If you have started collecting quotes for a chatbot, you have probably noticed the numbers make no sense together. One vendor points you at a $50-a-month widget. Another proposes a six-figure build. Both call the thing an AI chatbot, and neither is necessarily lying — they are describing different products that happen to share a name.
The honest answer to "how much does an AI chatbot cost" starts with which of three tiers you actually need: an off-the-shelf widget you configure, a custom bot grounded in your own documents and data, or a deeply integrated assistant that can take real actions in your systems. Each tier has its own market price band, its own running costs, and its own failure modes.
This guide lays out all three, the ongoing costs that continue after launch, the factors that push a quote up, and the ticket-deflection math that tells you whether any of it pays for itself. One caveat up front: every figure below is a broad market observation, not a quote. Prices vary widely by region, team seniority, and — more than anything — scope.
One name, three very different products
Tier 1: The off-the-shelf widget
This tier covers subscription chatbot builders and the AI add-ons inside helpdesk products you may already pay for. You point them at your website or upload a folder of documents, adjust the tone, paste a script tag into your site, and you are live — often within days.
Market pricing is subscription-based: free or low-cost entry tiers, climbing to a few hundred dollars per month as volume and seats grow. Increasingly there is also a per-resolution fee — commonly in the range of a dollar or two for each conversation the AI resolves on its own. The build cost is close to zero; the real costs are your team's time curating the content the bot reads, and monthly fees that compound quietly as volume rises.
This is the right tier when your questions are genuinely generic — hours, shipping, returns, basic how-tos — when the bot never needs to touch your internal systems, and when you want something live this week rather than this quarter.
Tier 2: A custom bot grounded in your data
The second tier is a custom-built assistant that answers from your own content using retrieval-augmented generation (RAG): your documentation, policies, and product data are indexed; each incoming question fetches the most relevant passages; and the model is instructed to answer only from them — and to say plainly when they do not contain the answer. You control the model choice, the prompts, the interface, and where your data lives.
In the current market, builds at this tier typically land in the low to mid five figures and take somewhere between four and ten weeks. What you are paying for is not the AI — model access is cheap — but the retrieval pipeline, the evaluation set that catches regressions, the escalation paths, and the unglamorous engineering that makes answers trustworthy enough to put in front of customers. The same architecture, taken to full product scale, sits behind Refine AI, a compliance platform we built where a RAG-powered service ingests company policies and auto-drafts security-questionnaire answers with confidence scores and human-review flags. It is a useful reminder that "reliable answers from your own documents" is a capability worth considerably more than support chat alone.
Tier 3: Deep integration — a bot that does things
The third tier is where the assistant stops just answering and starts acting: checking an order status against your database, booking or rescheduling an appointment, starting a return, opening a support ticket with the details pre-filled. Technically this means tool calling — you define functions against your systems, and the model decides when to invoke them — plus authentication before anything personal is shown, explicit confirmation before anything is written, and a clean human handoff with the transcript attached.
At this point you are no longer buying a chatbot; you are commissioning an AI integration and automation project, and a meaningful share of the budget goes into API development and integration work, because the assistant is only ever as capable — and as safe — as the endpoints behind it. Market pricing reflects that: a focused first phase with one or two actions commonly sits in the mid five figures, while assistants that act across several systems, channels, and languages run into the low six figures and beyond, almost always delivered in phases.
The running costs that outlast the build
Chatbot budgets fail more often on the second year than the first. Four line items continue for as long as the bot is live:
- LLM API usage. Model calls are metered per token. A typical grounded support conversation costs a few cents on a mid-tier model, which means monthly spend ranging from tens of dollars at low volume to a few thousand at tens of thousands of conversations. Model choice is a tenfold lever here — routing routine questions to a cheap, fast model and reserving the expensive one for hard cases is one of the highest-return optimizations available.
- Monitoring and evaluation. A test set of 50–100 real customer questions with known-good answers, rerun whenever the prompt, model, or content changes, plus a weekly transcript review. This is hours of someone's time every month. Budget it, or quality will decay silently.
- Content upkeep. A retrieval bot is exactly as accurate as the pages it reads. When your pricing changes and the docs do not, the bot now delivers the stale answer at scale. Someone must own the knowledge base, and that ownership is a running cost.
- Maintenance. Model providers deprecate versions, which forces migrations and full re-testing; dependencies need patching like any other software. The common rule of thumb of 15–25% of the build cost per year for maintenance applies to chatbots as much as to any custom application.
What pushes the price up
Within any tier, four factors explain most of the spread between the low quote and the high one:
- Integrations. Every system the assistant touches adds not just the happy path but authentication, error handling, retries, and sandbox testing. A modern, well-documented API integrates in days; a legacy system with no sandbox can consume weeks by itself.
- Human handoff done properly. "Talk to a human" sounds like one button. Done well, it means routing by topic and business hours, attaching the full transcript so the customer never repeats themselves, and falling back to ticket capture with a realistic response-time promise when nobody is available. That is real workflow engineering.
- Multi-language support. Modern models handle languages fluently, but fluency is the easy part. Your knowledge base, your evaluation set, and your escalation team all need coverage in each language you promise — the cost multiplies per language, not the model fee.
- Compliance and data handling. GDPR obligations, health or financial data, data-residency requirements, and audit logging shape the architecture itself. They can dictate which model providers you may use, where processing happens, and how transcripts are stored and redacted. As with any software project, compliance designed in from the start costs a fraction of compliance retrofitted later.
Price a chatbot by the number of systems it has to touch, not the number of questions it has to answer. Answering is cheap; safely taking actions against your CRM, your order system, and your calendar is what sets the budget.
The ROI math: ticket deflection
The clearest way to justify chatbot spend is deflection — conversations resolved without a human. Industry benchmarks commonly put the fully loaded cost of a human-handled support ticket somewhere between a few dollars and fifteen dollars or more for straightforward consumer queries, and considerably higher in B2B where responses take research.
Run the arithmetic on your own numbers. As a purely illustrative example: a business handling 2,000 tickets a month, deflecting 35% of them at a cost of $7 per ticket, saves around $4,900 a month — roughly $59,000 a year. Against a tier-2 build in the low-to-mid five figures plus modest running costs, that pays back within the first year. At 200 tickets a month, the same percentages produce a tenth of the savings, and an off-the-shelf widget is almost certainly the better call.
Two honesty rules keep the math credible. Define deflection strictly — the customer did not escalate and did not email support about the same issue a day later. And count the value beyond deflection without inflating it: around-the-clock coverage and pre-qualified sales enquiries are real benefits, but they are harder to price, so treat them as upside rather than the justification. Our guide to building a website chatbot that actually helps covers how to measure all of this without fooling yourself.
Buy or build?
Buy off the shelf when your questions are generic, your volume is modest, the bot never needs to act inside your systems, and you already pay for a helpdesk whose AI add-on covers the job. At low volume, no custom build beats a subscription on cost, and speed to live is measured in days.
Build custom when any of the following is true: the assistant must take actions against your internal systems; your data is proprietary, structured, or changes constantly; compliance or data residency constrains where processing can happen; per-resolution fees at your volume exceed what a custom bot would cost to run; or conversational AI is part of your product itself rather than a support channel beside it.
There is also a pragmatic middle path we often recommend: run an off-the-shelf widget for a quarter first. The subscription fee buys you the real distribution of customer questions, the topics where a generic bot fails, and a baseline deflection rate — which together turn a custom build from guesswork into a scoped project. That is some of the cheapest product research you can do.
Getting a number you can plan around
Generic price bands can only take you so far, because the honest answer always depends on your ticket volume, your systems, and your constraints. If you want a real figure, write a one-page brief: the ten to twenty question types pulled from your actual support inbox, every system the assistant must read from or write to, the languages you need, and any compliance requirements. Send it through our contact page and we will come back with an itemized scope and quote — including which tier you actually need, and whether a subscription product would serve you better than a build — within 2–3 business days.