Here is the short version. Off-the-shelf chatbot widgets run roughly $0–100/month and need almost no engineering. A custom chatbot trained on your own data typically starts around $6,000 to build, plus a modest monthly cost for LLM tokens and hosting. More complex RAG or tool-using agent systems run higher — usually from ~$15,000 — because they involve retrieval pipelines, integrations, and guardrails. The rest of this post explains where those numbers come from so you can size your own project.
The two costs that make up any chatbot
Almost every quote you get can be broken into two parts, and confusing them is the most common budgeting mistake we see.
- Build cost (one-time). The engineering to design, connect, ground, test, and ship the chatbot. You pay this once (plus any later feature work).
- Ongoing cost (monthly). What it costs to run the thing — mostly LLM tokens for each conversation, plus hosting, vector storage, and monitoring.
A cheap tool hides both numbers inside a flat subscription. A custom build separates them, which is why it looks more expensive up front and is often cheaper at scale. Keep the two buckets separate whenever you compare options.
Chatbot types and what they cost
The single biggest driver of price is which kind of chatbot you actually need. Here is how the common tiers break down in 2026.
| Chatbot type | What it does | Build cost | Ongoing cost |
|---|---|---|---|
| Off-the-shelf widget | Scripted flows or generic AI answers on a hosted platform; little to no data grounding | $0 (setup only) | ~$0–100/mo |
| Custom, data-trained | Answers grounded in your docs, tuned tone/UX, connected to your site or app | from ~$6,000 | ~$50–500/mo usage |
| RAG assistant | Retrieval over a large knowledge base with citations, freshness, and access control | from ~$15,000 | ~$200–1,500/mo usage |
| Tool-using agent | Takes actions across your systems — books, updates, files things — with guardrails | ~$25,000+ | usage + integration upkeep |
These are honest starting ranges, not fixed prices. A small, single-source custom chatbot lands near the bottom; anything with many data sources, multiple integrations, or strict compliance moves up. If you are still deciding between "answers questions" and "does the work," our post on AI agents vs. chatbots draws the line clearly — the difference is worth thousands of dollars in scope.
What drives the build cost
Two custom chatbots can differ by 3–4x on build price. Here is what moves the number.
Data volume and messiness
Grounding a bot on one clean FAQ page is quick. Grounding it on 5,000 support tickets, a decade of PDFs, and three product manuals is not. The work is in cleaning, chunking, and indexing that content well enough that retrieval returns the right passage. More sources — and messier sources — mean more engineering.
Integrations
A bot that only reads your docs is cheaper than one that also pulls a customer's order status from your backend or writes a ticket into your helpdesk. Every system it touches is an API to authenticate, map, and error-handle. Integrations are usually the line item that pushes a project from "custom chatbot" into "RAG" or agent territory.
Guardrails and accuracy
Getting a demo working is easy; making it safe is where senior time goes. Grounding responses so the model only answers from approved content, adding citations, handling "I don't know" gracefully, filtering unsafe inputs, and logging everything for review — this is real work, and it is exactly what separates a toy from something you can put in front of customers.
Custom UX
A drop-in widget looks like every other drop-in widget. A branded interface — streaming responses, suggested prompts, handoff to a human, mobile-friendly layout, analytics — adds design and front-end work. It is often worth it, but it is a cost.
Ongoing cost: the token math
The monthly number scares people until they see the arithmetic. Ongoing cost is dominated by LLM token usage, and tokens scale with traffic.
A rough model: each chatbot reply consumes tokens for the user's question, the retrieved context you feed the model, and the answer itself. A grounded response commonly runs a few thousand tokens end to end. At 2026 API prices for a solid mid-tier model, that is a fraction of a cent to a few cents per message depending on the model and how much context you retrieve.
Multiply by volume:
- A few hundred conversations a month: often under $50 — genuinely cheap.
- Thousands a month: tens to a few hundred dollars.
- High traffic or large retrieved context: hundreds to low thousands.
On top of tokens you have hosting (usually modest), a vector database for retrieval (often $0–100/mo at small scale), and monitoring. The practical levers are model choice (a smaller model for easy questions, a larger one only when needed) and how much context you retrieve per query — feeding the model less, more relevant text cuts cost and often improves accuracy. We pick models to match the job rather than defaulting to the most expensive one; see how we approach that in AI development.
Off-the-shelf vs custom
You do not always need a custom build, and we will tell you when you don't.
Off-the-shelf makes sense when your needs are standard, your content is small and stable, and "good enough" answers are fine. Our own no-code sister product, CertifChat, exists for exactly this case — you point it at your content and get a working assistant without an engineering project. If a hosted tool solves your problem, use one; paying for custom work you don't need is just a more expensive mistake.
Custom makes sense when you need answers grounded in your private data, real integrations, control over tone and UX, tighter accuracy and guardrails, or ownership of the system rather than renting it. The build cost buys you a bot that behaves the way your business actually works — and the ongoing cost is usually lower per conversation than a per-seat or per-message SaaS plan once volume climbs.
The honest framing: off-the-shelf trades control for speed; custom trades up-front cost for fit and long-term economics.
Hidden costs to watch
The sticker price is not the whole picture. Budget for these too.
- Maintenance. Your content changes. The bot's knowledge base has to be re-indexed as docs, prices, and policies update — either on a schedule or automatically.
- Tuning. The first version is never the best version. Reviewing real conversations, fixing weak answers, and adjusting retrieval is ongoing, higher-value-early work.
- Model drift and upgrades. Providers release new models and deprecate old ones. Occasionally re-testing and swapping models keeps quality up and cost down.
- Monitoring. You want logging and alerts so you catch bad answers before customers do. Cheap to set up, expensive to skip.
None of these are large, but pretending they are zero is how a "$50/month tool" quietly becomes a support headache.
Frequently asked questions
Is a custom chatbot worth it vs a $50/mo tool?
If a $50/mo tool answers your questions accurately, keep it — that is the right call for standard needs and small, stable content. A custom build (from ~$6,000) earns its cost when you need answers grounded in your private data, real integrations, brand-controlled UX, or stronger accuracy and guardrails than a generic tool provides. At higher conversation volumes, custom is often cheaper per interaction than per-seat SaaS pricing, on top of fitting your business better.
What are the ongoing costs of running a chatbot?
Mostly LLM tokens, which scale with traffic — roughly a fraction of a cent to a few cents per message. Light usage often stays under $50/month; heavier or higher-context usage runs into the hundreds. Add modest hosting, a vector database for retrieval, and monitoring. The big levers are model choice and how much context you retrieve per query.
How do I stop it hallucinating?
You constrain it. The standard approach is retrieval-augmented generation (RAG): the bot answers only from approved content you retrieve at query time, cites its sources, and says "I don't know" instead of guessing. Good grounding, tight retrieval, and honest fallbacks eliminate most hallucinations — it is the core of solid RAG development and a major reason custom builds cost more than a raw widget.
How long does a custom chatbot take to build?
A focused, single-source custom chatbot can ship in a few weeks. RAG assistants and tool-using agents take longer because of retrieval pipelines, integrations, and guardrail testing. We usually scope a small pilot first to prove accuracy and ROI before committing to the full build.
Getting a number for your project
The only way to a real figure is a defined scope: which data, which integrations, how much traffic, and how strict the accuracy bar. After a short discovery call we give a fixed, itemized estimate and a phased plan so you can start lean and expand once the bot proves itself.
If you want to price a build, our AI chatbot development service is the place to start — and if you already know you need retrieval over a large knowledge base, RAG development covers that end to end. Get in touch with your use case and rough traffic, and we will come back with an honest range rather than a hype number.