AI Chatbot for Your Website: 20 Questions Answered (2026)
We get asked about chatbots more than anything else on this list, and roughly half the businesses who ask us shouldn't build one. That's not a sales tactic. It's just what the numbers look like once you sit down and count the questions a bot would actually be answering.
This is the version of the conversation we'd have on a call, written down. Twenty questions, in the order people usually ask them.
1. What is an "AI chatbot" in 2026, exactly?
A chat window on your site, backed by a language model that has been given your content to read before it answers.
The chat window isn't the interesting part — that's been the same for fifteen years. What changed is what sits behind it. The old bots matched keywords against a script someone wrote by hand. The current ones read your pages, your price list, your policies, and your FAQ, and answer from that material in plain language.
The technical name for the pattern is retrieval. In practice it means the bot looks things up before it speaks, rather than answering from memory.
2. How is this different from the chatbot everyone hated in 2019?
That one made you pick from a menu, then said "I didn't quite catch that" until you gave up and looked for the phone number.
The failure mode has moved. Modern bots rarely get stuck. They get confidently wrong instead. A badly built one will invent a returns policy you don't have, in fluent English, and sound completely sure of itself.
That's a worse failure in some ways and a fixable one in others. It's fixable because it's a design problem, not a model problem. See question 4.
3. Do I actually need one?
Count your inbox.
If you get five enquiries a week and each one is different, no. You'll spend money to save yourself twenty minutes and you'll lose the personal contact that's probably winning you the work.
If you get forty enquiries a week and thirty of them are four versions of the same three questions — opening hours, do you deliver here, how much, do you have this in stock — then yes, and it'll pay for itself inside a quarter.
The dividing line isn't volume on its own. It's repetition.
4. Will it make things up about my business?
If it's built carelessly, yes.
The fix is boring and it's mostly instruction and scope. The bot answers from your documents and nothing else. It's told, explicitly, that if the answer isn't in the material it has, it does not guess — it says so and hands over. You test it against the questions you're afraid of, not the ones you hope for.
The single most important line in the whole build is the one telling the bot what to do when it doesn't know. Most vendors treat that as an afterthought. It's the thing that decides whether the bot is an asset or an embarrassment.
5. What does it cost to build?
Depends almost entirely on whether it needs to touch your other systems.
A bot that answers questions from your existing content and hands off cleanly is the cheap end — typically low four figures USD, and we price in LKR for Sri Lankan clients so nobody's paying an exchange rate for the privilege. That's a couple of weeks of work.
A bot that checks live stock, books a slot, looks up an order, or writes anything into a system you rely on is a different project. You're paying for integration, permissions, error handling, and testing — usually several times the first number.
Be suspicious of anyone quoting a flat price before they've asked what systems you run.
6. What does it cost to run every month?
Two lines: hosting, and model usage.
Hosting is small — it sits alongside your site. Model usage is charged per conversation and it's cents, not dollars, for a normal FAQ workload. A busy small business site might run twenty to sixty USD a month in model costs. Costs have fallen every year since 2023 and we don't expect that to reverse.
Watch the SaaS pricing instead. Several platforms charge per "resolution," and their definition of a resolution is generous to themselves. Read that clause before you sign anything.
7. How long does it take?
Two to three weeks for a content-based bot. Six to ten for anything with real integrations.
The build isn't the slow part. Getting your content into a state where a machine can answer from it is the slow part. Half our chatbot projects turn into a documentation project for a week, because it emerges that the delivery policy lives in the owner's head and three staff members each describe it differently.
That week is not wasted. It's usually the most valuable week of the project.
8. What do you need from me?
The twenty questions you get asked most, answered properly. Your current pricing rules, including the exceptions. Anything you'd correct a new employee on in their first month. And a name — who the bot escalates to, and how fast that person replies.
If you can supply those four things, the rest is our problem.
9. Which model should we use? Does it matter?
Less than the vendor marketing suggests, for this kind of work.
Answering "do you deliver to Negombo" from a page that says so is not a hard task, and most current models do it well. Model choice starts mattering when the bot has to reason across several documents, work in a second language, or decide which tool to call.
What matters more: don't get locked in. The model should be a setting you can change, not a foundation you'd have to rebuild on. We've swapped models under live bots in an afternoon. That's how it should be.
10. Can it book appointments, take orders, or check stock?
Yes, and this is where the project gets serious.
Our standing advice is to earn write access. Start read-only — the bot can look up an order status or check whether something's in stock, but it can't change anything. Live with that for a month. Then, if the logs justify it, let it book.
A bot that reads the wrong number gives someone bad information. A bot that writes the wrong number creates a booking nobody made, in a slot you've now lost.
11. What happens when it can't answer?
It hands over, and how well it does that is the whole ballgame.
Handoff can be a form, an email, a WhatsApp thread, a ticket, or a live agent if you have staff for that. It should carry the conversation with it so the customer doesn't repeat themselves. Nothing annoys people faster than explaining their problem twice.
A bot that says "I don't have that — here's Nuwan, he'll know, he usually replies within the hour" is more valuable than one that guesses. Customers forgive not knowing. They don't forgive being misled.
12. Will it work in Sinhala, Tamil, or Arabic?
Yes, with real caveats.
Model quality in these languages has improved a lot and is still uneven. Sinhala and Tamil handle common phrasing well and get shakier with technical or regional vocabulary. Arabic raises the register question — Modern Standard reads correctly and can sound stiff to a customer typing in Gulf dialect.
The rule we follow: never ship a language nobody on the project can read. Get a native speaker to run fifty real questions through it before launch. That test catches more than any benchmark will.
13. Is it safe to connect to my customer data?
It can be, if you're disciplined about access.
The bot gets the minimum it needs and nothing more. Scoped credentials, not admin keys. Read-only wherever read-only will do. Every query logged. And the customer data it touches should be data that customer is already entitled to see about themselves.
The mistake we see is people wiring a bot straight into a CRM on day one because the integration was one click. One click is exactly how a chat widget ends up able to read your whole contact list.
14. What about data protection law?
We're not lawyers, and for anything with real exposure you want one. But the practical shape of it is consistent across Sri Lanka's PDPA, GDPR, and the UAE regime:
Tell people they're talking to a bot and that the conversation is stored. Have a retention period and actually enforce it — most chat logs don't need to exist after 90 days. Check your model provider's terms on training: you want your customer conversations excluded, and that's a contract term, not a setting. If you're handling EU or UK residents' data, get a data processing agreement in place.
None of this is expensive. It's just easier to do at build time than after a complaint.
15. Can someone trick it into saying something embarrassing?
Someone will try within the first week. Usually a bored teenager, occasionally a competitor.
You reduce the damage by limiting what the bot is capable of saying in the first place. It has no authority to offer discounts, agree to refunds, or make commitments. It stays on topic and declines everything else. You filter output and you monitor logs, especially early.
Assume screenshots. Build the bot you'd be comfortable seeing quoted out of context, because eventually it will be.
16. Who's liable if it tells a customer the wrong thing?
Practically speaking, you are. It's on your site, in your name, and the customer has no idea which vendor built it.
The legal position is genuinely unsettled and varies by jurisdiction — again, not lawyers, and this is one to ask a real one about if you're in a regulated trade. What we can tell you is the design answer: don't let the bot make binding statements. Prices it quotes are indicative and it says so. Anything contractual routes to a human before it's confirmed. Availability comes from the live system or it doesn't get quoted.
Most of the horror stories are businesses that let a bot promise something the business then had to honour.
17. Will it hurt my SEO?
Not if it's built properly, and it might quietly help.
The direct risk is performance. A heavy third-party widget that loads on every page and shifts your layout while it initialises will cost you Core Web Vitals. Lazy-load it, defer it, and measure before and after.
The indirect benefit is bigger. The log of questions your bot couldn't answer is the cleanest keyword research you'll ever get, written by actual customers in their own words. Turn the top twenty into pages. That's a content plan that came free with the chatbot.
18. Do I still need human support staff?
Yes. What changes is what they spend their day on.
The bot takes the repetitive volume — hours, location, delivery, stock, order status. What's left is the complaints, the negotiations, the unusual cases, and the customers who want to talk to a person because the purchase matters to them. That work is harder and it's also where your margin is.
Anyone selling you headcount reduction on a small support team is selling you something. On a team of two, you don't remove a person. You get the two people back.
19. What's the difference between a chatbot and an "AI agent"?
A chatbot answers. An agent acts.
The marketing has blurred this to the point of uselessness, so here's the practical test: can it change something? If the worst outcome of a mistake is bad information, it's a chatbot. If the worst outcome is a cancelled order, a wrong booking, or an email sent in your name, it's an agent, and it needs to be built like one — permissions, audit trail, a way to undo things, and a human in the loop for anything irreversible.
Same underlying technology. Very different risk profile, and very different price.
20. How do I know if it's working?
Most chatbot dashboards lead with "conversations handled." That number is worthless. It counts the person who typed "hi" and closed the tab the same as the person who booked a table. Worse, the standard "resolved without human intervention" metric quietly records as a success every session where somebody gave up, went looking for your phone number, and needed a human.
Here's what actually goes on the dashboard:
- Handoff rate, and the reason behind each one. A 30% handoff rate isn't a failure. Thirty percent of handoffs because the bot couldn't find your opening hours is.
- Unanswered questions, ranked by frequency. This list is usually the most useful thing we hand you — it tells you what your website should have said in the first place.
- Assisted conversions — sessions where someone chatted and then bought or booked.
- Average messages to resolution. If that number is climbing, something's broken.
- Thumbs up or down at the end of the chat. Blunt, but people are honest with a thumb.
Give it 60 days. If the numbers aren't moving, we fix it, or we tell you it's the wrong tool for your business. We'd rather have that conversation than keep invoicing for something that isn't earning its keep.
Where to start
You don't need a full build to find out whether this is worth it.
Write down the 20 questions you get asked most. The real ones, including the boring ones. If most can be answered from information that already exists on your site or in your head, a chatbot will pay for itself quickly. If most need judgement, negotiation, or someone the customer actually trusts, it won't — and we'll tell you that.
We'll run that assessment on a call at no charge, and give you a straight answer either way.
Related reading: Small Business Website Security: 19 Questions Answered
About the author
Lyle Heartman is the founder of ELRQ Agency, which builds web platforms, AI automation, and security-hardened integrations for businesses in Sri Lanka, the Gulf, and North America. Get in touch for a free chatbot readiness assessment.