Intercom's Fin charges roughly ninety-nine cents every time it resolves a customer conversation. Sit with that for a second, because it quietly reframes the question everyone actually asks — "how much do chatbots cost?" — from a number you pay once into a tax you pay forever, per answer, for as long as the thing is switched on. The person asking that question wants a single figure they can put in a budget cell. The honest answer is that the figure they want is the one that matters least.
This is the piece I wish more buyers read before they sign anything. I run a boutique engineering studio with a Python and AI-ML practice that has been shipping production LLM systems for the last couple of years, and the most expensive mistake I watch people make is not overpaying for a build. It's mistaking the build price for the cost of the chatbot. They are not the same number. They are not even the same kind of number.
The conventional answer, and why all three tiers of it are misleading
Ask the internet how much a chatbot costs and you get a tidy three-tier menu. Tier one: free, or nearly — a no-code widget you bolt onto a site in an afternoon. Tier two: a SaaS subscription, somewhere between $50 and $500 a month depending on seats and message caps. Tier three: a custom build, which agencies will quote anywhere from $15,000 for something simple to well past $150,000 for a genuine AI assistant wired into your systems. All three numbers are real. All three are also the part of the iceberg above the waterline.
The reason this menu misleads is structural, not dishonest. Every tier quotes you the cost of creating the chatbot and stays conveniently quiet about the cost of operating it. A car salesman who quoted only the sticker and never mentioned fuel, insurance, or the mechanic would be committing fraud. In chatbots we call it a pricing page.
The build is a rounding error. The meter is the product.
Here is the part the quote hides. A modern chatbot is a thin wrapper around a large language model you rent by the token, and that rental is metered on the way in and on the way out. As of this writing a frontier-class model runs on the order of a few dollars per million input tokens and $10–$15 per million output tokens; the budget tier sits under a dollar. Those sound like trivial sums until you internalize two facts: output tokens — the part the bot writes — cost three to five times more than input, and a conversational bot re-sends the entire chat history on every single turn so it can "remember" what was said. A ten-message conversation doesn't cost ten messages' worth of tokens. It costs something closer to the sum of one, then two, then three message-loads stacked up. Context is not free, and it compounds.
Bolt on the things that make a chatbot actually useful and the meter spins faster. Retrieval-augmented generation — the standard way to make a bot answer from your own documents instead of making things up — stuffs retrieved passages into the prompt on every query, which is more input tokens every time. A bot that searches your knowledge base, calls a tool, and reasons over the result might burn five or ten model calls to produce one reply. The per-message cost that looked like a fraction of a cent in the demo is now a real number multiplied by every conversation you will ever have. That is why per-resolution pricing exists: the vendors who meter you that way have simply done the arithmetic out loud.
The cost that appears on no invoice: being wrong
Inference is the cost you can at least forecast. The one that wrecks budgets is the cost of the bot being confidently incorrect. A chatbot that quotes a refund policy that doesn't exist, promises a discount you never authorized, or hands a healthcare user wrong information is not a software bug — it's a liability, and liabilities don't show up on the AI vendor's pricing page. The work of preventing that is the actual engineering project, and it is permanent. Evaluation harnesses, guardrails, a human-fallback path for when the bot is unsure, observability tooling like Langfuse or Phoenix to catch quality drift before a customer does — none of that is optional on a system that talks to real people, and all of it is recurring engineering time, not a one-off.
This is where I'll plant a flag from our own bench: on our AI builds every prompt change and pipeline change goes through review by another senior engineer before it ships, the same as application code, precisely because an unreviewed tweak to a system prompt can silently degrade thousands of answers and you won't see it on a dashboard until the complaints arrive. People price the model and forget that a probabilistic system that improvises in front of customers needs a correctness budget bigger than its inference budget. The bot is cheap to run and expensive to trust.
This liability is not hypothetical, and the case law is already here. In 2024 a Canadian tribunal ordered Air Canada to honour a bereavement-fare discount its own support chatbot had simply invented; the airline argued the bot was a separate entity responsible for its own statements, and that argument lost. A car dealership's GPT-powered bot was talked into agreeing to sell a new SUV for one dollar, cheerfully adding "no takesies backsies." A delivery firm's bot was coaxed into swearing at a customer and composing a poem about how bad the company was. None of these were build-cost problems. All of them were owning-cost problems: the price of shipping a confident system that can be wrong, manipulated, or both, with your brand on the speech bubble.
So when someone asks how much does a chatbot cost, the honest line item to add is the one no vendor quotes: what it costs you the first time the bot is wrong at three in the morning with your name on it. Price that, and the build quote stops looking like the number that matters.
Integration is the iceberg, not the model
A chatbot that can chat is a toy. A chatbot that can check an order status, issue a refund, look up a patient record, or escalate to a human with full context is a system — and the cost lives almost entirely in that second clause. The language model is a commodity you can swap in an afternoon. Wiring it safely into your CRM, your billing, your auth, your permissions, and your audit trail is where the weeks go. When we scope an AI assistant, the model choice is one of the smallest decisions; the integration surface and the failure paths are where the real estimate comes from. Anyone quoting you a chatbot price without asking what it has to connect to is quoting a demo, not a deliverable.
And then it rots
Software you wrote in 2015 still runs in 2026. A chatbot built on a model that gets deprecated will simply stop working on a date the vendor chooses, and the replacement model will behave differently enough that prompts tuned for the old one quietly get worse. Models deprecate, prices reprice overnight, and the prompt that was perfect six months ago drifts. Maintenance on an LLM product is not the gentle bit-rot of normal software — it's an active subscription to someone else's release schedule. Budget for a quarterly tune-up or budget for a bot that degrades while you're not looking.
So what does a chatbot actually cost?
Here are numbers I'll defend, stated as run-rates rather than stickers. A genuinely simple FAQ widget on a SaaS plan: $50–$500 a month, and for many businesses that is the correct, unembarrassing answer — don't custom-build what a subscription solves. A custom retrieval-grounded assistant wired into your systems: a build in the tens of thousands to low six figures depending entirely on how many systems it touches and how badly a wrong answer hurts you, plus an ongoing bill that splits roughly into three streams — inference (driven by traffic volume and how much context each answer drags along), human fallback and review (driven by your tolerance for the bot being wrong), and maintenance (driven by how often the underlying model changes underneath you). The single biggest swing factor in all of it is volume: a bot answering a hundred conversations a month and one answering a hundred thousand are different financial animals wearing the same UI.
The four questions that actually move the number
If you want a real estimate instead of a price band, answer four questions, because these are the variables that swing the figure by an order of magnitude. One: does the bot just retrieve and answer, or does it take actions in your systems? Actions are where the cost lives. Two: how good does retrieval have to be, a flat FAQ or grounded RAG over your live, messy, contradictory documents? Three: what is the blast radius if it's wrong, and therefore how much evaluation, guardrailing, and human-in-the-loop does it need? A bot answering store hours needs little; a bot quoting policy or money needs a lot. Four: what is the volume, because that is your inference bill. A team that quotes you without asking these is quoting the demo, not the system.
The strategy-versus-shipping split runs through all our AI writing; what AI consulting really is states it plainly. Our AI engineering services describe the shipping half.
Stop pricing the build. Underwrite the run.
The position is simple and I'll commit to it: "how much do chatbots cost" is the wrong question, and the tidy build quote that answers it is the cheapest, most knowable, least decision-relevant figure in the whole exercise. The number that decides whether your chatbot is a smart investment or a slow bleed is the cost per conversation multiplied by your real traffic, plus the standing cost of keeping it correct, plus the maintenance tax of renting someone else's model. Ask a vendor that and watch the polished answer evaporate. The cheap part of a chatbot is building it. Everything after "go live" is the price — and unlike the quote, it never stops arriving.