A good AI chatbot for business answers routine customer questions on your website or WhatsApp, using your own content, at any hour, and passes anything it can't handle to a person with the conversation attached. It can't fix a broken process, make promises your team wouldn't make, or replace staff who handle complaints and judgement calls. If you treat it as a well-trained front desk rather than a replacement for your support team, it can take a real share of repetitive work off your people. This guide covers what that looks like in practice, with a focus on businesses in Sri Lanka and the Gulf.
What a modern AI chatbot actually is
Older chatbots worked from decision trees. You clicked a button, it showed the next set of buttons, and anything outside the script got "Sorry, I didn't understand." Today's assistants are built on large language models such as GPT or Claude, which understand questions written in normal language, including typos, mixed languages and half-finished sentences.
On its own, a language model knows a lot about the world but nothing about your business. It doesn't know your delivery areas, your return policy or that you closed the Kandy branch last month. So a business chatbot needs two more pieces: your own content to answer from, and rules about what it may and may not do.
What it does well, and where it falls short
| Good fit | Poor fit |
|---|---|
| Opening hours, locations, delivery areas and fees | Complaints from upset customers |
| Product or service questions answered from your catalogue | Refund or compensation decisions |
| Order status, when connected to your order system | Anything involving legal, medical or financial advice |
| Booking or enquiry forms collected in conversation | Questions your own staff can't answer consistently |
| Explaining policies in plain language | Negotiating prices or custom deals |
| Collecting details before handing to a person | Situations where a wrong answer is costly and hard to undo |
A useful test: if a new staff member could answer the question correctly after reading your website and help pages, the bot probably can too. If the answer depends on experience, discretion or a manager's approval, it shouldn't be the bot's call.
Grounding it in your own content
The single biggest factor in answer quality is the content behind the bot. The usual approach is called retrieval: when a customer asks something, the system searches your documents for the relevant passages and gives only those to the language model, with an instruction to answer from them. If the answer isn't there, the bot should say so and offer a person.
That means the work starts with your content, not the AI. Gather your FAQs, policies, price lists, product details and the replies your team sends most often. You'll usually find gaps and contradictions, such as two different delivery times on two pages. Fix those first. A bot will repeat whatever it's given, confidently.
For live information such as stock, order status or appointment slots, the bot needs a connection to the system that holds it, through an API. Copying that data into documents goes stale within a day.
Handing off to humans
Every good support bot has a clear way out. Decide in advance when it should hand over:
- The customer asks for a person.
- The bot can't find the answer in your content.
- The topic is on a blocked list, such as refunds, complaints or legal matters.
- The customer seems frustrated, or has asked the same thing twice.
When it hands over, it should pass the full conversation and anything it has collected, such as the name, order number and issue, so the customer doesn't repeat themselves. Outside working hours, it should say honestly when someone will reply rather than leaving the customer waiting in silence.
Website chat or WhatsApp?
In Sri Lanka and across the Gulf, many customers would rather message a business on WhatsApp than fill in a form or use website chat. A bot on WhatsApp meets them where they already are, and the conversation stays on their phone.
To automate WhatsApp you need the WhatsApp Business Platform, accessed through Meta directly or through an approved provider, not the ordinary WhatsApp Business app on a phone. There are rules to plan around. Businesses can reply freely within a window after the customer's last message, but messages sent outside that window must use pre-approved templates. Meta also charges for some conversations, and those charges change, so check the current rules when you plan.
Website chat is simpler to set up and gives you more control over the design. Many businesses run both, using the same knowledge base behind each.
Arabic, Sinhala, Tamil and Malay
Large language models handle Arabic well, and they generally cope with Sinhala, Tamil and Malay too, though quality in Sinhala and Tamil can be less consistent than in English or Arabic. A few things matter more than the model itself:
- How people really type. In Sri Lanka, many customers write Sinhala or Tamil in English letters, often mixed with English words. In the Gulf, people mix Arabic and English in one message and write in dialect, not formal Arabic. Test with real messages from your inbox, not textbook sentences.
- Which language to reply in. Usually the customer's own, but decide what happens with mixed messages.
- Content in each language. If your policies only exist in English, the bot will translate on the fly. For anything sensitive, such as terms or prices, have a native speaker review the source content.
- Right-to-left layout. A website chat widget must display Arabic properly, including mixed Arabic and English lines.
There's more on local expectations on our Sri Lanka and UAE pages.
What drives the cost
We don't publish prices because the range is wide, but the cost depends on a few predictable things:
- Channels: website only, or WhatsApp and others too.
- Integrations: answering from documents is simpler than checking orders, bookings or a CRM in real time.
- Content preparation: clean, current content is quicker. Scattered PDFs and old pages take longer.
- Languages: each language needs its own testing.
- Volume: language models are billed by usage, so running costs rise with the number and length of conversations.
- Ongoing care: someone has to review conversations and keep content current.
Off-the-shelf chatbot tools are a sensible start for simple FAQ bots. A custom build makes more sense when you need deep integration with your own systems, control over where data is stored, or careful handling of several languages.
How to measure whether it's working
Agree on the measures before launch, so you're judging the bot on the same terms later:
- Resolution rate: the share of conversations that ended without a handover and without the customer coming back about the same issue.
- Handover quality: whether staff got enough context to help straight away.
- Accuracy: a weekly sample of conversations checked by someone who knows the answers.
- Customer feedback: a simple thumbs up or down at the end.
- Unanswered questions: the list of things it couldn't answer, which tells you what content to add.
Reading real transcripts each week is the most useful habit. Numbers tell you something is off. Transcripts tell you why.
Common mistakes
- Launching without a handover path. Customers who can't reach a person get angry quickly.
- Letting it answer beyond its content. A bot that guesses will eventually invent a policy.
- Hiding that it's a bot. Say so clearly. People are fine with bots that are honest and helpful.
- Feeding it stale content. Prices and policies change. Make updating the bot part of the change.
- Testing only in English. If your customers write in Sinhala, Tamil or Gulf Arabic, test in those.
- Treating launch as the finish line. The first month of real conversations is where most of the improvement happens.
Where to start
Pull the last few hundred customer messages from your inbox, WhatsApp and website forms. Group them by topic. The most common, most predictable questions are your bot's first job. Everything else can wait.
If you'd like help scoping it, see our AI solutions work, or contact us for a quote with a rough idea of your channels, languages and message volume.
Questions people ask
Will an AI chatbot replace my support team?
No. It handles routine, predictable questions so your team can spend time on complaints, exceptions and customers who need a person. Most businesses keep the same team and change what they spend their time on.
Can an AI chatbot make things up?
It can, if it's allowed to answer from general knowledge. Limiting it to your own content, telling it to say when it doesn't know, and reviewing conversations regularly keeps this under control.
Can the chatbot reply on WhatsApp?
Yes, through the WhatsApp Business Platform rather than the regular WhatsApp Business app. You'll need a verified business account, and messages sent outside the customer service window must use pre-approved templates.
Is customer data safe with an AI chatbot?
It depends on the setup. Check where conversations are stored, whether the AI provider may use your data for training, and how long data is kept. Avoid asking customers for card numbers or passwords in chat.