How Much Does an AI Chatbot Cost? A Full Breakdown

An AI chatbot costs a one-time setup amount plus monthly running costs, and for most businesses the model or API fees are only a small part of the total. The bigger items are usually preparing the content the bot answers from, integrating it with your website and helpdesk, and the ongoing staff time for maintenance and reviewing conversations. The useful figure is not the monthly fee but the cost per resolved conversation, compared with what the same contact costs when a person handles it.

This guide breaks down each cost item, shows a complete worked example with hypothetical numbers and explains which assumptions move the result most. To calculate the model usage for your own case, open the LLM API cost calculator in a second tab.

The formula behind every chatbot budget

First-year cost = setup cost + 12 × monthly running costs

Cost per resolved conversation = monthly running costs ÷ conversations resolved without a person

The second figure is the one to compare. A chatbot that is cheap per month but resolves almost nothing is expensive per result.

Three ways to get a chatbot

How you build the chatbot decides how you are billed.

Route	What it means	Typical billing model	Good fit
AI feature in your helpdesk or website software	Switch on the AI assistant your existing tool offers	Add-on per seat, per conversation or per resolution	You already use the tool and the standard features cover your needs
Chatbot platform	A dedicated product you configure with your content	Monthly plan with conversation or message limits, plus overage	You want a bot quickly, with little development
Custom build on an LLM API	Your own application calling a model, usually with retrieval from your documents	Tokens per request, plus hosting and development	High volume, special integrations or strict control over data flow

Pricing models and prices change often, so compare current offers directly with vendors. What matters for your budget is how the bill scales: per seat, per conversation, per resolution or per token. A per-resolution model can look attractive until you check how the vendor defines a resolution.

One-time setup costs

Setup is where most of the real work happens, regardless of the route:

• Content preparation: collecting, updating and structuring the help articles, FAQs and policies the bot answers from. Outdated content produces outdated answers.
• Configuration and integration: connecting the bot to your website, helpdesk, CRM or order system, and setting up handover to a person.
• Instructions and guardrails: writing the bot's system prompt, scope and escalation rules. Our guide on how to write a system prompt shows a complete example for customer support.
• Testing: building a set of real questions, including awkward and out-of-scope ones, and checking the answers before launch.
• Privacy and legal review: what personal data the bot processes, where it is stored, and what you tell users.
• Team briefing: how agents take over conversations and report wrong answers.

Internal hours count as cost even when no invoice is paid.

Monthly running costs

Cost item	What it covers
Model usage or platform fees	Tokens via an API, or the chatbot plan and any overage
Hosting and infrastructure	Servers, vector database, logging, for custom builds
Content maintenance	Updating answers when products, prices or policies change
Monitoring and review	Reading a sample of conversations, fixing failure patterns
Technical maintenance	Integration fixes, model updates, re-testing prompts
Escalations	The human time for conversations the bot hands over

The last line is often left out. Escalated conversations still cost staff time, and a bot that collects details before handing over can make them faster or, if it handles the handover badly, slower.

Estimating model usage

If you build on an API, the token bill depends less on the user's messages than on everything sent with them. Each reply from the bot is a separate request that includes:

• the system prompt with rules and tone,
• passages retrieved from your knowledge base,
• the conversation so far, which grows with every turn,
• the user's latest message.

A conversation with four questions therefore means four requests, each larger than the last. LLM API pricing explained covers how input, output and caching are billed, and what tokens are explains how to estimate them from text.

Worked example: a website support chatbot

A company plans a support chatbot built on an LLM API with retrieval from its help centre. All figures are assumptions for illustration. The model prices are placeholders, not those of any real provider.

Usage

Assumption	Value
Conversations per month	2,000
Model calls per conversation	4
Input tokens per call (system prompt 1,000 + retrieved text 1,500 + history 800 + message 50)	3,350
Output tokens per call	250
Example prices per million tokens (input / output)	1.00 / 4.00

Per conversation: 4 × 3,350 = 13,400 input tokens and 4 × 250 = 1,000 output tokens.

• Input: 13,400 × 1.00 ÷ 1,000,000 = 0.0134
• Output: 1,000 × 4.00 ÷ 1,000,000 = 0.0040
• Per conversation: 0.0174
• Per month: 2,000 × 0.0174 ≈ 35

Running costs per month (staff time at a fully loaded 45 per hour)

Item	Cost
Model usage	35
Hosting, vector database, logging	100
Content and technical maintenance, 4 hours	180
Reviewing a sample of conversations, 4 hours	180
Total	495

Setup cost

Item	Hours
Content preparation	30
Configuration and integration	40
Test set and testing	20
Privacy and legal review	8
Team briefing	6
Total internal	104 hours × 45 = 4,680

Add 3,000 for an external developer, and setup comes to 7,680.

Results

• First-year cost: 7,680 + 12 × 495 = 13,620
• Running cost per conversation: 495 ÷ 2,000 ≈ 0.25
• If half of all conversations are resolved without a person: 495 ÷ 1,000 ≈ 0.50 per resolved conversation

Two things stand out. First, the token bill is about 7% of the running costs in this example; people's time is the larger part. Second, the cost per resolved conversation doubles if the resolution rate halves, which is why that rate deserves more attention than the model price.

Is it worth it? Compare with the cost of a human contact

Suppose a person needs six minutes per chat or email at a fully loaded 45 per hour. That is 4.50 per contact.

• Optimistic case: 1,000 conversations per month are resolved by the bot and would otherwise have reached the team. That is 100 hours or 4,500 per month. Net of running costs: 4,005 per month, payback of setup in about two months.
• Cautious case: the bot resolves only 25% of conversations, and half of those people would never have contacted you anyway. That leaves 250 contacts actually avoided, 25 hours or 1,125 per month. Net: 630 per month, payback after about twelve months.

Same chatbot, same costs, very different business case. The difference comes entirely from two assumptions: the resolution rate and how many conversations would really have reached a person. Measure both in a pilot before you scale. The AI ROI calculator lets you test these assumptions with your own numbers, and how to calculate the ROI of AI automation explains the method.

What drives chatbot costs up

• Long conversations. Every turn resends the history. Summarising or trimming older turns helps.
• Large retrieved context. Sending ten long passages when two short ones would do multiplies input tokens. Good retrieval is a cost lever as well as a quality lever; see RAG vs. fine-tuning.
• Model choice. A larger model can cost several times more per request. Test whether a smaller one is good enough for routine questions.
• Integrations. Looking up order status or bookings needs development and maintenance, but often decides whether the bot can resolve anything at all.
• Several languages and channels. Each adds testing and content work.
• Poor content. Gaps in the knowledge base lead to escalations and wrong answers, which cost more than the bot itself.

For tactics on the token side, such as caching the system prompt or routing simple questions to a smaller model, see how to reduce LLM API costs.

Costs that are easy to forget

• Wrong answers. A bot that confidently gives a wrong refund policy creates follow-up contacts, goodwill gestures and sometimes legal exposure. Budget time for reviewing conversations, as described in our guide to human-in-the-loop AI.
• Transparency and data protection. Under the EU AI Act, transparency rules that apply from August 2026 require that people are informed when they interact with an AI system, unless it is obvious. The overview of deployer obligations explains who carries which duty. Data protection rules apply to the personal data in chat logs.
• Vendor changes. Plans, limits and model versions change. Re-test after every significant change.
• Exit costs. If you switch platforms, can you export your content, configuration and conversation logs?

A budget checklist

• Estimate conversations per month from your current chat, email and phone volume.
• Decide the route: helpdesk feature, platform or custom build.
• List setup items with hours, including content work and testing.
• Estimate model usage per conversation with the LLM API cost calculator, or get the vendor's price for your volume.
• Add hosting, maintenance, review time and escalations to the monthly costs.
• Calculate cost per conversation and per resolved conversation.
• Compare with the cost of a human contact, using a cautious resolution rate.
• Run a pilot and replace assumptions with measured figures before scaling.

Summary

The price tag of the software or the token bill is rarely what decides whether a chatbot pays off. Setup work, content quality, review time and above all the share of conversations the bot actually resolves matter more. Calculate the cost per resolved conversation, test it against a cautious scenario, and let a pilot decide.