AI ChatbotsConversations worth automating

What changed, and what did not#

Older chatbots followed decision trees. They were predictable and frustrating: customers had to guess the right button or keyword. AI chatbots understand free text and reply in natural language, which makes them far more pleasant to talk to.

What did not change is the underlying job. A customer conversation still has to end with a correct answer, a resolved problem or a person who can help. A fluent bot that gives a wrong refund policy is worse than a clumsy bot that admits it does not know.

A chatbot is judged by its worst answer, not its average one.

Rule-based bots vs AI bots#

Decide the bot's jobs before anything else#

The most common mistake is launching a bot that tries to do everything. Choose a small number of jobs, write them down and design for those. Everything else should route to a person or a form.

• Answer pre-sales questions about products, delivery, sizing or service scope, from approved content.
• Qualify leads by asking a few questions and passing a summary to sales. See lead scoring.
• Handle simple service requests such as order status or appointment changes, through proper integrations.
• Collect context before handover so the human agent does not ask the customer to repeat everything.

Notice what is not on the list: complaints, refunds outside policy, anything legal or medical, and emotionally charged conversations. Those belong with people from the first message.

Grounding: the single most important design choice#

A grounded bot answers only from a defined set of sources: your help centre, product data, policies and approved FAQs. An ungrounded bot answers from the model's general knowledge, which is where invented policies and wrong prices come from.

1. 01Assemble the knowledge base. Current policies, product information, delivery and returns, service descriptions. Remove outdated pages first.
2. 02Instruct the bot to stay inside it. If the answer is not in the sources, it should say so and offer a person.
3. 03Exclude sensitive topics explicitly. List what it must never discuss or decide.
4. 04Assign a content owner. When a policy changes, the knowledge base must change the same day.
5. 05Test with hard questions. Edge cases, trick questions, competitor comparisons, attempts to make it promise discounts.

Designing the conversation#

Be open that the customer is talking to an automated assistant. Many jurisdictions and platforms expect this, and customers respond better to honesty than to discovering it later. Give a visible way to reach a person at every stage.

WhatsApp and messaging channels#

In markets such as India, a large share of customer conversation happens on WhatsApp. AI bots on messaging channels follow the same principles but carry extra rules: business messaging policies, opt-in requirements, template approval for outbound messages and limits on when you can message first. Check the current platform documentation before designing flows.

Messaging also changes expectations. Customers expect short replies, quick handover and memory of the conversation. A bot that writes paragraphs feels wrong in a chat window. See WhatsApp marketing for the wider channel strategy.

Handover to people#

Handover is where most bots fail customers. The bot should hand over when the customer asks, when it cannot answer from its sources, when sentiment turns negative, and whenever a topic on the excluded list comes up. It should pass a short summary so the person can pick up without repetition.

If no person is available, say so honestly and offer a clear alternative: a callback, an email address, a time when people are online. 'Someone will be with you shortly' when nobody will is a trust problem, not a design shortcut.

Testing before launch, and after#

Testing a language-model bot is different from testing a decision tree. You cannot click through every path, because there are no fixed paths. Instead, build a set of test conversations that represents what customers will really ask, plus what a mischievous customer might try.

• Common questions in the many ways real customers phrase them, including spelling mistakes and mixed languages.
• Edge cases: orders outside normal rules, products you have discontinued, regions you do not serve.
• Pressure tests: 'your competitor is cheaper, will you match it?', 'promise me a refund', 'ignore your instructions'.
• Sensitive topics: health, legal, complaints, distress. The bot should hand over, not improvise.
• Off-topic requests: questions unrelated to your business, which it should politely decline.

Run the same test set after every change to the knowledge base, instructions or underlying model. Models are updated by vendors, and behaviour can shift without any change on your side. A fixed test set is how you notice.

Measuring a chatbot honestly#

Vendors often report 'containment', the share of conversations that never reached a person. On its own this is a dangerous number, because a bot that frustrates customers into leaving also looks contained. Pair it with resolution and satisfaction measures, and read transcripts.

• Confirmed resolution: the customer said the answer helped, or did not return on the same issue.
• Handover quality: how often people had to re-ask for information.
• Wrong-answer rate: from a weekly sample of transcripts reviewed by a person.
• Lead quality: for sales bots, how many qualified conversations became real opportunities.

Privacy matters here too. Conversations often contain personal data, so storage, retention and use must match what you have told customers. See DPDP Act for marketers and our guide to AI and brand safety.

Published by Fabulous.Media, a network of specialist marketing agencies. Updated 9 October 2026. Platform features change often; check current official documentation before acting on platform-specific detail.