AI chatbot development cost in 2026
Chatbot quotes range from a few thousand to six figures for the same brief. The four tiers, the ongoing costs nobody mentions, and how to estimate yours.
"How much for an AI chatbot?" has become as unanswerable as "how much for a website?" — for the same reason. The word covers four genuinely different products.
The four tiers
Rule-based bot — $3,000–$10,000. Decision trees, buttons, scripted paths. No language model. Handles known questions reliably and everything else badly. Still the right answer for narrow, high-volume flows like order status.
Retrieval-based assistant — $10,000–$40,000. An LLM grounded in your content: docs, policies, product data. Answers in natural language, cites real sources, declines gracefully outside its scope. This is what most businesses actually want when they say "AI chatbot", and it is where we do most of our work.
Multi-agent conversational AI — $50,000–$150,000. The assistant takes actions, not just answers: checks inventory, books the appointment, raises the ticket, processes the return. Every action is an integration with its own permissions, error handling, and audit trail.
Enterprise omnichannel — $100,000+. The above across web, app, WhatsApp, voice, and email, with SSO, role-based access, compliance logging, and human handoff into an existing contact-centre stack.
Where the money goes
Rarely on the model. Typically:
- Content and data preparation, 20–30%. Your knowledge is spread across a website, a PDF folder, a wiki, and three people's heads. Getting it into a form a model can retrieve reliably is the single most underestimated line item.
- Integration work, 20–40%. Any bot that does something is really an integration project wearing a chat interface.
- Guardrails and evaluation, 15–25%. Scope enforcement, refusal behaviour, prompt-injection resistance, and a test suite that catches regressions when you change a prompt.
- Conversation design, 10–15%. What it says when it does not know, how it escalates, what it never does.
- The interface, 10–15%. Usually the cheapest part, and the part clients imagine is most of it.
The ongoing costs
Build cost is not total cost.
| Line | Small business | Mid-market | Enterprise |
|---|---|---|---|
| Model/API usage | $30–$150/mo | $500–$1,500/mo | $3,000–$15,000+/mo |
| Hosting & infra | $20–$100/mo | $200–$800/mo | $1,000+/mo |
| Maintenance | 15–20% of build/yr | 15–25%/yr | 20–30%/yr |
Usage cost scales with conversation volume, context length, and which model you pick. Long system prompts sent on every message are the most common source of a surprising bill — and the easiest to fix with caching.
How to estimate yours
Answer four questions:
- Does it answer, or does it act? Acting multiplies cost, because each action is an integration.
- How many sources of truth? One clean docs site is cheap. Twelve inconsistent PDFs are not.
- How many channels? Each channel has its own delivery quirks, formatting, and testing.
- What is the cost of a wrong answer? A marketing site FAQ and a regulated financial assistant need very different guardrail budgets — and that difference is real engineering, not paperwork.
Build versus buy
Before commissioning anything custom, price the alternative honestly.
| Option | Typical cost | Where it fits |
|---|---|---|
| Off-the-shelf widget (Intercom, Tidio, etc.) | $50 – $500/mo | Generic FAQ answering, fast to launch |
| Platform + your content (SiteGPT-style) | $100 – $1,000/mo | Grounded answers, limited customisation |
| Custom retrieval assistant | $10,000 – $40,000 build | Your data, your rules, your UI, your integrations |
| Custom agent | $50,000+ build | It takes actions, not just answers |
Buy when your questions are generic and your content is already tidy. Build when the assistant needs your systems, your tone, your guardrails, or data that cannot go through a third party.
The honest test: if a platform trial answers 80% of your top twenty questions acceptably, the remaining 20% rarely justifies a custom build on its own.
A worked example
A B2B software company, 1,400 support tickets a month, fully loaded cost of $6 per ticket.
Repetitive tickets (est. 45%) 630/mo
Realistic containment rate 55%
Tickets deflected 347/mo
Monthly saving (347 x $6) $2,082
Model + hosting cost -$320
Net monthly saving $1,762
Build cost (retrieval assistant) $22,000
Payback 12.5 months
That is a reasonable, unexciting result — and it is roughly what an honest projection looks like. Beware of anyone modelling 90% containment in month one.
Two levers change this materially: content quality (a tidy docs site pushes containment up) and ticket cost (support-heavy industries pay back much faster).
The costs nobody quotes
- Content cleanup. If your policies contradict each other, someone has to resolve that. It is often the most valuable part of the project and it is rarely in the quote.
- Evaluation set maintenance. The test suite that catches regressions needs updating as your product changes.
- Human handoff integration. Connecting to your helpdesk so escalations arrive with context is real work.
- Ongoing conversation review. Someone should read the failures monthly. Budget an hour or two.
- Model version upgrades. Providers deprecate models. Staying model-agnostic makes this a config change instead of a project.
How to keep the running cost down
Usage bills are usually fixable, not fixed:
- Cache your system prompt. Long instructions resent on every message are the most common source of a surprising bill.
- Retrieve less, better. Five well-chosen passages beat twenty mediocre ones on both cost and accuracy.
- Route by difficulty. A small model handles simple questions; escalate only when needed. This often halves spend.
- Cap conversation length. Long histories resent in full grow cost quadratically.
Common questions
How long does it take to build? A retrieval assistant is typically 4–8 weeks including content preparation. Agents run longer, because each action is an integration.
Can we use our existing documentation as-is? Usually partly. Well-structured docs work well. PDFs, slide decks, and tribal knowledge need preparation first, and that preparation is a real line item.
What about data privacy? Enterprise API tiers from major providers offer zero-retention and regional processing. Get it in the contract. If data genuinely cannot leave your infrastructure, self-hosted open models are viable but raise both build and running cost.
Will it replace our support team? No. It removes the repetitive half so your team handles the cases that need judgement. Deployments sold as headcount replacement tend to disappoint on both sides.
What if it gives a wrong answer? That is a design question, not an inevitability — scope, grounding, and refusal behaviour decide it. Budget for guardrails rather than hoping.
The ROI test
Conversational AI is one of the few AI investments with a straightforward payback calculation. Take your monthly support ticket volume, estimate the share that is repetitive, and multiply by your cost per ticket. If automating half of that does not cover the build within a year, you are solving the wrong problem — and we will say so.
Send us your ticket volume and your top twenty repeated questions. That is enough for a real number, and the consultation is free.
