A customer messages at eleven at night to ask whether you deliver to Sharjah. Another wants the fee for the weekend batch. A third asks where the order is. Your team answers the same twenty questions every day, and the answers are all written down somewhere. An AI chatbot for business looks like the obvious fix, and sometimes it is. It is also sold with more promise than it can keep.
We are a product engineering team at Infopark in Kochi. We build chatbots and assistants grounded in a company's own content for businesses in India, the UAE and Australia. We also say so when a clearer FAQ page or a better contact form would do the job. This guide explains what a chatbot is good at, where it fails, and how to tell which case you are in.
The short answer
An AI chatbot is worth having when many people ask questions whose answers already exist in your documents or systems, and when a wrong answer can be caught and corrected without much harm. It is a poor fit when questions are few, when the answers are not written down, or when the reply is a decision with legal, medical or financial weight. It works best as the first line, with a clear route to a person behind it.
How an AI chatbot works, in plain terms
There are two kinds of chatbot, and the difference matters. A rule-based bot follows a script. The customer taps a button or types a keyword, and the bot returns the reply written for that path. An AI chatbot uses a language model, so it can read a question in the customer's own words and write a reply.
The language model
A language model is software trained on a very large amount of text. Given some text, it produces the words most likely to come next. That is enough to make it good at reading a loosely worded question and writing a clear reply. It also explains the main weakness. The model is not looking facts up. It is producing plausible text, and plausible text can be wrong.
On its own, the model knows nothing about your prices, your policies or your stock. Asked about them, it may still produce an answer.
Grounding it in your own content
The usual fix is called retrieval. Your documents, such as policies, product sheets, fee structures and help articles, are indexed. When a question arrives, the system searches that index for the most relevant passages. It hands them to the model with the question and an instruction to answer only from them. The reply can then point to the document it came from.
Retrieval reduces wrong answers. It does not remove them. If the right passage is missing, out of date or contradicted by another document, the reply will reflect that.
Connecting it to your systems
A chatbot can also be given a small set of actions: look up an order by its number, check an appointment slot, raise a ticket, record an enquiry in the CRM. Each action is a defined connection to your software with its own permissions. The chatbot decides when to use one. Your system decides whether it is allowed.
What an AI chatbot does well
A chatbot does well on work that is repetitive, already documented and low in risk. In practice that means the following.
Answering routine questions at any hour, from content you have already written.
Understanding the same question asked in many different ways, including with spelling mistakes and informal wording.
Collecting details before a person takes over: name, requirement, order number, preferred time.
Looking up a single record, such as an order or an appointment, when it is connected to the system that holds it.
Summarising the conversation, so the staff member who takes over does not ask the customer to repeat it.
Helping your own staff find things in policies, manuals and past proposals.
The last item is often the safer first project. Staff can judge whether an answer is right, and a mistake stays inside the company.
What an AI chatbot cannot do
The limits come from how the technology works. Better wording in the instructions narrows them, and none of them goes away.
It can be wrong and sound certain
A language model writes a wrong answer in the same fluent, confident tone as a right one. It does not reliably know when it is guessing. A customer who is told the wrong refund period or the wrong fee will hold you to it. Any design has to assume that some answers will be wrong, and decide in advance where that is acceptable.
It only knows what you give it
If the delivery policy lives in one manager's head, the chatbot cannot know it. If last year's price list is still in the folder, the chatbot may quote it. Most of the work in a chatbot project is putting the content in order, and that work continues after launch.
It should not make decisions that need judgement
Approving a refund outside policy, advising on a medical symptom, quoting for a custom job, handling an angry customer: these need a person. A chatbot can collect the facts and pass the case on. It should not be the one to decide.
A chatbot on a public website will also be tested by people who try to make it promise a discount or say something foolish. Limit what it can say and what it can do.
It costs money every time it is used
Model providers generally charge by the amount of text processed. Every conversation has a cost, and long conversations and long documents cost more. The amount per conversation is usually small, and it scales with use, so it should be measured from the first day.
It handles private information
Customers type names, phone numbers, addresses and sometimes health or payment details into a chat box. You need to know where that text goes, how long it is kept, and whether the model provider may use it for training. The terms differ between providers and between plans, so read the ones that apply to you. Our post on data privacy and regulation covers the wider duties.
Rule-based bot, AI chatbot or a person: a comparison
Most businesses end up with a mix of the three. The table shows what each is suited to.
| Rule-based bot | AI chatbot | A person | |
|---|---|---|---|
| How it answers | Follows a script of buttons and keywords. | Reads free text and writes a reply from your content. | Understands context and uses judgement. |
| Questions it handles | Only the ones you scripted. | Routine questions whose answers are in your documents or systems. | Anything, including complaints and exceptions. |
| When it does not know | Fails visibly and offers the menu again. | Should say so and hand over. Without care, it may guess. | Asks a colleague. |
| Availability | All hours. | All hours. | Working hours, unless you staff shifts. |
| Cost profile | Low to set up and to run. | A build, then a usage cost for every conversation and upkeep of the content. | Salary and training. Cost rises with volume. |
| Upkeep | Scripts edited by hand when something changes. | Documents kept current, and answers checked against a test set. | Training and supervision. |
| Main risk | Customers get stuck in menus and give up. | A confident wrong answer. | Slow replies and missed messages at busy times. |
A rule-based bot is not the inferior option. For booking a slot, checking a balance or choosing between four departments, buttons are faster for the customer and cannot give a wrong answer.
When you do not need an AI chatbot
Skip the chatbot if enquiries are few. A person who answers ten messages a day does it better than any software, and learns something about customers each time.
Skip it if one page would answer the questions. When most people ask about price, timings and location, put those on the website where they can be seen, and on the WhatsApp Business profile. A chatbot that exists because the website hides the answers is solving the wrong problem.
Skip it, for now, if the answers are not written down. Write them first. The exercise is useful on its own, and it is the first step of a chatbot project in any case.
Skip the custom build if a tool you already pay for includes a chatbot. Many helpdesk, website and messaging products now do. Try that first. A custom build earns its cost when the chatbot has to read your own systems, follow your own permissions, or sit inside your own product or customer app.
How to decide
Decide on volume, documentation and the cost of a mistake. Answer these questions about your own business.
How many enquiries arrive in a week, and what share of them are the same few questions?
Is the answer to each of those questions written down in one current document?
What happens if the chatbot gives a wrong answer to each one? Who notices, and what does it cost?
Do customers message outside working hours, and do you lose business when nobody replies?
Who will keep the content current and read the transcripts each week?
If volume is high, the answers are documented and mistakes are cheap to correct, a chatbot is worth scoping. If any one of the three is missing, fix that first.
What a sensible first build looks like
A sensible first chatbot is narrow. It answers one group of questions well, hands over cleanly, and is measured from the start.
Start with the questions. Collect real enquiries from WhatsApp, email and the phone log, and group them by topic.
Put the content in order. One current document per topic, with old versions removed.
Pick a narrow scope. Delivery and returns, or fees and admissions. Everything else goes to a person.
Build the handover first. The customer can ask for a person at any point, and the person receives the conversation so far.
Test before launch. Keep a fixed set of real questions with known correct answers, and run it again whenever the content, the instructions or the model changes.
Read the transcripts after launch. They show what customers ask that nobody wrote down.
This is the same six-stage process from discovery to growth we follow on every build, with working software every fortnight. The first stage is where we find out whether a chatbot is the right answer at all.
Frequently asked questions
Can an AI chatbot replace a customer support team?
No. It can take the routine questions off the team and collect details before a handover. Complaints, exceptions and anything that needs judgement still need people.
Can an AI chatbot work on WhatsApp?
Yes. A chatbot can be connected to WhatsApp through the WhatsApp Business Platform, and the same assistant can also serve a website and a mobile app.
Will it answer in Malayalam, Hindi or Arabic?
Current language models read and write many languages, and quality varies from one language to another. Test it with real questions in each language you need, including the mixed English and Malayalam or Hindi that customers type, before you offer it.
Will our data be used to train the model?
It depends on the provider and the plan. On the builds we do, your data is not used to train a model unless you ask for it, and we document which data crosses which boundary before anything is connected.
How do we know whether the chatbot is working?
Track how many conversations end without a handover, how many wrong answers the weekly review finds, and what each conversation costs. Compare them with the same period before launch.
Conclusion
An AI chatbot for business is a good first line for routine, documented questions. It is not a support team, and it should not be left to make decisions. If your enquiries are few or your answers are not written down, start with the content and a clearer website. If the volume is real and the answers exist, build something narrow, keep a person within reach, and measure it.
We are happy to look at the questions your team answers every day and give a straight view, including when the answer is that you do not need a chatbot. Tell us what your customers keep asking and we will start from there.
Human. Digital. Future.
