Train your AI brain
A menu answers the questions you predicted. The AI brain answers the ones you did not — but only as well as the knowledge and the example phrases you give it.
Add your own Gemini or OpenAI key, fill in store knowledge (shipping, returns, sizing, payment), then create intents with five to ten phrases each in the words customers actually use. Test each intent for confidence before activating it — an inactive intent never matches in production.
Before you start
- WhatsApp connected, for live replies.
- A Gemini or OpenAI API key. Keyword-only intents work without one.
- A published flow as the fallback router — Build your first flow.
- Fifty real inbound messages to read before you start writing phrases.
Step 1 — load the knowledge
The AI cannot invent your return window. Before any intent work, fill in store knowledge, because this is what grounds every generated answer and it is the difference between a useful assistant and a confident fabricator.
- Shipping — dispatch time, delivery time by region, courier partners, what happens to a delayed order.
- Returns and exchanges — the window, the condition requirement, who pays return shipping, how refunds are issued.
- Payments — which methods you accept, whether COD is available and where, any COD fee.
- Sizing and product care — the questions your category always gets.
- What you do not do — no international shipping, no exchanges on sale items. Knowing the limits stops invented answers.
Step 2 — set the persona
Intelligence → AI → Bot personality sets the assistant’s name and tone. Keep it plain and close to how your brand already writes. A bot with a wacky persona ages badly and makes complaints read worse than they are.
Step 3 — train intents on real phrases
- Read your inbox first Open the last fifty inbound messages and group them. The groups are your intents, and the messages are your phrases.
- Create the intent Intelligence → Intents → Add intent. Name it for the question, such as "Shipping time".
- Add five to ten real phrases Copy how customers actually write — Hinglish, abbreviations, no punctuation, typos. "kitne din me ayega" is a better training phrase than "What is your delivery time?". Phrases you invented yourself are the main cause of low confidence scores.
- Choose what it does Reply with text, start a flow, assign to an agent, or escalate. Not everything should be answered by the AI — "I want a refund" is better routed to a human.
- Activate it Toggle Active. An inactive intent is invisible in production no matter how well trained it is.
- Test it Use the test tool to type sample messages and read the matched intent and its confidence. Low or wrong matches mean more phrases, or splitting the intent.
| What you see | What to do |
|---|---|
| Right intent, high confidence | Nothing. Move on. |
| Right intent, low confidence | Add three to five more phrases closer to how the message was worded. |
| Wrong intent matched | Two intents are competing. Split them into narrower ones, or deactivate the duplicate. |
| No match at all | Expected for genuinely new questions. Make sure your flow fallback hands these to a human. |
How to check it worked
- Your AI key is saved and the knowledge tabs accept input.
- The test tool matches your sample phrases to the right intents.
- Each intent you want live is toggled Active.
- A real inbound message gets an AI reply when no agent has taken over.
- An escalation intent puts the thread in the agent queue.
If the AI answers badly
- Test matches the wrong intent or reports low confidence
- Add three to five more phrases in customers’ own words.
- Split broad intents — "Shipping" and "Returns" should not be one intent.
- Deactivate duplicate intents competing for the same phrases.
- The AI answers but the facts are wrong
- Fill the gap in store knowledge — an empty field is where invention happens.
- Add the limits explicitly, including what you do not offer.
- Change that intent to a fixed reply instead of a generated one.
- The AI never replies at all
- Check the intents are Active.
- Check no agent has taken over the thread in Live Chat.
- Confirm your AI key is saved and valid with its provider.
- Confirm a flow is Published as the fallback router.
- AI brain reference — Intents, knowledge and persona.
- Build your first flow — The fallback router.
- AI WhatsApp chatbot guide — What to automate and what not to.
Frequently asked questions
Do I need my own OpenAI or Gemini key?
For AI-generated replies, yes. TopEdge uses your own key, so usage is billed to you by that provider and you keep control of which model answers your customers. Keyword-matched intents with fixed replies work without any key.
How many example phrases does an intent need?
Five to ten, taken from real customer messages rather than written by you. Phrases in your own phrasing score poorly because customers do not write like the person who built the store — expect Hinglish, abbreviations and no punctuation.
Should the AI handle refunds and complaints?
No. Route anything involving money, a complaint or a legal question to a human. A confidently wrong answer about a refund costs far more than the time it saved, and customers escalate it publicly.
Updated 2026-10-04