Almost every guide to building an AI chatbot without code describes the same twenty minutes: sign up, paste your website URL, embed a script, done. That part genuinely is easy now. It is also about a tenth of the work, and it is not the part that decides whether the thing helps you or embarrasses you.
The two things that decide it are the quality of what you feed the bot, and the fact that in law, what your bot says is what your business said. Both are covered below, along with the four different ways these platforms bill you — because that is where the budget goes wrong. Checked 20 August 2026.
Start here: you are responsible for what it tells people
In 2024 a Canadian tribunal decided a case that every small business deploying an AI chatbot should know about. A passenger followed advice given by an airline’s website chatbot about bereavement fares; the advice was wrong. The airline argued, among other things, that the chatbot was effectively a separate entity accountable for its own statements. The tribunal rejected that outright and held the company liable for negligent misrepresentation.
That is the whole legal position in one sentence: your chatbot is not a third party. It is your business speaking. If it quotes a price, promises a refund window, describes a warranty or states a policy, you may be held to it.
Three practical consequences, and they should shape the build rather than be bolted on afterwards:
- Never let it improvise on money or policy. Refunds, pricing, warranties, delivery guarantees, eligibility — these get scripted answers or a handoff, never generated ones.
- Default to refusing. A bot that says “I’ll get a human to confirm that” is worth far more than one that answers everything.
- Log every conversation. If a dispute arises, the transcript is the evidence, and you want it to exist.
The build is a documentation project
Here is the thing platform marketing hides. When you point a bot at your website and it gives wrong answers, the bot is usually working correctly — it is faithfully reproducing documentation that is stale, contradictory or missing.
So the real sequence is:
- Pull your last 200 real customer messages. Email, chat, DMs, whatever you have. Do not guess what people ask.
- Count them. You will almost always find that fifteen to twenty questions cover the large majority of volume. That list is your scope.
- Write one clean, current answer for each. Short, unambiguous, dated. If two pages on your site disagree, fix the pages — this is the step people skip and then blame the AI for.
- Mark which answers are commitments. Anything involving money, time limits or eligibility goes in the scripted-or-escalate pile from the section above.
- Only now open a chatbot platform.
Do steps one to four properly and almost any tool will perform well. Skip them and no tool will. This is the same principle behind building a custom GPT for your business: the configuration is trivial, the source material is everything.
The four meters — and the perverse one
Chatbot platforms do not sell the same unit, which makes headline prices meaningless against each other. There are four models in use.
| Model | Example rates (Aug 2026) | Suits |
|---|---|---|
| Per resolution | Around $0.99 per resolved conversation on top of seats from about $29/seat, with a monthly minimum; another major helpdesk charges roughly $1.50–2.00 per automated resolution | Teams already on that helpdesk, with high volume and good docs |
| Per conversation | AI add-ons from roughly $39/month for around 50 conversations up to about $289 for 500, on top of a base plan from about $29 | Small businesses with predictable, moderate volume |
| Per message or credit | Builder tools from roughly $19–40/month, with extra credits around $40 per 1,000 | Simple website bots; cheapest way to start |
| Flat subscription | Fixed monthly fee regardless of volume | Anyone who values a predictable bill over a low one |
Now the part worth pausing on. Under per-resolution pricing, the better your bot performs, the more you pay. Improve your documentation, raise your automation rate, and your bill rises in direct proportion to your success. That is not a scandal — it is outcome-based pricing working as designed — but it means the savings model everyone assumes (“AI handles 60% of tickets so I pay 60% less”) does not apply. You are buying capacity, not savings.
Two further cautions. Vendors define “resolution” differently, and their definition is part of the price — some count a conversation the bot ended, others count only end-to-end outcomes, and at least one counts an AI interaction as a ticket as well. And marketed automation rates run ahead of reported real-world ones; independent write-ups have put achieved resolution meaningfully below the headline figures vendors advertise, though most such tests come from companies selling alternatives, so treat the specific percentages as directional.
Before committing, run the arithmetic on your own volume for each model, the way we did for automation platforms in automating work with AI tools. The cheapest headline is frequently the most expensive outcome.
Configuring it so it behaves
Four settings matter more than everything else in the interface.
Closed domain. Restrict answers to your uploaded material only. Every serious platform has this. Turn it on. A bot that answers general questions is a bot that will eventually answer a question about a competitor, a medical matter or a refund policy you never wrote.
An escalation trigger you can name. Define precisely when it hands over: low confidence, a named topic, an angry tone, a second failed attempt, or the customer simply asking for a person. Write these as rules, not hopes.
A short, boring persona. Skip the personality. Instruct it to be brief, factual, and to say when it does not know. Charm is where public chatbot incidents come from.
Fallback text you wrote yourself. The default “I’m sorry, I didn’t understand” is a dead end. Yours should offer the next step — a human, a form, a phone number.
Bear in mind the general limitation from what AI agents can and can’t do: these systems are safe where a wrong answer is cheap to spot and cheap to reverse. Support answers usually qualify. Actions — issuing refunds, changing bookings, cancelling orders — usually do not, at least not without a human in between.
Test it properly before it goes live
Take fifty real past customer questions — the messy ones, with typos and missing context — and run them through. Score each answer as correct, wrong, or correctly escalated.
| Result | What it means |
|---|---|
| Correct | Good. This is your real automation rate, not the marketing one |
| Correctly escalated | Also good. Escalation is a success, not a failure |
| Wrong but confident | The only real failure. Fix the source document, then retest |
Then launch quietly — one page, off-peak, or a percentage of traffic if the platform supports it. Read every transcript for the first week. You will find three or four documentation gaps you would never have predicted, and fixing those is worth more than any model upgrade.
What a no-code chatbot will not do
- Replace your support person. It absorbs repetitive questions. The hard, angry and unusual ones still arrive, and now they arrive already annoyed at having talked to a bot first.
- Fix bad documentation. Repeated for emphasis, because it is the single most common cause of a disappointing deployment.
- Handle account-specific questions without integration into your systems, which is where no-code stops being no-code.
- Stay accurate on its own. Change a price, a policy or a shipping time and the bot keeps citing the old one until someone updates the source. Put a review in the calendar.
- Survive a hostile visitor. Public bots get probed. Closed-domain settings, scripted answers on commitments, and logging are your defence.
- Guarantee a saving. See the metering section — on outcome pricing, success costs money.
Should you build one at all?
Yes if you answer the same fifteen questions constantly, your documentation is decent or fixable, and your customers are online rather than on the phone. Start on the cheapest capable tier, monthly billing, and give it thirty days.
No if your volume is low, your answers are genuinely bespoke, or your documentation is a mess you do not have time to fix. In that case a well-written FAQ page and a saved-replies file will do more for less — which is the same conclusion we reached about tool sprawl generally in cutting business costs with AI and free vs paid AI tools.
Once the bot is answering reliably, the next problem is the system around it — routing, handoff quality and keeping the human experience intact. That is covered in building a no-code customer support system and AI customer service without losing the human touch. For where this sits in a wider stack, see best AI tools for small business owners.
Frequently asked questions
Do I need coding skills to build an AI chatbot?
No. Modern platforms let you upload documents or point at a website, configure behaviour through a form, and paste an embed snippet. The difficult work is preparing accurate source material and defining escalation rules, neither of which involves code.
Am I legally responsible for what my chatbot says?
Generally yes. A Canadian tribunal held an airline liable in 2024 for wrong advice its website chatbot gave, rejecting the argument that the bot was a separate responsible entity. Treat chatbot statements as statements by your business, and script or escalate anything involving money or policy. This is general information, not legal advice.
How much does an AI chatbot cost per month?
It depends entirely on the billing model. Builder tools start around $19–40 a month on message credits; conversation-based AI add-ons run roughly $39 for 50 conversations up to about $289 for 500; per-resolution helpdesks charge around $0.99 to $2.00 per resolved conversation on top of per-seat fees. Price your own volume against each model.
Why does my chatbot give wrong answers?
Almost always because the source material is wrong, outdated or contradictory. The bot is reproducing what it was given. Fix the underlying document rather than adjusting the prompt.
What is a realistic automation rate?
Lower than the marketing. Measure your own by running fifty real past questions through before launch and scoring the answers. Count correct escalations as successes — a bot that hands over cleanly is working, not failing.
Should the chatbot handle refunds or bookings?
Not autonomously. Those are irreversible actions with financial consequences. Let it gather details and prepare the request, then have a person approve it.
Sources
- British Columbia Civil Resolution Tribunal — the body that decided the 2024 airline chatbot case
- Intercom — official pricing, including per-resolution rates
- Tidio — official pricing and AI add-on tiers
- Zendesk — official pricing and automated resolution rates
Pricing checked 20 August 2026 and changes frequently; several vendors moved to outcome-based billing during 2025 and 2026. Verify current rates and each vendor’s definition of a billable resolution before committing. Nothing here is legal advice.
