AI automation
How much does an AI chatbot for a website cost in 2026, and when does it need a human?
Published 2026 price guides put a scripted website bot at about $2,000–$15,000 and an LLM chatbot that answers from your content at $24,000–$60,000 or more, plus $400–$6,000 a month to run. VITON13 launches a one-channel chatbot with human handoff from $73. Refunds, complaints, legal questions and anyone who asks for a person go to a human.
What does each kind of website chatbot cost to build?
Quotes land in four bands, set less by the vendor than by what the bot may do. These are ranges vendors publish themselves: asking prices, not audited averages.
| Type | What it does | Published build price | Build time | Source |
|---|---|---|---|---|
| Scripted FAQ bot | Buttons and prepared replies; no language model | $4,500–$12,000 (RaftLabs); $2,000–$10,000+ for FAQ and basic support (Cleveroad) | 3–5 weeks | RaftLabs, Jun 2026; Cleveroad, Jan 2026 |
| Lead-qualification bot | Asks set questions, routes the lead to sales | $8,000–$25,000+ | Not stated | Cleveroad, Jan 2026 |
| LLM chatbot grounded in your content | Reads free text and answers from your documents | $24,000–$60,000 (RaftLabs); $35,000–$120,000+ for generative AI bots (Cleveroad) | 8–14 weeks | RaftLabs, Jun 2026; Cleveroad, Jan 2026 |
| Enterprise AI agent | Acts inside your systems: bookings, orders, accounts | $72,000–$210,000 | 16–28 weeks | RaftLabs, Jun 2026 |
| VITON13 chatbot launch | One channel, answers you approve, request capture, handoff to a person | From $73 AI-assisted or $93 human-led | 5–8 working days | VITON13 service page |
Per RaftLabs, bots built on a language model cost four to eight times more than scripted ones: prompts, retrieval and output checks all need building and testing.
Why does a VITON13 chatbot start so far below those ranges?
Those bands price custom software: a bespoke interface, a retrieval pipeline, integrations, weeks of a team's time. Our launch is narrower: an existing chatbot platform picked in the written scope, one channel, a knowledge base you sign off. AI drafts answers from your documents; a specialist reads each one before launch.
| Your situation | What fits | Starting price | Window |
|---|---|---|---|
| Up to 30 recurring questions or 20 pages of documents, one channel | AI chatbot with human handoff | $73 AI-assisted or $93 human-led | 5–8 working days |
| Hundreds of documents; answers must cite their source | RAG knowledge assistant | $113 AI-assisted or $173 human-led | 10–18 working days |
| The bot must act: check an order, book a slot | AI agents for business | $73 AI-assisted or $93 human-led | 5–8 working days |
| Unsure a chatbot is the right first AI project | AI Use-Case Session: 90 minutes, one process, a written plan | From $53 | 1 working day |
What does a chatbot cost to run each month?
RaftLabs puts running a chatbot at $400–$6,000 a month, with model API fees the largest, least predictable line. A small site's bill has four lines, each paid to a different provider.
| Line | What moves it | Published price | Source |
|---|---|---|---|
| Model tokens | Conversation length, model, volume | Claude Haiku 4.5: $1 per million input tokens, $5 per million output; about $37 per 10,000 support conversations of roughly 3,700 tokens (Anthropic's example) | Anthropic pricing |
| Chat platform | Seats, AI usage, plan | Intercom: Fin AI Agent $0.99 per outcome; seats from $29 a month billed yearly | Intercom pricing |
| WhatsApp channel | Template messages you send | Replies inside an open customer service window are free; marketing templates are always charged | Meta, WhatsApp Business Platform pricing |
| Upkeep | New questions, price changes | 5–10 hours a month reviewing edge cases and updating answers | RaftLabs, Jun 2026 |
How a provider charges matters as much as how much. On list prices, 1,000 conversations of Anthropic's example length use about $3.70 of Haiku tokens; 1,000 Fin outcomes cost $990. The platform fee buys an inbox, seats and reports a bare model lacks, so compare whole bills.
Chatbot, live chat or both: which costs less per conversation?
A bot is cheap only if the conversation ends well. In a Gartner survey of 3,566 customers (February–March 2026), 87% called access to a human agent essential when a company uses generative AI for service, and only 27% would try a chatbot again after a bad experience. Their complaints: the bot misread the problem, gave generic text or made a person harder to reach.
| Setup | What you pay for | Where it wins | Where it fails |
|---|---|---|---|
| Bot only | Tokens or a per-outcome fee | Instant answers at 3 a.m., no queue | Dead ends; a failed visitor rarely returns |
| People only | Shift hours, chats or not (US median customer service wage: $21.53 an hour, May 2025), plus a chat seat | Judgment, exceptions, tone | Office hours and queues; the same twenty questions all day |
| Bot first, person on a rule | The bot's price on routine chats, a person's time on the rest | Routine chats close fast; hard ones reach someone who decides | Only as good as its handoff rule |
For most small sites, the third setup is the one worth buying: the question is which chats reach a person, and how fast.
Which conversations must go to a person?
Write the handoff rule before anyone builds the bot. These triggers belong in nearly every rule:
- The visitor asks for a person. Hand over the first time; making someone ask twice loses them.
- Money in dispute: refunds, cancellations, chargebacks, a complaint about an order.
- Anything legal, medical or financial, and any decision that changes what a customer pays or receives.
- No approved answer: the question falls outside the knowledge base, or the bot has already missed once.
- A bigger deal than usual: a quote above a threshold you set, a group booking, a company enquiry.
- Anger, distress or anything touching safety.
The handoff carries the transcript, the contact and the reason, so nobody makes the visitor start over; out of hours, the bot says when a person will reply. RaftLabs reckons proper escalation adds 20–30% to development time, which is why cheap quotes drop it. Ask where those hours sit in yours.
What must the bot tell visitors, and what does the law ask?
In the EU, Article 50(1) of the AI Act requires systems that interact directly with people to let them know they are dealing with AI, unless that is obvious; Article 50(5) wants the notice clear and distinguishable by the first interaction. The AI Act Service Desk dates these rules from 2 August 2026.
- Disclosure: the opening message says it is an automated assistant and how to reach a person, on every channel, wherever the visitor is.
- Personal data: link your privacy notice before the bot asks for a name or number, and agree how long transcripts are kept.
- Scope: answers come only from approved content; beyond it, the bot says so and offers a person.
- Promises: no discounts, delivery dates or refunds your policy does not state. Visitors read the bot's reply as your company's.
- Training: where possible, stop the platform training its models on your conversations.
What should a chatbot quote spell out before you sign?
Two quotes for “a chatbot” can differ tenfold because they describe different jobs. Any line a vendor will not put in writing is a cost you meet later.
- Knowledge: which documents, how many, who keeps them current.
- Channels, languages and flows, each named.
- Captured fields and where they land.
- Handoff: triggers, channel, hours, the out-of-hours message.
- Testing: how many scripted conversations, including attempts to drag the bot off topic.
- Running fees by provider, whose name the accounts are in, and where transcripts are kept and for how long.
- After launch: the fix period, the first report, what counts as new work.
When is an AI chatbot not worth buying yet?
- Few chats arrive. At a handful a week, a WhatsApp button or a contact form costs less and loses nothing.
- Your answers are not written down. Write the FAQ first; a bot cannot answer from nothing.
- Prices or stock change daily and live in someone's head, so the bot would quote stale figures.
- The questions need professional judgment. We do not build bots that give medical, legal or financial advice.
- Nobody can take the handoff. A bot that offers a person and never produces one is worse than no bot.
- You want a promise of more leads. We report what the bot did; nobody can honestly promise what it will earn.
AI-assisted or human-led: what changes for a chatbot?
Same package, same window of 5–8 working days; what differs is the specialist time spent on the conversation itself.
| AI-assisted | Human-led | |
|---|---|---|
| Starting price | $73 | $93 |
| Answers | Drafted by AI from your documents, checked by a specialist | Conversation and tone designed by a specialist |
| Tests before launch | 20 scripted conversations | 40 scripted conversations |
| After launch | 5 working days of fixes and a report on day 10 | 5 working days of fixes and weekly transcript reviews for 2 weeks |
| Fits when | Your FAQ exists and the questions repeat | Tone matters, or the questions are sensitive or varied |
How VITON13 does it
We launch an AI chatbot on one channel that answers only from content you approve, collects the request and passes the conversation to a named person when it cannot help or the visitor asks. Scope, provider fees and the handoff rule are agreed in writing before the build.
AI chatbot with human handoff
- Price
- $73 AI-assisted or $93 human-led
- Timeline
- 5–8 working days
The final price is fixed in writing after a short scope review.
AI Use-Case Session
- Price
- $53 standard implementation or $73 custom implementation
- Timeline
- 1 working day
AI-assisted: AI drafts inside defined steps, a person directs and checks every result. Human-led: a specialist does the work and AI assists.
What is included
- 1 channel: your website, WhatsApp Business or Telegram
- A knowledge base of up to 30 question-and-answer pairs or 20 pages of documents
- 3 flows: answering questions, capturing a request, handing over to a person
- Up to 5 captured fields sent to 1 destination: email, a Google Sheet or a CRM with an accessible API
- 1 handoff channel (email, Telegram or WhatsApp) with the transcript and a working-hours rule
- 20 scripted test conversations before launch (40 human-led)
- 5 working days of fixes, an admin guide for editing answers and a report on day 10
- Delivery window: 5–8 working days once content, channel access and the handoff contact are in
Not included
- Platform, model and WhatsApp Business fees and Meta verification, paid by you to each provider
- Extra channels, quoted each
- Payments or bookings inside the chat, quoted as an integration
- Bots that give medical, legal or financial advice
- Outbound bulk messaging
Revisions: 2 rounds on answers and flows before launch
Get an estimate in 1 working dayWritten estimate within 1 working day
Example
INTERNALOur own product
VIT LIVE: our own assistant, rebuilt so it stops guessing prices
An early version of the assistant in VITON13's signed-in VITON ID cabinet gave the model only site-search hits: page titles and paths. Asked what a website costs, it had no figure, so it hedged or invented one. The rebuild reversed the order: a catalogue of real items with real prices answers first, in cards; fixed replies cover prices, ordering and bonuses; the self-hosted model comes last and must say a price is confirmed after the brief instead of guessing.
What this does not prove
- It is not a public widget: only signed-in members use it, so it says nothing about visitor conversion.
- It has no lead capture and no handoff to a person, the two parts this guide calls essential; the launch package adds both.
- There are no usage figures, and it does not show the model running in production or grade its answers.
Read the case: A VITON ID assistant that answers from its own catalogue first
Questions and answers
How long does a chatbot launch take?
5–8 working days, counted from the day your answers, channel access and the person taking handoffs are ready. Roughly two days on content and rules, two on building, two on scripted tests. Launches slip when approved answers arrive late.
Who owns the chatbot, the answers and the conversation logs?
You do. The platform chosen in the written scope runs under your account, the answers are your content, and an admin guide shows how to edit them. How long conversations are stored is agreed in writing and set before launch, so changing supplier does not cost you your history.
What happens if the bot gives a wrong answer?
It is built to answer only from the approved knowledge base and to offer a person when a question falls outside it. Scripted tests before launch include attempts to push it off topic. After launch, five working days of fixes cover faults, and unanswered questions are listed so you can add the missing answers.
Will the chatbot bring more leads or sales?
Nobody can promise that honestly, and we do not. We measure what the bot does: questions answered, conversations handed over and why, contacts captured with complete fields, and questions it could not answer. The first report comes on day 10, so you judge the effect against your own numbers.
Do I pay VITON13 every month?
Not for the launch package: it is a one-off job. Monthly costs go to the platform, the model provider and, on WhatsApp, Meta, from your own accounts, and the estimate lists them. New channels, integrations or changes after the fixes period are quoted before any work starts.
How does payment work, and can I cancel?
A written scope, timeline and final price arrive within one working day of your request. Nothing is charged before the total and payment route are confirmed; routes include cards, PayPal and bank transfer, depending on your country. Cancelling is free until you approve the scope; after that, only completed work is charged.
Sources
- RaftLabs — Chatbot development cost —
- Cleveroad — Chatbot development cost guide —
- Anthropic — Claude API pricing —
- Intercom — Pricing —
- Meta — WhatsApp Business Platform pricing —
- BLS — Customer service representatives —
- Insurance-Canada.ca — Gartner: customers want access to human agents —
- Insurance-Canada.ca — Gartner: few would retry a chatbot after a bad experience —
- AI Act Service Desk — Article 50 —
- AI Act Service Desk — Implementation timeline —