AI Agent

The AI Agent is an auto-reply assistant for your WhatsApp inbox. When a customer messages you, it reads their question, checks what it knows about your business, and either replies automatically or hands the conversation to a human. Done well, it absorbs the repetitive questions — hours, pricing, order status, returns policy — so your team only touches the conversations that need them.

It lives under AI Agent in the sidebar and has four tabs: Settings, Knowledge base, Test playground, and Auto-reply.

How it decides what to do

For every inbound message the agent produces an answer and a confidence score. You set the rules for what happens next:

• If confidence is above your threshold, it sends the reply.
• If confidence is below it, or the customer trips a handoff condition, it stops and hands off to a human instead of guessing.

This is the core safety idea: the agent should say "let me get a person" rather than invent an answer.

Step 1 — Build the knowledge base

The agent only knows what you tell it. Go to Knowledge base and add material. There are three ways to feed it:

• Add an FAQ — a question customers ask and the answer the bot should give. This is the highest-signal format; start here.
• Add a web page — paste a page URL (your pricing page, policies, about page) and the agent ingests its content.
• Add a knowledge snippet — paste raw content: a policy, an about section, shipping terms, any reference text.

Step 2 — Configure the agent

Open Settings and tune behaviour:

• Persona — sets the bot's identity and tone (who it is, how it speaks). Match your brand voice.
• Confidence threshold — how sure the agent must be before it replies on its own. Higher = more cautious (more handoffs, fewer wrong answers). Start conservative and lower it as you trust it.
• Extra instructions (optional) — house rules: things to always or never say, formatting preferences, languages to use.
• Handoff keywords — words that force a human handoff immediately (e.g. refund, complaint, human, agent).
• Handoff message — what the customer sees when the agent passes them to a person.
• Max auto-replies before handoff — a cap on how many times the agent answers in one conversation before it brings in a human, so it never loops forever.

Step 3 — Test before going live

Use the Test playground to rehearse safely. Send the agent a message and you'll see:

• The preview of how it would respond,
• Its confidence for that answer,
• Which knowledge it used,
• Whether it would hand off to a human.

Crucially, nothing in the playground is sent to customers — it's a sandbox. Probe it with your trickiest real questions and edge cases. If answers are wrong, add or fix knowledge; if it's too eager or too timid, adjust the confidence threshold.

Step 4 — Turn on auto-reply

When you're happy with the playground results, enable the agent on the Auto-reply tab. From then on it responds to inbound messages within the rules you set. Conversations it hands off appear in the shared inbox for your team, with context intact.

A good rollout plan

1. Load 15–20 FAQs and your key policy pages.
2. Set a high confidence threshold and generous handoff keywords.
3. Hammer the playground with real questions.
4. Turn on auto-reply and watch the inbox for the first few days.
5. Gradually lower the threshold as accuracy proves out, adding knowledge wherever it guessed wrong.

Next steps