How Much Does an AI Chatbot Cost in 2026? A Full Pricing Breakdown

Short answer: between about $4,000 and $300,000 to build, plus roughly $0.004 to $0.075 every time someone has a conversation with it.

That second number is the one almost nobody publishes, and it is the one that decides whether your chatbot is profitable. Most pricing guides give you a build range and stop at "ongoing model costs vary by usage." They vary in a way you can calculate, and this article shows the arithmetic.

What follows: build cost by chatbot type, the per-conversation token maths, a full year-one budget for a realistic project, and the point at which building beats paying a per-resolution SaaS platform.

What it costs to build

Build cost tracks one thing above all others: how much the bot has to know, and how many systems it has to touch.

Scripted or rule-based bot: $4,000 – $20,000

Decision trees and buttons. No language model, so no per-conversation cost. Still genuinely useful for narrow, predictable flows like booking or order status, and dramatically cheaper to run. If your queries are that predictable, do not pay for AI.

AI chatbot MVP: $15,000 – $45,000

One use case, a language model, a small knowledge base, one channel. This is the right first step for most businesses: it proves whether people will actually use the thing before you commit to the full build.

Custom RAG chatbot: $35,000 – $90,000

Answers from your own documents, help centre, product catalogue or policies, with retrieval so it cites real content instead of inventing it. This is where most serious business chatbots land in 2026, and where the cost of preparing your content starts to matter as much as the engineering.

Multi-channel conversational AI: $75,000 – $180,000

Web, WhatsApp, in-app and sometimes voice, with CRM and helpdesk integration and a handoff to human agents. Each channel is a separate build, not a toggle.

Enterprise platform: $150,000 – $300,000+

Multiple bots, role-based access, audit logging, SSO, data residency, and the review processes that come with regulated industries.

The running cost nobody shows you

Here is where most pricing guides go quiet. Every conversation costs money, because every message sends tokens to a model and gets tokens back.

The formula is simple:

(input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate)

Take a realistic support conversation: six back-and-forth turns. Each turn sends the system prompt, the conversation so far, and the documents your retrieval system pulled in — around 1,500 input tokens per turn, so roughly 9,000 across the conversation. The bot's replies come to about 1,200 output tokens in total.

Two things worth taking from that chart.

First, the spread is roughly 19x between the cheapest and most expensive tier for identical work. Model choice is not a detail; it is the single biggest lever on your running costs.

Second, and less obvious: around 60% of the cost is input, not output. People assume they are paying for the bot's answers. They are mostly paying to re-send retrieved documents and conversation history on every single turn. That is why trimming what you retrieve, and how much history you carry, cuts bills faster than shortening replies.

At 10,000 conversations a month, that works out at roughly $40 on a budget model, $300 on mid-tier, and $750 on a frontier model. At 100,000 conversations, multiply by ten.

The full year-one budget

Build cost plus token cost still is not the whole picture. Here is a complete first year for a realistic project: a RAG support chatbot handling 10,000 conversations a month.

Three line items catch people out.

Content preparation. A RAG chatbot is only as good as what it reads. If your help centre is out of date, contradicts itself, or lives in PDFs, someone has to fix that before the bot is useful. This is unglamorous work and it is almost never in the original quote.

Monitoring and evaluation. You need to know when the bot is confidently wrong. Without a review process, the first person to discover a bad answer is a customer.

Maintenance. Budget 15% to 25% of the build annually. Models get deprecated, your products change, and prompts that worked in January drift by June.

Note the shape of it: the build is about two thirds of year one, but under half of a three-year total. Year two onwards, at the same volume, runs closer to $18,000.

Buy or build? The crossover point

You do not have to build anything. Per-resolution platforms will sell you a working AI support agent today, and one of the better-known ones charges around $0.99 per resolution, on top of seat fees.

That is excellent value at low volume and expensive at high volume. Here is where the lines cross.

Below roughly 4,600 conversations a month, buy. You will be live in days, you carry no build risk, and the maths does not favour a custom project. Most small and mid-sized businesses sit comfortably in this zone and should stop reading here.

Above it, the per-resolution model starts working against you, because your costs scale linearly with success while a custom build's costs barely move. At 20,000 conversations a month the gap in this model is around $7,000 a month, which pays back a $45,000 build inside a year.

Two caveats on that, because the chart is a model and not a promise. Change the resolution rate, the model tier or the build cost and the crossover moves substantially. And the comparison ignores the things you are actually buying from a SaaS platform: it works on day one, someone else maintains it, and there is no project to go wrong.

What moves your number up or down

The accuracy one deserves expanding, because it is the most commonly underestimated. Getting a chatbot to answer 80% of questions correctly is relatively quick. Getting from 80% to 95% costs more than the first 80% did, because the remaining questions are the ambiguous, multi-step and edge-case ones.

That matters commercially. If your bot handles billing or medical or legal queries, 95% may be the floor. If it answers "what are your opening hours," 80% with a clean handoff is fine and far cheaper. Decide which you are building before you agree a budget.

How to get a quote you can actually compare

Agencies quote wildly different numbers for the same request, usually because they are each imagining a different project. Answer these before you ask, and the quotes become comparable:

• What is the one job this bot must do? Name a single primary use case.
• What will it answer from — how many documents, and how current are they?
• Which systems must it read from or write to, by name?
• Which channels: web, WhatsApp, in-app, voice?
• Roughly how many conversations a month do you expect?
• What happens when it cannot answer? Who picks up, and how?
• Any compliance constraints on where the data can live?

Also ask every agency for their estimate of your monthly running cost. If they cannot produce a number, they have not thought about the part of the project that never stops costing you money.

Getting a real number for your project

Duple IT Solutions builds AI chatbots and automation for clients across the US, UK, Canada and Australia. We start with a short scoping call that produces both figures: what it costs to build, and what it will cost per month to run at your volume.

That includes telling you when an off-the-shelf platform is the better answer. Under a few thousand conversations a month, it usually is, and we would rather say so than sell you a build you will regret.

Get a costed chatbot scope →