at once
    A small model and caching A large model with long conversations
    Content that is organised and up to date Knowledge scattered across emails and people's heads
    A simple handover to a person Complex escalation rules
    Starting with one use case Wanting to cover everything from day one

    Off the shelf or custom built

    Some products are sold by the month or by usage. Two examples with published prices:

    Off the shelf Custom built
    Getting started Fast and cheap Weeks, plus an initial build
    Monthly cost A subscription or pay per outcome, which grows with volume Model usage and hosting, usually low
    Connection to your systems Whatever the product offers Whatever you need
    Where the data lives Where the provider decides Where you decide
    Control over behaviour Limited to its settings Full, and so is the responsibility

    An off-the-shelf product makes sense if your case is standard and the volume is low. Custom built pays off when the bot has to read from or write to your systems, when pay-per-outcome costs shoot up with volume, or when you need the data in the EU.

    Data and the law: what you cannot skip

    1. GDPR. Conversations contain personal data. The model provider and the platform are data processors: the GDPR requires you to choose processors with sufficient guarantees and to govern the relationship with a contract (Article 28), and if the data leaves the European Economic Area the conditions in Chapter V apply (Article 44). As of October 2026, Anthropic's direct API only offers global or US-only inference, and OpenAI offers processing in Europe with certain retention controls and prior approval. Google states that on the paid tier of the Gemini API content is not used to improve its products; for models with a free tier, that free tier is used.
    2. The EU AI Act. Since 2 August 2026, Article 50 of the Artificial Intelligence Act applies: a system meant to interact with people must be designed so they know they are talking to an AI, unless it is obvious, and they must be told at the latest at the first interaction. On top of that, Article 4, which has applied since 2 February 2025 and was rewritten by Regulation (EU) 2026/1744, requires measures to support AI training for the staff who use it.
    3. Human review. The bot must not promise what it cannot deliver. Anything with consequences, such as refunds or contract changes, should be confirmed by a person. We go into more detail in AI agents for business.

    How to ask for a quote you can compare

    1. List the 20 questions you are asked most and where each answer comes from.
    2. Say what actions the bot should take, if any, and in which system.
    3. Choose one channel to start with.
    4. Work out the volume: conversations per month and messages per conversation.
    5. Ask for the price split into build, model usage, hosting, channel and maintenance.
    6. Ask where the data is processed and stored, and who each processor is.

    If you want us to work it out with your numbers, see our AI solutions service.

    Sources

    We check every article against primary sources before publishing it. If you spot a mistake, write to info@quantumnovaagency.com and we will correct it.

    Terms in this article

    Book your free diagnosis.

    Emailinfo@quantumnovaagency.comWhatsApp+34 678 428 710
    LocationTorrevieja (Alicante) · all of Spain

    What happens next

    1. 1We reply within 24 hours.
    2. 2We have a call to look at your case.
    3. 3We send you a proposal with a fixed price.

    By submitting this form you agree to your data being handled under the Privacy Policy.