Blog / Automation & AI

AI that answers customers on WhatsApp: how to train it on your content

“Press 1 for hours, 2 for the price list”: the old menu chatbots are the reason many people distrust automated replies. But what you can put on WhatsApp today is something else: an AI that has read your content and answers the way a well-prepared employee would — to the real question, written the way a real customer writes it, with the abbreviations and the typos. What makes the difference between the two experiences is the training: what you give it to read, what you let it do, and when you make it step aside.

What a trained AI is (and what changes versus the button-driven chatbot)

The traditional chatbot follows a tree of options: if the customer leaves the expected path, it gets stuck. A conversational AI, instead, understands the question and looks for the answer in the materials it was trained on: your website pages, the PDFs of your catalogs, the price lists, the terms of sale. It answers “how much is shipping to Sicily?” even if no one ever planned that button, and it does so in the language the customer writes in. The limit is just as clear: the AI knows only what you’ve given it. The quality of the answers is the quality of the materials — this whole article, in the end, is about that.

Step 1 — Gather the right content

Training starts with an inventory. You don’t need everything: you need what customers actually ask.

• The website: product and service pages, FAQs, shipping and return terms, contacts and hours
• The PDFs: catalogs, brochures, technical sheets, manuals — everything you send as an attachment today
• The price lists: up-to-date prices, with the exceptions spelled out (variants, surcharges, ongoing promotions)
• The answers you already give: ask whoever runs the chat for the twenty most frequent questions and the answers that work — it’s the most valuable material of all
• The house rules: what the AI can promise and what it can’t (discounts, delivery times, exceptions)

A good test before you begin: take your last fifty real conversations and check that the answer to each one exists in the materials you’ve gathered. Where it’s missing, write it: you’re building the knowledge base, and it’s work that pays off even if the AI never used it.

Step 2 — Train, test, correct

Once the materials are loaded — website links, PDFs, documents — the AI is up and running in no time. But up and running doesn’t mean ready for customers: first it has to be put to the test the way a difficult customer would. Ask it the tricky questions: abbreviations, dialect words, double questions (“do you have model X and how long does it take to arrive?”), borderline requests (“can you give it to me for less?”). Where it gets it wrong or makes things up, correct the source material: almost always the error comes from missing or ambiguous information in the documents, not from a whim of the model. And plan for updates: every time prices, hours, or terms change, the training must be aligned the same day.

Step 3 — Decide what to delegate (and what not to)

The dividing line is simple: the AI excels where the answer is already written down; the agent is needed where the answer has to be built or where there’s emotion involved. An angry customer who gets polite but automated replies gets angrier: that’s the human moment, and the AI must know how to recognize it.

Step 4 — Escalation to an agent (the piece that can’t be missing)

Every AI in production must have a clear way out to a person. There are three ingredients. The recognition of cases to pass on: the explicit request (“I want to speak to a person”), sensitive topics, mounting frustration, and questions with no answer in the materials. The clean handoff: the AI pauses on that conversation and notifies the team, and the customer knows a colleague is coming — no loop of evasive replies. And the coverage: someone who picks up the conversation within a reasonable time, even just to say they’ll reply tomorrow morning. An AI that admits “I’ll have a colleague answer this for you” builds more trust than one that improvises just to answer.

Best practices: the first thirty days

Start in supervised mode: the AI replies, but the team reviews the conversations and notes where it hesitates or gets it wrong. Keep the initial scope narrow — the most frequent informational questions — and widen it only when the answers are solid. Always state that it’s an automated assistant: customers happily accept it if the answers are useful and instant, much less so if they find out they’ve been deceived. And appoint a training owner: a person who updates the materials when something changes, because an abandoned knowledge base goes stale in a few weeks.

Common mistakes (up close)

• Training it on the whole website, including old pages: the AI cites the promo from two years ago because no one removed it from the materials
• Outdated price lists: a wrong price stated with confidence is worse than “I don’t know”
• Passing it off as human: sooner or later it shows, and lost trust doesn’t come back
• No escalation: the customer trapped in the bot is the customer who doesn’t come back
• Turning it on everywhere from day one: better to start from a scope, measure, and widen
• Not re-reading the conversations: the first week must be reviewed carefully — that’s where you find what’s missing in the training

With SendApp, the Agent AI is trained on website links, PDFs, and documents, answers on WhatsApp and the other connected channels in the customer’s language, and passes the conversation to an agent when needed: you see everything from a single inbox and refine the training as you go.