I get this question three or four times a week. It comes in every form: "how much for a chatbot", "what does a custom AI bot cost", "we have a quote for twenty thousand euros, is that fair". The honest answer is that the price range for a real custom AI chatbot in 2026 spans an order of magnitude, and the spread is not arbitrary — it maps directly to specific scope decisions you can understand in ten minutes.
This guide is based on pricing I have actually delivered across three very different chatbot projects in the last two years: an e-commerce support bot for a Barcelona fashion store (EUR 3,000, 5 weeks), a Telegram bot with monetised AI model aggregation (EUR 6,500, 6 weeks), and an Instagram audit bot with a 10,000-line codebase (EUR 14,000, 10 weeks). All fixed price, all solo developer, all in production.
The four price tiers and what separates them
Before the math, the map. Custom AI chatbots fall into four honest tiers in 2026. Picking the right tier is more important than picking the cheapest developer.
Tier 1 — Scripted bot with LLM flavour (EUR 800 to 2,000)
A widget on your site that answers from a fixed knowledge base of 10-50 FAQs. OpenAI or Claude is used to rephrase the answers naturally, not to reason. No connection to your real-time data (inventory, orders, CRM). Typical delivery: 1-2 weeks.
Fair for a landing page, a small service business, or a product with a stable FAQ. If your customers ever need to know "is this in stock" or "where is my order", this tier will frustrate them.
Tier 2 — RAG-grounded bot with limited actions (EUR 2,500 to 5,000)
Connects to your actual content — product catalog, documentation, policies — via retrieval-augmented generation. The bot reads the same database your staff reads. Supports 1-3 external actions like "check order status" or "look up stock". Escalation to a human when uncertain. Typical delivery: 3-5 weeks.
This is where most of my e-commerce and SaaS work sits. The fashion store bot I mentioned was in this tier — 3,000 SKUs, stock tool call, order lookup, human handoff. This is the tier that pays back in two to three months for most businesses.
Tier 3 — Agentic bot with domain-specific workflows (EUR 6,000 to 15,000)
Multi-step reasoning. The bot plans actions, calls multiple tools, and can complete transactions end-to-end — book an appointment, process a return, generate a quote. Includes admin panel, user management, billing integration if monetised. Typical delivery: 6-10 weeks.
The Telegram AI aggregator sits here. Five different AI models, credit system, Stripe plus crypto plus Telegram Stars payments, admin panel, queue prioritisation. Roughly 4,000 lines of code across the bot and the admin.
Tier 4 — Domain platform with a chatbot interface (EUR 15,000 to 40,000)
At this tier the "chatbot" is just the interface to a full application. Custom algorithms, heavy data pipelines, multiple external integrations, bespoke analysis. The Instagram audit bot — 132 tests, 10,000 lines of Python, four-stage data-retrieval pipeline — is this tier.
If someone is quoting you EUR 20,000 for a "chatbot" that is really a platform with a chat interface, they may well be honest. If someone is quoting you EUR 20,000 for what belongs in Tier 2, they are probably padding.
Where the money actually goes
Every chatbot I have shipped has roughly the same cost structure. Knowing this helps you read quotes critically.
Discovery and scope (8-12% of total). Not glamorous but non-negotiable. A thirty-minute call plus a written scope that lists the specific questions the bot must answer and the specific actions it must perform. If a developer skips this step and offers a flat "build me a chatbot" quote, you will pay for the misunderstanding later in change requests.
Backend integration (25-35%). Connecting the bot to your real data — product database, order system, CRM — is usually the largest line item. This is where custom work lives and where templates stop helping. A Shopify store is easier; a bespoke legacy database is harder. The quote should name the integrations by system and spec them.
Retrieval layer and prompts (15-20%). Embedding your content, building the retrieval query, writing and testing the system prompts. For a 1,000-5,000 item catalog this is a few days of work and a few euros of OpenAI calls for embeddings. For a 100,000-document corpus it is a substantial engineering task and should be quoted separately.
Frontend widget and UX (10-15%). The chat widget itself, typing indicators, message history, file uploads if needed, escalation button, mobile responsiveness. This is the fastest part of the build now — there are good open-source widgets you can adapt.
Admin panel and logs (10-15%). Your ability to see what the bot said, retrain it, flag bad responses, and handle escalations. Tempting to cut from the first version; regrettable not to have by week six.
Testing, deployment, and handover (10-15%). A test set of representative questions with expected behaviour, production deployment, monitoring, documentation. A quote that does not include these is quoting you a demo, not a product.
Running costs: what the bot costs after it ships
The capex is half the conversation. Here is what three real bots cost me to run for clients, per month, averaged over six months of production.
Fashion e-commerce bot (1,500-2,000 conversations/month):
- OpenAI API: EUR 140-180
- Hosting delta (VPS + Redis): EUR 25
- Embeddings re-indexing: EUR 8
- Total: EUR 175-215/month
Telegram AI aggregator (50,000+ AI calls/month at scale):
- OpenAI + image model pass-through: EUR 800-1,200 (billed to users, not the owner)
- Infrastructure (bot server, worker queue, Postgres): EUR 45
- Stripe fees: 2.9% of revenue
- Owner's net cost: around EUR 60/month because the bot's own pricing absorbs API costs
Instagram audit bot (pay-per-audit):
- Data retrieval APIs (Apify + fallback): EUR 0.05-0.15 per audit
- OpenAI for analysis: EUR 0.03 per audit
- Infrastructure: EUR 35
- Break-even at roughly 200 paid audits/month
The pattern: for internal-use bots (your own customer support), expect EUR 100-300/month in variable costs. For monetised bots where users pay per use, the economics flip and the bot can pay for its own infrastructure and then some by month two or three.
Hidden costs that nobody mentions in the initial quote
If I am writing an honest guide, these deserve equal billing with the sticker price.
Prompt maintenance. A system prompt that was great in January will drift by July as your product catalog changes, your policies evolve, and the underlying model gets updated by OpenAI. Budget two to four hours a month, or EUR 150-300, for prompt hygiene.
Retrieval quality drift. Embeddings generated by text-embedding-3-small will be replaced by something better within a year. Re-embedding a catalog is cheap; re-tuning the retrieval logic to match is an afternoon of work.
Escalation volume. If your bot handles 70% of conversations, the remaining 30% still need a human. Do not use the bot to justify firing your support team — use it to let them focus on the hard cases. Pretending otherwise is how projects get cancelled at month three.
Training data curation. If you want the bot to improve, somebody has to review logs, flag good and bad responses, and feed them back into the prompt. One hour a week for the first three months, tapering after. If you do not do this the bot freezes at its week-one performance, which is usually fine but rarely great.
Red flags in a chatbot quote
A quote for a bot deserves the same scrutiny as any software project. The red flags I have seen repeatedly:
- No named integrations. The quote says "connects to your systems" without naming the systems or the data it will read or write. That clause will be renegotiated painfully in week three.
- "Unlimited revisions". Either a hidden hourly rate that will surface later, or a developer who cannot scope their own work. Good fixed-price quotes are specific: two rounds of prompt refinement after launch, then hourly.
- No escalation path specified. The bot should not try to handle refunds or legal questions. If the quote does not describe what happens when the bot is unsure, the answer is that the bot will just guess, and you will find out from an angry customer.
- No test set. "How will we know the bot works?" should have a concrete answer. A list of 30-50 representative questions with expected behaviour is the minimum. Without one, "it works" is whoever-shouts-loudest.
- Proprietary platform lock-in. If the bot is built on a vendor's platform where you do not own the prompts, the embeddings, or the conversation logs, your exit cost is infinite. Clarify ownership in writing.
When NOT to build a custom chatbot
Not every business needs one. I have talked two prospective clients out of the project in the last year alone.
If you have fewer than 20 support tickets a day, the math does not work. A part-time VA at EUR 15/hour handling ten tickets a day costs you EUR 300/month. A custom bot starts paying back around 100 tickets/day. Below that, scripted FAQ and a well-organised help centre win on cost.
If your inquiries are genuinely unique each time, RAG will struggle. A bespoke tailor gets questions that have never been asked before. Those are hard for any retrieval system and need human judgement. Use the bot for the 20% that are repeatable (hours, location, booking), escalate the rest.
If your data is a mess, fix that first. A chatbot trained on your inconsistent product descriptions will confidently give inconsistent answers. Two weeks of catalog cleanup often delivers more than a month of AI plumbing.
Timeline reality check
Developers optimise their quotes. Here are realistic delivery windows based on actual work I have shipped, not calendar optimism:
- Tier 1 (scripted + LLM rephrase): 1-2 weeks to a working demo, 3-4 weeks to a production-ready deploy with your branding, logs, and analytics.
- Tier 2 (RAG + tool calls): 4-6 weeks from kickoff to production, assuming your data is accessible. Add 2 weeks if I have to reverse-engineer an old database or build a new export pipeline.
- Tier 3 (agentic + admin panel): 8-12 weeks. The admin panel, billing integration, and multi-step flows are the long poles.
- Tier 4 (platform with chatbot UI): 10-16 weeks. This is a full product, not a widget.
Anyone quoting you a Tier 2 bot in two weeks is either reusing 80% of a previous project (fine, ask them) or cutting corners on testing and deployment (not fine).
FAQ
Can I use ChatGPT's enterprise plan instead of building custom?
For internal knowledge search across documents, yes — ChatGPT Enterprise or a Claude workspace can replace a custom RAG bot for employees. For a customer-facing bot tied to your own data and branding, no. Those enterprise products are not designed to be embedded on your storefront.
What about no-code platforms like Voiceflow, Chatbase, Botpress?
Chatbase and similar "upload your docs" bots are fine for Tier 1 use. They break down fast on Tier 2 and above because you cannot wire them into a bespoke backend without escape hatches that negate the no-code value. For under EUR 3,000 of scope, they are competitive; above that, custom wins.
Will GPT-5 or Claude 4 make my bot obsolete?
Model upgrades improve the floor of quality without changing the architecture. The bot I built in 2024 got measurably better when GPT-4o was released — same code, better model. That is the main reason I build on the API and not inside a proprietary platform.
How do I evaluate a developer's portfolio for this kind of work?
Ask for a live demo of a bot they have shipped, not a video. Ask who owns the prompts and embeddings (the client should, always). Ask to see a log of a conversation where the bot failed, and what they did about it. Anyone who has shipped real production bots will have stories; anyone who does not will pretend their bots have never failed.
Is it better to hire a freelancer, an agency, or an in-house dev?
For a single chatbot project, freelancer wins on price by 2-3x over an agency for the same deliverable. In-house makes sense once you are building five or more AI products and need continuity. I wrote a longer piece on this — the short version is: if the project has a clear end, freelance; if it is a programme, hire.
How do I budget for maintenance?
Budget 15-25% of the build cost per year for maintenance — that covers prompt updates, library upgrades, the occasional bug, and small feature additions. If your bot is deeply integrated with your business logic, the higher end of that range. If it is a support widget answering FAQs, the lower.
If you are ready to scope one
The fastest way to get a real number is a 15-minute call where I ask you the same five questions I ask every new client: what volume of conversations, what data the bot needs to read, what actions it needs to perform, what languages, and what "good enough" looks like for your business. With those answers I can usually place a project in a tier and return a fixed-price proposal within 24 hours.
If you are comparing quotes, the most useful thing you can do is share this guide's tier framework with the other developers and ask which tier they are quoting. A developer who cannot answer that question is probably not the right choice, regardless of price.