Chatbot pricing is confusing because “chatbot” spans everything from a $30/month website widget to a six-figure enterprise assistant. Here is the itemised version, so you can place your project on the map before anyone quotes you.
The three tiers, priced
Tier 1 — Off-the-shelf widget ($0–$500/month, no build cost). The bot bundled with your helpdesk or website platform, answering from a template plus whatever pages it crawled. Fine for FAQs and lead capture. Its ceiling arrives fast: no deep grounding in your data, no real actions, generic tone. If your queries are simple and volume is low, genuinely — start here. This is the “buy” square of custom vs off-the-shelf.
Tier 2 — Custom grounded assistant ($15k–$60k build). A bot built on retrieval over your documentation, policies and product data (the RAG approach), with guardrails, your tone, graceful human handoff, and a measured accuracy number before launch. Lives on your site, in your app, on WhatsApp or Slack. This tier is where “resolves real queries” starts, and it is the most common shape we build.
Tier 3 — Assistant that acts ($40k–$120k+). Everything in tier 2, plus system integrations: check the order, change the booking, issue the routine refund, file the ticket — with permissions, approval checkpoints and audit logs. The price tracks the number of systems touched and the oversight the actions demand, because at this point you are in agent territory.
What moves the number inside a tier
Knowledge sources. One clean help centre is cheap. Seven systems, PDFs and an internal wiki that contradicts itself: the pipeline work grows accordingly — often it is the project.
Integration count. Each system the bot reads from or writes to adds engineering and testing. The write-integrations cost more than the reads, because mistakes do.
Accuracy bar and stakes. A bot suggesting help articles can be 90% right. A bot quoting refund policy to angry customers needs evaluation, guardrails and escalation design — the difference is real engineering, and it is where cheap quotes quietly cut.
Channels. Web only is simplest; adding WhatsApp, Slack or helpdesk integration adds moderately per channel.
The monthly bill nobody quotes
Running costs are modest but real, and you should see them itemised in any serious proposal:
- Inference — for most support bots at moderate volume, tens to a few hundred dollars per month. Model prices keep falling; usage keeps rising; budget flat.
- Monitoring and evaluation — dashboards and periodic eval runs; small if built in from day one (and suspicious if absent).
- Knowledge upkeep — someone updates the docs the bot answers from. Free, except organisationally.
- Improvement cycles — reviewing bad answers and shipping fixes; a few hours a month that compound into the bot people stop complaining about. A common planning figure across the system: 15–20% of build cost per year.
The ROI arithmetic (do this before quoting anyone)
Count your monthly conversations, the share that are routine, and your cost per human-handled ticket. A support team handling 2,000 tickets/month at even $4–8 each, with 40–60% routine, is spending $3k–$10k monthly on questions a grounded bot resolves. Against a tier-2 build, payback lands inside a year in most of the cases we scope — when the routine share is honestly measured first. If your volume is 60 tickets a month, buy the widget and spend your budget elsewhere; we will tell you the same on a call.
What to ask any vendor (including us)
What is the measured accuracy on our real queries before launch? What happens on a wrong answer — show the guardrail and the escalation path. What does month twelve cost, itemised? And who owns the pipeline, prompts and evals afterwards? (Right answer: you. Full checklist: how to choose an AI development company.)
Want a number for your actual case instead of ranges? Tell us your ticket volume and knowledge sources — we reply within a business day with an honest estimate, including “buy the widget” when that is the truth.