AI Chatbots for Business: Should You Use One? (An Honest 2026 Guide)

In January 2024, a customer trying to trace a missing parcel got the support chatbot of DPD, a large European delivery company, to swear at him and call DPD the “worst delivery firm in the world”. He posted the screenshots, they spread widely, and DPD told the BBC that an error after a system update was to blame and that the AI element had been immediately disabled.

Almost every business owner I speak to in Mumbai and Navi Mumbai now gets pitched an “AI chatbot” at least once a month. Some of them have heard the DPD story; most have not. This article answers the question they actually ask me: should a small Indian business put an AI chatbot in front of its customers, and if so, how do you do it without becoming the next screenshot?

1. What a 2026 chatbot actually is

There are two kinds of chatbot, and vendors often blur them. The older kind is rule-based: a menu of buttons and fixed answers (“Press 1 for timings”). It never says anything you did not write, but it cannot cope with a question phrased in a way you did not expect. The newer kind is AI-based: a large language model that reads the customer’s message in their own words and writes a reply, ideally using only information you have given it.

The AI kind is what changed the market. It can understand “bhai Sunday ko khula hai kya?” as easily as “Are you open on Sunday?”. The catch is that it writes new sentences every time, so it can also write wrong ones. Everything in the rest of this article is about getting the first benefit without the second problem.

The WhatsApp option most owners have not noticed

In May 2026, Meta announced Business AI on WhatsApp for small businesses in India. Meta says it works inside the WhatsApp Business app, is set up from the Tools tab under “Your Business AI” with no coding, learns from your business profile, catalog and documents, and can answer questions, recommend products, capture leads and book appointments in all native languages in India. Meta also says the owner can take over any conversation, change how the AI behaves, or switch it off.

WhatsApp’s own Business app features page now lists this as an AI agent (“Meta Business Agent”) that answers questions and captures leads. Two things Meta did not state in the announcement: the full eligibility criteria and whether it will stay free. Check the app and Meta’s help pages before you plan around it.

2. What customers really think of chatbots

Vendors quote satisfaction numbers from their happiest clients. The most useful independent data I have found points the other way. In July 2024, Gartner published a survey of 5,728 customers, conducted in December 2023, with two uncomfortable headline figures.

Source: Gartner press release, 9 July 2024, as reported by Insurance-Canada.ca. Global survey, not India-specific; fieldwork December 2023.

Read that last row twice. Customers were not mainly worried about AI being stupid. They were worried about being trapped. Gartner’s own recommendation was that a chatbot should say upfront that it will connect the customer to an agent if it cannot solve the problem, and that the handoff should continue the conversation rather than make the customer start again.

There is a counterweight. Meta, citing a 2025 Kantar study, says 91% of online adults in India chat with a business every week. That is Meta’s own reporting of research it commissioned or cited, so treat it as a strong signal of habit rather than a neutral measurement. Put the two together and my reading is this: Indian customers are comfortable messaging businesses, and they will accept fast automated answers, but they punish anything that blocks them from a person.

3. Which jobs to give a bot, and which to keep

The safest way to decide is to sort your incoming questions by two tests. Is the right answer the same for everyone? And does a wrong answer cost little? If both are yes, the bot can take it. If either is no, a person should.

Illustration based on my own work with Mumbai clients, not a published benchmark.

In practice, the first four rows are often most of the volume for a clinic, salon, coaching class or shop, which is why a bot can genuinely help. The last three rows are where the reputational damage happens, which is why the handoff matters more than the AI.

Lead capture is the quiet winner

The use I recommend most to service businesses is not answering questions at all. It is collecting the details your team needs to call back: name, area, what they need, budget range, preferred time. An interior designer who gets a 10 pm enquiry and a tidy summary at 9 am is in a better position than one who sees “hi, price?” and has to start from zero.

4. Where to put it: WhatsApp, website or both

For a business that serves customers in India, my default is WhatsApp first. It is where customers already message you, and a bot on WhatsApp handles real conversations rather than waiting for a website visitor to notice a chat bubble. A website widget makes sense when you get a lot of website traffic with questions that are not reaching WhatsApp, or when you sell to customers abroad.

Sources: Meta newsroom, May 2026; Meta for Developers, WhatsApp pricing; Intercom pricing, all checked October 2026. Intercom is shown as one published global example, not a recommendation; prices are in US dollars.

The Intercom example is worth a closer look because of how it defines what you pay for. Its pricing page counts an “outcome” when a customer confirms the issue is resolved, when they do not ask for more help after the bot replies, or when the bot completes a workflow, including a handoff. That is a reasonable commercial definition, but notice the middle one: a customer who gives up and leaves also “did not ask for more help”. Whatever tool you use, do not confuse a quiet customer with a satisfied one.

See also: for opt-in rules, templates and message pricing on WhatsApp, read WhatsApp marketing: how to do it right in India.

5. How chatbots go wrong, with real examples

Every failure I have seen falls into one of three patterns. Knowing them is more useful than any feature list.

They can be talked into things

In December 2023, people visiting the website of Chevrolet of Watsonville, a car dealership in California, discovered its ChatGPT-powered assistant would follow their instructions. According to Jalopnik’s report, one user got it to write a Python script, and others got it to agree to sell a new Tahoe for one dollar after telling it to call the offer binding. The deal was not legally binding, and the vendor said most manipulation attempts had failed, but the screenshots were the story.

This is called prompt injection. The OWASP Top 10 for LLM Applications 2025 lists it as the first risk, defining it as user prompts that alter the model’s behaviour or output in unintended ways. DPD’s swearing bot, mentioned at the start, was asked by the customer to disregard its rules.

They give wrong answers with total confidence

An AI model is built to produce a fluent reply. If your price list is out of date, or a question falls outside what you gave it, it may fill the gap with something plausible and wrong. Rule-based bots fail by saying “I don’t understand”; AI bots can fail by inventing a discount.

You are responsible for what they say

In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline responsible after its website chatbot gave a customer wrong information about a bereavement fare. It is not Indian law, and I am not a lawyer, but the principle is a safe one to assume: if your bot says it, your business said it.

Sources: HCAMag, 24 Jan 2024, citing DPD’s statement to the BBC; Jalopnik, 19 Dec 2023; Civil Resolution Tribunal, 2024 BCCRT 149. None are Indian cases.

6. How to set one up properly

The setup work that matters is not technical. It is writing down what your business knows, deciding what the bot may not do, and designing the exit to a person.

Step 1: Write the knowledge it will use

Put together a single document: timings including holidays, address and landmarks, services with current prices or price ranges, booking rules, delivery areas and charges, payment methods, and your refund and cancellation policy in the exact words you want customers to see. If your catalog in the WhatsApp Business app is incomplete, fix it first, because Meta says its AI uses the catalog.

Step 2: Set the limits

OWASP’s mitigation list for prompt injection is written for developers, but three items translate directly for a small business: constrain the bot to a defined role and topics, give it access only to what it needs, and require human approval for anything high-risk. In plain terms:

• It answers only about your business. Anything else gets a polite “I can only help with questions about [business]”.
• It never offers a discount, accepts a price, promises a delivery date it cannot check, or confirms a refund.
• It does not give medical, legal or financial advice, even if your business is a clinic or a CA firm. It books the appointment instead.
• If it is not sure, it says so and hands over.

Step 3: Design the handoff

Following Gartner’s advice, tell customers at the start how to reach a person, and make sure your team sees the whole bot conversation when they take over so the customer does not repeat themselves. Decide who picks up handoffs, during which hours, and what the bot says outside those hours (“Our team will reply after 10 am tomorrow. I have noted your details.”).

Step 4: Test with your real customers’ language

Meta says its WhatsApp agent supports all native Indian languages, and modern models handle Hindi, Marathi and Hinglish reasonably well. Reasonably is not perfectly. Take 20 real messages from your chat history, including the messy ones with spelling mistakes and voice-note-style typing, and check every answer.

See also: for other ways to use AI in a small business beyond customer chat, read AI automation for small businesses: where to start.

7. Customer data and consent

A chatbot collects personal data: names, phone numbers, addresses and sometimes health or financial details customers volunteer. The government’s PIB explainer on the Digital Personal Data Protection Rules, 2025 says the Rules were notified on 14 November 2025 with an eighteen-month period for phased compliance, that businesses must give a clear consent notice explaining the specific purpose for which data is collected, and that individuals can withdraw consent at any time.

For a chatbot, that means three practical things. Tell people what you collect and why, in plain language. Collect only what you need for the job (a salon does not need a date of birth to book a haircut). And know where the conversation data is stored, which vendor holds it and who on your team can see it. This is my summary, not legal advice; check your obligations with your CA or a lawyer, especially in healthcare, finance and education.

8. Measure whether it is helping

Most chatbot dashboards lead with “conversations handled”. That number goes up whether the bot is helping or annoying people. I ask clients to track four numbers instead, once a month.

• Resolved without a person: conversations where the customer got what they needed. Spot-check 20 a month by reading them.
• Handoff rate and handoff speed: how often customers ask for a person, and how long they wait. A rising handoff rate on one topic tells you what to add to the knowledge document.
• Leads and bookings: enquiries captured with full details, and how many became paying customers.
• Complaints about the bot: any customer who mentions the bot negatively, in chat or in a Google review.

If bookings do not move and complaints appear, switch it off. A bot that saves your team an hour a day but costs you two customers a month is a bad trade for most small businesses.

Where this advice breaks down

• The best data is not Indian. The Gartner survey is global and dates from December 2023. Customer attitudes in India may differ, and may have shifted since. The failure cases are from the UK, US and Canada.
• Meta’s India figures are Meta’s. The 91% weekly chat figure and the business testimonials in Meta’s announcement are company-reported. Useful as signals, not as proof a bot will work for you.
• Very small businesses may not need one. If you get 15 messages a day and the owner answers them happily, quick replies and a greeting message in the WhatsApp Business app may be all you need.
• High-volume businesses outgrow simple setups. An e-commerce brand with thousands of daily order queries needs a bot connected to order data, proper monitoring and someone accountable for it. That is a software project, not a settings toggle.
• Tools change monthly. Meta’s WhatsApp AI was renamed within months of launch. Treat every product detail here as accurate for October 2026 and check before you commit.

A chatbot checklist before you switch it on

DPD’s bot did not fail because AI is useless. It failed because nobody was checking what it said after an update. Check yours, keep the human one tap away, and a chatbot can be a quiet, useful member of your team.