AI chatbot for SMBs: how it really works (and what it costs)
"AI chatbot" has become a buzzword. Some sell a scripted dialog box for $5,000. Others sell real LLM-powered for $500/month. Here's the concrete difference, what a modern chatbot actually does for an SMB, and which marketing promises to ignore.
3 generations of chatbots — where are you?
Gen 1 — Scripted chatbot (2015-2020)
"Hello! Choose: 1) See our products 2) Book appointment". Rigid decision tree. User types anything else? Bot doesn't understand. Useless in 2026. Don't pay for it.
Gen 2 — Keyword-based chatbot (2020-2023)
Detects keywords ("price", "shipping") and returns pre-written answers. Better than gen 1 but blocks the moment a question goes off-script. Still sold by many agencies in 2026 — don't fall for it.
Gen 3 — LLM chatbot (2023+)
Built on GPT-4, Claude, or equivalent. Understands natural language. Responds like a human. Can be fed YOUR data (catalog, FAQ, prices, procedures) using a technique called RAG (Retrieval-Augmented Generation). That's what we're talking about today.
What a modern AI chatbot does concretely
- Answers FAQs 24/7 in natural language, in your tone, in FR or EN.
- Qualifies leads by asking the right questions (budget, timeline, specific need) before transferring you the contact.
- Books appointments by connecting to your Calendly or calendar.
- Captures emails and phones naturally, with Law 25 consent.
- Sends transcripts to your CRM automatically (HubSpot, GoHighLevel, Notion).
- Notifies you by SMS if a hot lead arrives (high budget, urgent).
What it does NOT do (honest limits)
- It doesn't replace a salesperson on complex B2B deals ($50k+). It qualifies, you close.
- It can hallucinate if misconfigured. Without strict data grounding it can invent prices or features. Done well, kept below 1%.
- It doesn't handle complex emotion (angry customer, sensitive situation). Must escalate to human.
- It depends on OpenAI/Anthropic/etc APIs which can have outages. Plan a fallback.
Agencies pitching a chatbot that "understands everything" and "never gets it wrong" are lying. A good seller tells you the limits and how to manage them.
What it really costs
Initial setup
Well-configured LLM chatbot for SMB: $2,000-$5,000 setup. Includes: conversational design, your data ingestion (site, PDFs, FAQ), site integration, CRM connection, testing, training.
Monthly recurring
- OpenAI/Claude API: $30-$200/month based on conversation volume. 500 conversations/month ≈ $50.
- Workflow hosting (self-hosted n8n/Make): $0-$50/month.
- Maintenance and improvement: $300-$800/month. Human reviewing conversations, fixing mistakes, improving responses.
Realistic total: $500-$1,500/month for an active SMB. More if high volume (1,000+ conversations).
How to tell if you need one
- How many repetitive questions per week? Email, phone, forms combined. Under 20/week? Probably overkill. Above? ROI is real.
- How many leads lost because you don't reply nights/weekends? If you've heard "I went with someone who replied faster", you're paying this cost without seeing it.
- Are your questions predictable? Prices, hours, services, availability — yes. Personalized medical advice, complex quotes — no.
Simplified ROI math
If chatbot brings you 2 extra clients/month you would've lost, and your margin per client is $500: $1,000/month additional revenue. Cost: $800/month. ROI positive month 1.
If 0 additional clients but 10 hours/week saved not answering same questions × $30/hr × 4 weeks = $1,200/month saved. ROI positive too.
Want to see an AI chatbot working on a real case?
30-minute call — I show you how it works live, price your specific case, tell you honestly if you need one.
Book a free call →3 pitfalls when shopping a chatbot
- Long contract without test period. Require minimum 30 days to validate quality before annual commitment.
- Hidden API fees. Ask precisely: who pays the OpenAI/Claude API? If included but capped ("up to 500 conversations"), ask what happens past that.
- No conversation access. You must be able to read conversations to understand customers and improve the bot. Otherwise you're blind.
To sum up
A well-built AI chatbot in 2026 is an employee that works 24/7, doesn't get sick, costs less than a part-timer. But it's a tool, not a magic wand. It doesn't save a site that doesn't convert. Doesn't invent prospects. It multiplies what already works.