AI chatbot quotes swing from $50 to $150,000 - here is why

Ask three vendors what an AI chatbot costs and you can get $49 a month, $0.99 a conversation, and $120,000 - for what all three describe as "an AI chatbot for customer support." None of them is lying. They are quoting two fundamentally different products under one name: a SaaS agent you rent, and a custom system you own. Rent scales with how much you use it, forever. Ownership is a big number once, then a small one. Get the two models straight and the price stops looking random - it becomes a break-even you can actually calculate. Here is what each path costs in 2026, the running costs both like to leave implicit, and the volume at which the answer flips from buy to build.

Two ways to get an AI chatbot - and why their price tags look nothing alike

Every AI chatbot quote sits in one of two camps, and comparing across them is where budgets go wrong:

• Buy (SaaS): you subscribe to a ready-made agent - Intercom Fin, Zendesk AI, Ada, Salesforce Agentforce. You pay a monthly platform or seat fee and, increasingly, a charge for every question it answers. It is live in days, but your bill grows with every conversation, every month, with no end.
• Build (custom): you commission a chatbot built on models you call through an API - typically AWS Bedrock or Amazon Lex - trained on your own content and wired into your own systems. It is a one-time project, then a small monthly infrastructure bill. You own it, and your per-chat cost barely moves as volume climbs.

The reason the numbers look unrelated is that they scale on different variables. SaaS scales with usage. A build scales with complexity - how many systems it touches and how messy your knowledge is - and then largely stops. A brand handling 500 chats a month and one handling 500,000 pay almost the same to build, and wildly different amounts to rent. That single fact decides most of this choice. If you are still weighing whether you even need the newer agent architecture, our take on AI support agents versus traditional chatbots draws the line.

Option 1: buy a SaaS chatbot (what you actually pay)

SaaS is the fast path - live in days, no engineering. The pricing has moved from flat seats to usage-based "per resolution" billing, which is fairer at low volume and brutal at high volume. Here are the 2026 rates:

One line in a SaaS contract matters more than the headline rate: per resolution versus per conversation. Per resolution means you pay only when the bot fully answers a customer. Per conversation means you pay for every chat it touches - including the 30-40 percent it fails and hands to a human. At a 60 percent resolution rate, per-conversation pricing quietly costs you about 40 percent more than the sticker suggests. Ask which model you are on before you sign.

Worked SaaS math. A support team resolving 5,000 queries a month on a per-resolution agent at $0.99 pays about $4,950 a month, or roughly $59,000 a year - and it climbs every time your volume does. Double your customers and you double the bill. That is the trade for zero upfront cost and a same-week launch. The proof it can pay off is real: we documented a 30 percent support-cost reduction with a chatbot, but that math only holds when deflected volume outweighs the per-chat fee.

Option 2: build a custom chatbot (development cost by type)

Building means paying a team once to create a bot on your own stack, usually AWS Bedrock for the language model and Amazon Lex or a custom orchestration layer for the conversation flow. The one-time cost tracks the type of bot, not your traffic:

The single biggest cost driver on a build is not the model - API models are cheap to call. It is the work of turning your messy business information into a reliable retrieval system: extracting documents, cleaning duplicates, designing how content is chunked and tagged, building the ingestion pipeline, tuning retrieval so answers are actually grounded in your sources, and handling permissions. This is why rule-based bots are dying - they cannot do this, and customers can tell. We argue that case in why rule-based chatbots are dead in 2026. A build only makes sense if you invest in the retrieval layer, not just the chat window.

The running cost of a build (the number both sides forget)

"Build" is not free after launch - but the monthly cost is far smaller than most people expect, and it is where the ownership model earns its keep. On AWS, the running bill has three parts:

• Conversation orchestration - Amazon Lex, $0.00075 per text request. A store handling 10,000 customer interactions a month pays about $7.50 in pure Lex cost. This line is almost never your problem.
• The language model - Amazon Bedrock, per token. Claude Haiku 4.5 on Bedrock is about $1 per million input tokens and $5 per million output tokens ($0.001 and $0.005 per 1,000). A busy RAG chatbot fielding 10,000 questions a day lands around $1,400-2,000 a month in model calls - and prompt caching and batch discounts cut that further.
• Supporting infrastructure - a few hundred dollars a month. A vector database for retrieval, embeddings, hosting, logging, and monitoring. Small, but real - budget it so it is not a surprise.

Add it up and a mid-volume custom chatbot runs on the order of $2,000-4,000 a month, flat, regardless of whether conversations climb 20 percent. Contrast that with SaaS, where a 20 percent volume rise adds 20 percent to the bill. If you want the architecture behind those numbers, we lay it out in AWS Lex for e-commerce chatbots, and our AI-on-AWS practice is where we build and run them.

Build vs buy: the break-even math

Strip away the brand names and the choice is arithmetic. Put your monthly resolution volume against both models over three years:

• Low volume (2,000 resolutions/mo): SaaS at $0.99 is about $2,000/mo = $24,000/yr, live this week. A $75,000 build plus ~$1,500/mo running would take years to pay back. Buy.
• High volume (50,000 resolutions/mo): SaaS at $0.99 is about $49,500/mo = roughly $594,000/yr. A $100,000 build plus ~$3,000/mo running is about $136,000 in year one and ~$36,000/yr after. The build pays for itself in under three months and saves over half a million a year thereafter. Build.
• The crossover: somewhere around 5,000-10,000 resolutions a month, viewed over a multi-year horizon, the lines cross. Below it, renting is cheaper and faster; above it, owning wins because your cost stops scaling with success.

Two honest caveats keep this from being purely mechanical. Buying gets you live in days and needs no engineering team, which has real value when speed matters or volume is genuinely small. And a build is only cheaper if it is built well - a poorly scoped RAG bot that hallucinates costs you more in lost trust than any SaaS fee. The right move is to run your numbers, not a vendor's example. If you would rather have an agent that takes actions than just answers, that is a different and larger decision - our AI agent development team handles that end.

Frequently asked questions

How much does an AI chatbot cost in 2026?

It depends entirely on rent versus own. A SaaS chatbot runs $29-99 a month at the SMB end, or a per-resolution fee of roughly $0.99-2.00 per answered query at scale. A custom build is a one-time $5,000-15,000 for a rule-based FAQ bot, $30,000-120,000 for an LLM agent, or $75,000-120,000 for a production RAG chatbot with your knowledge base and CRM - then a few thousand dollars a month to run. Model both against your real conversation volume before deciding.

Is it cheaper to build or buy an AI chatbot?

Buy is cheaper below a few thousand resolutions a month - low upfront cost, live in days. Build is cheaper above roughly 5,000-10,000 resolutions a month over a multi-year horizon, because a one-time development cost plus a flat monthly AWS bill eventually beats a per-chat fee that grows with every customer. The exact crossover depends on your volume and how complex the bot must be.

What does it cost to run a chatbot on AWS?

Small. Amazon Lex charges $0.00075 per text request - about $7.50 for 10,000 interactions. The model calls on Amazon Bedrock (Claude Haiku 4.5 at roughly $1 per million input and $5 per million output tokens) are the main line: a RAG bot answering 10,000 questions a day costs about $1,400-2,000 a month. Add a few hundred dollars for the vector database, hosting, and monitoring. Prompt caching and batch pricing reduce it further.

Why is per-conversation pricing worse than per-resolution?

Per-conversation billing charges for every chat the bot touches, including the ones it cannot solve and escalates to a human. Per-resolution billing charges only when the bot fully answers the customer. If your bot resolves 60 percent of chats, per-conversation pricing effectively costs about 40 percent more than the same rate would on a per-resolution basis. Always confirm which model a SaaS vendor uses.

The short version: "AI chatbot cost" is not one number because it is not one product. You are choosing between renting a SaaS agent whose bill scales with your success, and owning a custom build whose cost is big once and small forever. Below a few thousand resolutions a month, rent. Above five to ten thousand, over a few years, own. If you want that modelled against your real volume - the SaaS bill you would pay versus a fixed-scope build and its monthly running cost - our AI chatbot team will run both numbers with you, and our AI development services is where the build happens.