A custom AI chatbot costs $1,500 to $3,000 for a website or support bot, $3,000 to $6,000 for a RAG assistant that answers from your documents, and $6,000 to $12,000 for an enterprise assistant with integrations and access control. Running it costs anywhere from a few dollars to a few hundred dollars a month, depending on chat volume and the AI model you choose.
This guide shows where that money goes: the line items in a real quote, three example projects with their build and running costs, the token math behind the monthly bill, and the compliance work US, UK, and EU buyers need to plan for. If you already know what you need, see our AI chatbot development services.
Key takeaways
- The chat window is the cheapest part of a chatbot. Data, integrations, and access control set the price.
- A RAG chatbot answering 10,000 questions a month costs about $12 in AI usage on a budget model, or $118 on a stronger one, plus about $25 for the database.
- Most of a good quote should go to retrieval quality and testing. If those lines are missing, the low price is not a bargain.
- EU users must be told they are talking to an AI from 2 August 2026. Budget for disclosure, data agreements, and data residency from day one.
About this guide. Written by Hashir Siraj, CEO of Emergen, who scopes and quotes our chatbot projects. Technical review by Huzaifa Imran, CTO, who designs our RAG pipelines. Emergen builds chatbots, so we sell what this guide prices; where a cheaper tool is the better choice, we say so. All model and database prices were checked against the providers’ pricing pages on 12 October 2026.
Chatbot development cost by type
| Chatbot type | Build cost | Timeline | What gets built | Typical running cost |
|---|---|---|---|---|
| Website/support chatbot | $1,500 – $3,000 | 1–2 weeks | Answers from your FAQs, help pages and policies; hand-off to a person or a contact form | A few dollars a month in AI usage, plus hosting |
| RAG knowledge assistant | $3,000 – $6,000 | 3–5 weeks | Document pipeline, search and reranking, answers with source links, a test set | $25–$150 a month for most volumes |
| Enterprise AI assistant | $6,000 – $12,000 | 6–10 weeks | Multiple sources, logins and roles, CRM/HR/ticketing integrations, audit logs, security review | $100+ a month, driven by users and model choice |
Which band is your chatbot in? A 1-minute check
- Does the bot need to answer from more than a few web pages (PDFs, a help centre, internal docs)?
- Does it need to look things up in, or write to, another system (CRM, orders, tickets, WhatsApp)?
- Should different users see different answers based on their role?
- Will staff use it for internal or confidential information?
- Must your data stay in a specific region or on your own servers?
All no: support bot band, or a builder tool may be enough. Yes to question 1 only: RAG assistant band. Yes to two or more of questions 2 to 5: plan for the enterprise band.
What’s inside a chatbot quote, line by line
Two quotes for “an AI chatbot” can differ by thousands of dollars because they include different work. Use this table to compare quotes side by side. If a line is missing, ask why.
| Line item | What it covers | Ask the vendor | Red flag if missing |
|---|---|---|---|
| Discovery and scope | Use cases, real user questions, success criteria | “Which questions must it answer on day one?” | Quote with no list of use cases |
| Content pipeline | Extracting text from PDFs and pages, splitting it into passages, keeping it in sync | “How do you handle tables, scans, and updated files?” | “We just upload your files” |
| Search and ranking | Finding the right passages by meaning and keywords, then reranking them | “Do you use hybrid search and a reranker?” | No mention of how answers are found |
| Prompts and guardrails | Instructions, tone, refusal rules, prompt-injection defences | “What happens if a user asks it to ignore its rules?” | No guardrails at all |
| Chat interface and channels | Website widget, WhatsApp, Slack or Teams | “Is each extra channel priced separately?” | Channels promised with no time allowed |
| Integrations | CRM, helpdesk, order system, calendars | “Read-only, or read and write?” | “Connect to your CRM” with no detail |
| Access control and logs | Logins, roles, audit trail | “Can a sales rep see HR documents?” | Internal data with no roles |
| Evaluation set | Real test questions with expected answers, re-run after changes | “Can we see the test results before launch?” | Testing described as “QA” |
| Deployment and monitoring | Hosting, error alerts, usage dashboard | “How will we know if answers get worse?” | No monitoring after launch |
| Handover | Code, prompts, documentation, admin training | “Do we own the code and prompts?” | A licence instead of ownership |
For the full vendor checklist, see 15 questions to ask an AI development company.
What pushes the price up
1. Scope: how many jobs the bot does
Answering pricing questions is one job. Add booking demos, checking order status, and creating support tickets, and you have four. Each job needs its own prompts, test questions, and fallbacks. Shipping one job first and adding the rest later is the single best way to keep the first invoice small.
2. Your data and how messy it is
Fifty clean help-centre articles take hours to load. Three thousand PDFs with scanned pages, tables, and outdated drafts take days, because each file type needs its own extraction and the outdated versions must be removed or the bot will quote them. Content that changes daily, such as prices or stock, also needs a sync job.
3. Integrations
“What are your opening hours?” needs no integration. “Where is my order?” needs read access to your order system, a way to check the customer is who they say they are, and a fallback when the order is not found. Writing data back, such as creating a ticket, adds error handling on top. Abee, our enterprise example below, connects to the client’s CRM and HR systems, which is a big part of why it sits in the top band.
4. Security and access control
If the bot can see internal documents, permissions must be applied at the search step, not just at login, so a sales rep never gets an answer built from an HR file. Add audit logs and defences against prompt injection (users tricking the bot into ignoring its rules). The OWASP Top 10 for LLM Applications lists the risks a security review should cover.
5. Hosting: provider API or your own servers
Calling a hosted model through an API is the fastest and cheapest route. OpenAI and Anthropic state that they do not train on API data by default, and your documents can stay in your own database. Running an open model on your own servers is possible, but it adds GPU costs and upkeep, so keep it for cases where a contract or regulator requires it.
6. Testing
A good evaluation set is 30 to 50 questions taken from your real inbox or helpdesk, each with the answer you expect, including the awkward ones: refunds, complaints, questions about competitors, and things the bot should refuse. It is re-run after every prompt, model, or content change. This is the line cheap quotes drop first.
Three example projects and what they cost
These are typical scenarios we scope, not specific clients. Running costs use OpenAI list prices on 12 October 2026.
A US dental group: website FAQ bot
- What it does: answers questions about opening hours, insurance, prices, and aftercare across 6 clinics, and hands booking requests to the front desk.
- What gets built: chat widget, about 80 FAQ entries, hand-off form, test set.
- Band: support bot, $1,500–$3,000, about 2 weeks.
- Running cost: about 2,000 chats a month on GPT-5.6 Luna costs roughly $1–2 in AI usage, plus hosting.
- Main risk: the bot giving clinical advice. It needs a firm rule to refer those questions to staff.
A UK SaaS company: help-centre RAG assistant
- What it does: answers product questions from 400 help articles and release notes, links to the source, and creates a helpdesk ticket when it cannot answer.
- What gets built: content pipeline with nightly sync, hybrid search and reranking, ticket integration, test set of 50 real questions.
- Band: RAG assistant, $3,000–$6,000, 3–5 weeks.
- Running cost: about 8,000 chats a month on Luna is roughly $9 in AI usage, plus $25 for the database.
- Main risk: answering from outdated release notes. The sync job and the test set catch this.
An EU logistics firm: internal staff assistant
- What it does: answers 600 staff questions on HR policies, safety procedures, and customer contracts, with each employee seeing only what their role allows.
- What gets built: connectors to document storage and the HR system, role-based search, audit logs, EU-hosted database, AI disclosure, security review.
- Band: enterprise assistant, $6,000–$12,000, 6–10 weeks.
- Running cost: about 6,000 questions a month on GPT-5.6 Terra is roughly $71 in AI usage, plus $25 for the database.
- Main risk: permission leaks between departments. Roles must be tested as hard as the answers.
What it costs to run a chatbot each month
Where the tokens go in one chat
AI providers charge per token, about ¾ of a word. In a typical RAG chat, the model reads far more than the user types:
| Part of the request | Tokens (approx.) |
|---|---|
| System instructions and rules | 600 |
| 4 retrieved passages from your documents (about 600 each) | 2,400 |
| Recent conversation history | 400 |
| The user’s question | 100 |
| Total sent to the model | 3,500 |
| Answer written by the model | 400 |
That is why the passages you retrieve, not the question, drive the bill. Prices used: GPT-5.6 Luna at $0.20 per million tokens in and $1.20 out; GPT-5.6 Terra at $2 in and $12 out (OpenAI pricing). The vector database runs on the Supabase Pro plan at $25 a month, using pgvector.
Monthly cost by chat volume
| Chats per month | GPT-5.6 Luna (budget) | GPT-5.6 Terra (stronger) | Database (Supabase Pro) |
|---|---|---|---|
| 1,000 | $1.18 | $11.80 | $25 |
| 10,000 | $11.80 | $118 | $25 |
| 50,000 | $59 | $590 | $25 |
At low volume, the database costs more than the AI. At high volume, model choice dominates. For the integration side of the budget, see our LLM integration cost and timeline guide.
Five ways to cut running costs
- Cache the fixed instructions. Repeated input is billed at $0.02 per million tokens on Luna instead of $0.20, and $0.20 instead of $2 on Terra. Caching a 600-token system prompt saves about $54 a month at 50,000 Terra chats.
- Route by difficulty. Send routine questions to Luna and only complex ones to Terra.
- Retrieve fewer, better passages. A reranker lets you send 3 strong passages instead of 8 weak ones, which cuts input tokens and usually improves answers.
- Cap answer length. Output tokens cost 6 times more than input tokens on both models.
- Reuse answers to repeat questions. Store answers to the top questions and serve them without calling the model.
Costs people forget to budget for
- WhatsApp fees. Meta charges its own per-message fees for some WhatsApp Business Platform messages, separate from your build and AI costs. See Meta’s WhatsApp pricing.
- Keeping content fresh. Someone must update the documents the bot answers from. Budget a few hours a month.
- Re-testing after model updates. Re-run your evaluation set before switching to a new model version.
- A stronger model later. OpenAI’s flagship GPT-5.6 Sol costs $5 in and $30 out per million tokens, 2.5 times Terra.
Compliance for US, UK and EU buyers
Compliance work changes both the build cost and the timeline, so raise it in the first call, not after launch.
- EU AI Act disclosure. From 2 August 2026, chatbots used in the EU must make clear to users that they are talking to an AI (Cooley summary of Article 50). This is a small build task but easy to forget.
- GDPR and UK GDPR. You need a data processing agreement with your developer and with the model provider, and a clear answer on what personal data the bot stores and for how long.
- Data residency. OpenAI offers data residency in Europe for eligible API customers, and Supabase lets you choose an EU region for your database.
- US sector rules. If the bot will handle health, financial, or children’s data, check the rules for your sector (for example, HIPAA for patient data) and confirm the model provider will sign the agreements they require.
How to know the chatbot works before you pay the final invoice
Put measurable acceptance criteria in the contract instead of “it works well”:
- A target share of the evaluation set answered correctly, agreed before the build starts
- Every answer that uses your documents links to its source
- When the bot is unsure, it hands off to a person instead of guessing
- Role tests pass: no user can get an answer from documents they cannot access
- A named person re-runs the evaluation set after each change, and you see the results
Build vs buy: when a chatbot builder is enough
Builder tools are a good start when the bot only answers from a few public pages and does not need your systems. Go custom when it must search private documents, connect to your CRM, helpdesk or WhatsApp, show different users different information, or keep data under your control. To decide on cost, add up 24 months of the builder’s per-seat or per-conversation fees and compare that with a one-off build plus the running costs above.
Why our ranges differ from other chatbot cost guides
Other guides quote much higher numbers. RaftLabs puts an LLM chatbot at $24,000–$60,000, and Monocubed puts generative AI chatbots at $75,000–$300,000. Those ranges cover larger teams and bigger scopes. Ours are lower for two reasons:
- Reused building blocks. We have already built the document pipeline, search, and access control for projects like Abee, so we do not start from zero.
- Tight scope. Each band covers one main use case. Extra channels, integrations, or custom model training are quoted on top, so you only pay for what you need.
A real example: Abee
Abee is an enterprise AI assistant we built for a hospitality business in France, used by 600+ people. It searches company documents, answers voice questions, connects to the client’s CRM and HR systems, and uses MFA and role-based access. The client reports 50% faster task completion and zero security incidents since launch.
What to send us for an accurate quote
Bring these to the call, and we can give you a fixed price and timeline. Rough answers are fine.
- The one job the bot must do first, in one sentence
- 10 real questions customers or staff ask today
- Where the answers live: links, a folder or an export, and roughly how many documents
- The systems it must connect to (CRM, helpdesk, WhatsApp, database)
- Who will use it, and whether different users should see different information
- Your expected chats per month, so we can estimate running costs
- Your data rules: NDA, DPA, region, or own servers
Our usual stack is OpenAI and Anthropic models, Supabase with pgvector or Pinecone for search, and Next.js and Node.js for the app. We sign NDAs and DPAs, offer an optional support plan after launch, and you own the code and prompts.
Frequently asked questions
A custom website or support chatbot costs about $1,500 to $3,000. A RAG assistant that answers from your documents costs $3,000 to $6,000. An enterprise assistant with integrations and access control costs $6,000 to $12,000. Running costs are separate and depend on chat volume and model choice.
At OpenAI list prices in October 2026, a RAG chatbot uses about $1.18 of AI per 1,000 chats on GPT-5.6 Luna, or $11.80 on GPT-5.6 Terra. Add about $25 a month for a Supabase Pro database, plus hosting. At 10,000 chats a month, that is roughly $37 to $143 in total before hosting.
A RAG chatbot usually costs $3,000 to $6,000 to build and takes 3 to 5 weeks. Messy or very large document sets, frequent content updates, several data sources, or role-based permissions can push it into the enterprise band.
For simple FAQ bots on public content, usually yes. Compare 24 months of the builder’s fees with a one-off build plus running costs. Custom development pays off when you need private documents, integrations, access control, or data kept under your control.
A website or support chatbot takes 1 to 2 weeks, a RAG assistant 3 to 5 weeks, and an enterprise assistant 6 to 10 weeks. Delays usually come from content preparation and waiting for system access, not from development.
In the EU, yes. From 2 August 2026, the EU AI Act requires chatbots to make clear to users that they are interacting with an AI system. It is good practice in the US and UK too, and it is a small build task.
Yes. OpenAI and Anthropic state they do not train on API data by default; documents can stay in your own database; OpenAI offers European data residency for eligible customers; and a self-hosted open model is possible when a contract or regulator requires it.
Get a fixed quote for your chatbot
A 15-minute call is enough to place your chatbot in a band, flag compliance needs, and tell you honestly whether a builder tool would do the job.
Keep reading: AI chatbot development services · How to choose an AI development company · AI SaaS development cost
