AI Chatbot Cost for Small Businesses: Buy vs Build Breakdown

Most pricing pages for AI chatbots are written by the vendors selling them, and each frames the question around its own model. This guide separates the three ways to get a chatbot, explains the pricing units you'll run into, and ends with a worksheet you can fill in with your own numbers.

Prices here were checked on the vendors' own pages in September 2026. They change often, so check again before you sign anything.

One question before any of that: is a chatbot the fix you need? If your real problem is that the website isn't producing inquiries, a bot won't solve it. Start with our diagnostic for websites that aren't generating leads.

RAG, short for retrieval-augmented generation, means the assistant searches your documents for relevant passages and hands them to the model along with the customer's question. It's how a bot 'trained on your documents' usually works. It's cheaper and easier to keep current than fine-tuning a model, because updating an answer means editing a document.

Most small businesses should start with the first option, move to the second when the bot keeps saying it can't check something, and consider the third only when volume, data rules or channel needs justify it.

Intercom defines an outcome as a conversation where the customer confirms the issue is resolved, doesn't ask for more help after Fin responds, or Fin completes a defined workflow; its pricing page says that last case includes handoffs, and that you're charged once per conversation. Read every vendor's definition of 'resolved' before comparing prices.

Token pricing needs a worked example. Assume, for illustration, a six-message support conversation in which each turn sends about 3,000 input tokens (the question, the chat so far and retrieved help-center passages) and returns about 250 output tokens. That's 18,000 input and 1,500 output tokens per conversation. Using list prices from Anthropic and OpenAI as of September 2026:

Discovery defines what the bot should and shouldn't answer, which systems it can read, when it hands off to a person, and how success will be measured.

Knowledge-base preparation is unglamorous, and it decides quality. Policies, FAQs and product data need to be current, consistent and written so each passage makes sense on its own. Contradictory documents produce confident wrong answers.

Integrations are where cost varies most: reading order status from Shopify, checking availability in a booking tool, creating a ticket or CRM record, sending WhatsApp messages. Each needs authentication, error handling and testing, and often an automation layer; our guide to n8n automation costs covers that piece.

Testing means a written set of real customer questions, including awkward ones, run before launch and after every change, with answers checked by someone who knows the business.

As an illustrative planning estimate, not a quote: a website assistant answering from about 50 help-center pages, with handoff to a person, might take 40 to 80 hours across the five phases. Adding an order-status integration and a WhatsApp channel could roughly double that. At a hypothetical $50 an hour, 40 to 80 hours is $2,000 to $4,000. Ask bidders for hours per phase so their quotes can be compared.

The last line in the list below is the one to watch. Policies, prices and products change, and a bot answering from last quarter's return policy can cost more in refunds and complaints than it saves. Give the knowledge base an owner and a monthly review, and budget their time.

For WhatsApp, Meta's pricing page says that as of July 1, 2025 it charges per delivered template message by category, while non-template messages are free. Replies inside a conversation the customer started therefore carry no messaging fee; proactive templates do.

For Shopify stores, start with what's free. Shopify Inbox is listed as a free app in the Shopify App Store. Its instant answers show preset questions in the chat window, and a default Track my order answer uses the customer's order number and email to show order status. You can display up to 100 instant answers.

Shopify's help page on instant answers adds two caveats worth copying into any chatbot plan: AI-generated answer suggestions are available only in English, and you're responsible for the accuracy of content you publish, even when it's generated automatically.

Helpdesk tools suit stores managing email, social and chat volume together. As of September 2026, Gorgias lists its Starter helpdesk at $10 a month for 50 tickets, with AI Agent interactions priced separately and counted as tickets (Gorgias pricing). Chat apps such as Tidio sit in between.

Go custom when answers depend on data the apps can't reach: a 3PL's tracking events, made-to-order lead times, sizing that depends on a product's cut, or return rules that vary by product type. For sizing in particular, clear size charts and fit notes on the product page help the bot and the customer alike.

Many requests for a chatbot are really 'our team can't find answers in our own documents'. For that, a business AI plan is cheaper and faster than any build. As of September 2026, ChatGPT Business lists a standard seat at $20 a month billed annually or $25 monthly and includes creating and sharing GPTs within a workspace. Claude Team lists a standard seat at the same $20 billed annually or $25 monthly.

OpenAI's business pricing page says business data isn't used for training by default on that plan. Check the equivalent terms for any tool before your team uploads client files.

Keep customers out of staff tools. A public bot needs guardrails, handoff to people, conversation logs and a clear scope, none of which a staff workspace is designed to provide.

Wrong answers are the biggest hidden cost. A bot that invents a refund policy or a delivery date creates refunds, complaints and, in some cases, legal exposure. Limit it to approved sources, make it say 'I don't know' and hand off, and log conversations so someone can review them.

Data privacy: know what customer data the bot sees, where the vendor stores it, whether it's used for training, and how long logs are kept. Don't give the bot data it doesn't need.

Vendor lock-in: platform bots keep your content and conversation history in their own format. Keep your knowledge base in a source you control, such as your help center or a document store, and confirm you can export conversations.

For a structured approach, NIST's AI Risk Management Framework is free and voluntary. AI RMF 1.0 was released on January 26, 2023, built around four functions (Govern, Map, Measure and Manage), and NIST published a Generative AI Profile, NIST-AI-600-1, on July 26, 2024. A small business doesn't need the whole framework, but its questions make a good pre-launch checklist.

Compare the monthly total with what those conversations cost today in staff time, and with the cost of slow replies. If the bot only pays off at an optimistic resolution rate, start with the free or cheapest option and measure before committing to more.

If after-hours phone calls are the bigger gap, see our guide to AI receptionists for law firms. For what happens once a bot captures a lead, see how to automate lead follow-up. Pixel2Tech scopes and builds AI assistants and the integrations around them through our AI solutions service.