AI Chatbots
A chatbot that answers from your own information, not from guesswork.
What is AI chatbots?
An AI chatbot for business is an assistant on your website or WhatsApp that answers customer questions using your own documents, prices and policies, hands over to a person when it cannot help, and keeps a readable record of every conversation.
Bad chatbots nearly all fail the same way. Nobody gave them your information, so they generate answers that sound plausible and are wrong, which costs more trust than the tool ever saves in time.
We build yours on top of your own material: the price list, the policies, the product details, the opening hours. When the answer is not in there, it says so and passes the customer to a person rather than filling the gap.
It runs on your website and on WhatsApp, which is where most Kenyan customers will actually message you. You can read every conversation, correct wrong answers, and add new information without calling us.
The two conditions we build under
Duara builds customer-facing chatbots from Nairobi for businesses whose enquiry volume has outgrown the person currently answering it.
We insist on two things before building: that the bot answers only from material you supply, and that reaching a human stays easy and obvious. Clients occasionally ask us to hide the human option to cut costs, and we decline, because it damages more than it saves.
Who this is for
- You answer the same questions over and over
- Customers message you outside working hours
- Your team is buried in WhatsApp enquiries
- Enquiries sit on one staff member phone
Industries we serve
- Retail and eCommerce
- Clinics and healthcare
- Schools and colleges
- Hospitality and travel
- Property and estate agency
- Professional services
Problems this solves
These four are what businesses describe when they call us about a chatbot, usually in this order.
- The same questions, endlessly. Opening hours, prices, availability and delivery times consume most of the time your team spends replying to customers.
- Enquiries arriving overnight. Messages sent at nine in the evening get answered at nine the next morning, by which point the customer has often asked somebody else.
- Everything landing on one phone. WhatsApp enquiries reach one person handset, so they stop when that person is busy, asleep or leaves the business.
- Chatbots that invent answers. A previous attempt produced confident nonsense because it was never given the business own information to work from.
What we deliver
A chatbot trained on your documents, live on your website and WhatsApp, that answers what it knows, admits what it does not, hands over cleanly, and gives you a screen to correct it.
A chatbot trained on your own documents and prices
Website and WhatsApp using the same answers on both
Clean handover to a person when it cannot help
A screen where you read conversations and fix wrong answers
Reporting on the questions it could not answer
Setup of the WhatsApp Business account required for automation
Features
- Answers grounded in your documents. Every response comes from material you supplied, and the system says so rather than generating when the answer is absent.
- One answer source, two channels. Website and WhatsApp draw from the same material, so a customer never gets different answers depending where they asked.
- Obvious route to a person. Asking for a human is easy and visible. Hiding it is the fastest way to turn a helpful tool into a complaint.
- Self-service corrections. You update the source document and the answer changes. Day to day edits do not require us.
- Unanswered question reporting. What customers asked that it could not handle, which tells you exactly what information is missing.
Benefits
- Instant replies at any hour. The enquiry arriving at nine at night is answered at nine at night, which is frequently what decides the sale.
- Your team stops repeating itself. Routine questions are absorbed, leaving staff on the enquiries that actually need judgement.
- Enquiries stop living on a personal phone. Conversations sit in one place the business owns rather than one handset.
- You learn what customers actually ask. The unanswered question report is usually the most useful output in the first month.
Technologies we use
- Retrieval from your documents. The system looks up your material first and answers from what it finds, rather than generating from memory.
- WhatsApp Business API. The official route required for automated replies at volume, including the account setup and verification.
- Website widget. A lightweight embed that does not slow your pages, which matters for both ranking and conversion.
- Conversation logging. Every exchange stored and searchable so you can audit what was said to whom.
- Handover routing. Rules deciding which conversations reach which person, including outside working hours.
The technology matters less than the grounding. A cheaper model answering from correct documents beats an expensive one guessing.
How the process works, step by step
A chatbot is usually answering real customers within three to four weeks, including WhatsApp account verification.
-
Step 1
Free enquiry review
We look at your actual recent enquiries to see how many a chatbot could genuinely answer.
-
Step 2
Gather your material
We collect the prices, policies, product details and FAQs the bot will answer from.
-
Step 3
Build and internal testing
We build it and your team tests it against real past questions before any customer sees it.
-
Step 4
Supervised launch
It goes live with your team watching the conversations and correcting answers as they appear.
-
Step 5
Tuning from real questions
The unanswered question report drives what gets added, usually weekly at first.
Quality assurance and how we work
What we hold to on customer-facing chatbots specifically.
- It never invents an answer. The system is configured to say it does not know and offer a person, rather than generating something plausible.
- The human route stays visible. We decline requests to hide or bury the option to reach a person, because it costs more than it saves.
- You test before customers do. Your team runs it against real past enquiries and signs off before it faces a single customer.
- You can correct it without us. Editing the source material changes the answers. We would rather you did that than waited on our availability.
- Conversations belong to you. The full log is yours, exportable, and never used to train public AI models.
Frequently asked questions about AI chatbots
Will the chatbot make up answers?
No. It answers from the documents you supply, and where the answer is absent it says it does not know and offers a person. That behaviour is configured deliberately.
Can we change what it says?
Yes. You update the source documents and the answers change. Day to day edits do not require us.
Can the chatbot run on WhatsApp?
Yes, through the WhatsApp Business API, which is the official route for automated replies at volume. We handle the account setup.
Do we need a WhatsApp Business account?
Yes for automated replies. Setting it up and getting it verified is part of the work rather than something you arrange separately.
Can customers still reach a person?
Always, and easily. We decline requests to hide that option because it produces complaints rather than savings.
What happens when the office is closed?
It answers what it can and captures the rest as messages routed to whoever is on duty next.
How long before the chatbot is live?
Usually three to four weeks including WhatsApp verification, which is often the slowest part.
What does it cost to run each month?
It depends on message volume and WhatsApp charges. We give you the monthly figure before you commit.
Will it slow down our website?
No. The widget is lightweight and loads separately, so page speed and search ranking are unaffected.
Will the chatbot reply in Swahili?
Yes, including conversations that switch between English and Swahili, which is how most customers actually write.
How do we know what it got wrong?
You read the conversations. The screen shows everything it said, and you correct answers directly at the source.
What is the unanswered question report?
A list of what customers asked that it could not handle. In the first month it is usually the most useful thing the system produces.
Can it take orders or bookings?
Yes, where you want it to. That is scoped separately because it needs to connect to your ordering or calendar system.
Does it learn automatically from conversations?
No, and deliberately so. Learning from unreviewed conversations is how chatbots acquire bad answers. You approve what gets added.
Is our customer data safe?
Conversations stay in your system, are exportable, and are never used to train public AI models.
What if a customer is angry?
Escalation rules pass those conversations straight to a person, rather than letting the bot attempt to handle them.
Can it work with our existing helpdesk?
Usually. We connect to the common systems so conversations and handovers land where your team already works.
How many questions can it answer?
It depends entirely on how much material you have. More documented information means more it can handle.
Do we need someone monitoring it?
Closely for the first fortnight, lightly afterwards. The correction workflow is designed to take minutes rather than hours.
What if it does not work for us?
The enquiry review before we build is there to prevent that. If the proportion it could handle is too low, we say so then.
Glossary of terms
Chatbot terms explained without the vendor vocabulary.
- Grounding
- Restricting a system to answer only from supplied documents rather than from its general training.
- Hallucination
- A confident, fluent answer that is simply untrue, which is what ungrounded chatbots produce.
- Handover
- The point where a conversation transfers from the automated system to a member of staff.
- WhatsApp Business API
- The official interface required for automated messaging at volume, as distinct from the ordinary app.
- Intent
- What the customer is actually trying to achieve, as opposed to the literal words they typed.
- Fallback
- What the system does when it cannot answer, which should be an offer of a person rather than a guess.
Related services in AI Implementation
Talk to us about AI chatbots
Send us a week of your actual customer messages. We will tell you what proportion a chatbot could genuinely handle before you commit to anything.
Start your project