A chatbot on Air Canada's website once told a grieving customer he could book a flight at full price and apply for a bereavement discount retroactively. That policy didn't exist. The airline argued the chatbot was "a separate legal entity" responsible for its own words. A tribunal disagreed and made Air Canada pay.
That case is now the standard warning in every serious guide to deploying one of these things — because the lesson isn't "don't use chatbots," it's "don't deploy one without controlling what it's allowed to say."
Here's what's actually working for small businesses right now, what it costs, and how to train a bot so it helps instead of hallucinating your returns policy.
Where Things Actually Stand in 2026
The adoption numbers vary a lot depending on who's counting, but a few figures show up consistently enough to trust.
Around 38% of small businesses using AI have adopted customer service chatbots specifically, making it the second most common AI use case after writing and content generation. Separately, roughly 64% of small businesses say they plan to adopt an AI chatbot by the end of 2026, up from 38% in 2024 — so a lot of the growth is still ahead of us.
On performance, the honest range is 50-70%, not the 80-90% some vendor pages imply. Current-generation platforms, powered by LLMs and tied into a real knowledge base, resolve roughly 50 to 70 percent of inbound conversations autonomously without a human handoff. That resolution rate depends more on the quality of the knowledge base than on which platform you pick (see the best AI chatbot platforms for small businesses).
On cost, the case for doing this at all is straightforward: a standard support ticket resolved by an AI agent costs roughly $0.46 versus $4.18 for a human-handled ticket — call it a 9x gap. For a small business fielding even 50-100 questions a day about hours, pricing, or order status, that difference adds up fast.
Quick Stats
| Metric | Figure | Source |
|---|---|---|
| Small businesses using AI chatbots for customer service | 38% | 2026 survey data |
| Small businesses planning chatbot adoption by end of 2026 | 64% | Industry benchmarks |
| Typical AI resolution rate (no human handoff) | 50-70% | Platform analysis, 2026 |
| Cost per AI-resolved ticket vs. human-handled | $0.46 vs. $4.18 | Forrester TEI 2026 |
| Customers who trust businesses to use AI ethically | 42% (down from 58% in 2023) | Salesforce |
What a Chatbot Actually Costs in 2026
This is the part most guides gloss over, and it's where budgets get blown. Pricing for these tools falls into roughly three tiers, and the "starting price" advertised on a homepage is rarely the number you'll actually pay once resolutions or seats are factored in.
| Platform | Entry price | How it scales | Best fit |
|---|---|---|---|
| Chatbase | $19/mo | Message credit tiers | Solo founders testing the concept |
| Crisp | Free–$95/mo | Flat workspace pricing | Small teams wanting predictability |
| Tidio (Lyro AI) | $29/mo base + $39/mo AI | Conversation volume tiers; sharp jump for Plus tier | Shopify and small e-commerce |
| Intercom (Fin AI) | $29/seat/mo | Seats + per-resolution fees ($0.99 each) | Growing support teams that want resolution accuracy |
| Drift (Salesloft) | Custom (~$2,500/mo+) | Sales-led contract | B2B teams doing account-based sales |
For an SMB or a small SaaS operation, the realistic range sits at $30-200 a month on tools like Tidio or Crisp. The trap is per-resolution pricing: as your bot gets better at answering questions, your bill goes up proportionally — a bot going from a 25% to a 75% resolution rate can triple your monthly cost on the same volume of conversations.
This is usually the point where DIY setups hit a ceiling — not because the tools are bad, but because nobody accounted for the seat fees, the resolution fees, and the time it takes to configure and maintain the thing properly.
The Honesty Section: What These Tools Can't Do Yet
AI-powered customer service actually fails at roughly four times the rate of other AI use cases, according to industry surveys — and it's rarely the underlying AI model that's the problem.
The recurring root causes of failure: * Automating without an escape route for the customer * Hallucinations caused by poorly constrained data * Training the bot on documentation alone instead of real customer conversations * Ignoring industry-specific requirements * Measuring deflection rate (the customer stopped asking) instead of actual resolution
Since the Air Canada ruling, tribunals and courts have generally treated a chatbot's output as the company itself speaking — meaning if your bot promises a discount, a refund, or a policy that doesn't exist, you may be on the hook for it (read AI chatbot vs AI agent).
The trust gap backs this up. Only 42% of customers currently trust businesses to use AI ethically. People aren't rejecting chatbots outright — 62% actually prefer a chatbot over waiting for a human agent for simple stuff (read 80% of callers who hit voicemail hang up) — but they're wary of being misled or stonewalled by one.
What to Actually Train Your Chatbot On
Training isn't really about teaching it to talk; it's about telling it exactly where it's allowed to have an opinion and where it has to hand off to a human.
- Start with your actual FAQ and policy pages — hours, shipping, returns, pricing tiers, warranty terms.
- Feed it real past conversations, not just documentation. Actual questions come phrased messier than your FAQ page anticipates.
- Define explicit scope boundaries. List what the bot is allowed to answer, and what it should immediately hand off — refunds, complaints, anything involving money changing hands outside a listed policy.
- Use retrieval-grounded architecture (RAG), not a raw model. RAG grounds every response in your actual documents and knowledge base rather than letting the model improvise.
- Version and date your knowledge base. Update it whenever pricing, policies, or products change.
- Build a visible, easy escalation path. A customer should never feel trapped in a loop.
- Disclose that it's a bot. Simple, and increasingly a legal requirement.
Reality Check
| What marketing says | What's actually true |
|---|---|
| "Resolves 80% of tickets automatically" | 50-70% is realistic with a strong knowledge base |
| "Set it up once and forget it" | Needs quarterly review minimum; pricing/policy changes cause quiet drift |
| "AI eliminates the need for support staff" | Reduces the workload; complex or emotional cases still need a human |
| "More conversations resolved = pure savings" | On per-resolution pricing, better performance can triple your bill |
A Simple Decision Framework
- Under 50 customer questions a month: A free or near-free tier (Crisp, Chatbase, Tidio's free plan) is genuinely enough.
- 100-500 customer questions a month: A flat-fee mid-tier tool like Tidio with Lyro or Crisp Pro is the sweet spot.
- 1,000+ monthly conversations: Intercom's Fin or Zendesk AI starts to make sense, just model the bill at your actual volume first.
Where a lot of small business owners get stuck isn't picking the tool — it's the setup. Writing the scope boundaries, cleaning up the knowledge base, and testing it against weird questions takes more time than onboarding videos suggest.
That's exactly the kind of groundwork Brandywebs handles as part of a website build or a standalone automation project — custom sites start at $999, and chatbot setup can be scoped in alongside it.