How to Use Trengo to Answer WhatsApp With an AI Agent

The short answer: if most of your WhatsApp questions are about orders, returns and opening hours, run Trengo's AI Agent in Supervised Mode first: it drafts every reply for a person to approve, it carries no AI surcharge, and its 30-day window covers a full month of real questions (Supervised Mode). On Autopilot, each conversation the agent handles adds €0.30 on monthly billing (checked on 26 October 2026 on the pricing page). The steps: connect the number, build the knowledge, write behaviour and handover rules, test in Preview, supervise for a month, then switch to Autopilot with an AI notice in the first message.

Disclosure: I may earn a commission through some links on this site. It does not change what this page says.

Three numbers before you start:

1. What Autopilot adds to the bill. For example, 400 WhatsApp conversations a month with half closed by the agent is 200 x €0.30 = €60 a month on top of Boost's €349, using the pricing page. Meta's own charge comes on top: since 1 October 2026 each service message past the first 1,000 per number a month is billable (WhatsApp pricing change).
2. The answers you already have. Your returns policy, delivery times and FAQ pages are the agent's knowledge base on day one. Nothing needs writing from scratch; it needs splitting into one topic per snippet.
3. The two rates to watch. The AI Agent dashboard reports a resolution rate (closed by the agent alone) and an assistance rate (handed to a person). Note your current first-response time now, so you have something to compare them with.

What you need

• A Trengo account. The free trial lasts 7 days with no card (trial page), and the plans are compared on the pricing page.
• A Facebook account and a Meta Business account you can verify within 30 days, and a phone number that meets Meta's conditions (WhatsApp setup).

Step 1: Connect the WhatsApp number

In a web browser, not the desktop app, go to Settings > WhatsApp Business, click "Connect your number", accept the requirements and follow the "Continue with Facebook" flow, per the WhatsApp setup article. Until your Facebook Business account is verified, the number runs in Meta's trial phase. A number on the API cannot be used in the WhatsApp Business App any more, unless you connect it through Coexistence, which keeps the app working beside Trengo for a number already active on it (Coexistence).

Step 2: Build the knowledge base

Add snippets for the answers that must be exact, such as returns, refunds and shipping, one topic per snippet. Add your website as a supporting source, upload PDF, TXT or CSV files, or connect your help centre (knowledge base article). Trengo's advice is to use snippets for critical or sensitive information and websites only as a secondary source. Without connected knowledge, the agent cannot answer knowledge questions at all.

Step 3: Write the behaviour and the handover rules

The agent reads three layers of instruction, per Instruct your AI Agent:

• Behaviour: a job description, your company context and a tone of voice, with what to avoid.
• Handover criteria and additional instructions: by default the agent escalates when it cannot answer or when the customer asks for a person. Add your own, such as payment disputes, and list the details to collect before handover, such as order number and email; it asks only for what is missing.
• Skills: step-by-step processes for one situation, such as an order-status question. A Skill overrides the other layers when it applies.

Put the AI notice in additional instructions now, so it is there when Autopilot starts: "Begin your first reply with: You are chatting with an AI assistant from [brand]; type 'person' at any time to reach the team." Article 50 of the EU AI Act, which applies from 2 August 2026, requires that people are informed they are interacting with an AI system at the latest at the first interaction, unless it is obvious (Regulation (EU) 2024/1689 on EUR-Lex, Articles 50 and 113).

Step 4: Configure and connect the channel

On the configuration screen, select the knowledge sources, then connect the WhatsApp channel (configuration article). Each channel takes one AI Agent, and a channel already used in an AI Journey should not be connected here as well. Set a default language for messages the agent cannot detect, switch on "Generate summary on escalation" so the person taking over sees the context, and set the reminder and auto-close times; the reminder must come before the auto-close.

Step 5: Test it in Preview

Run whole conversations, not single questions: rephrase the same request, send incomplete messages, attach a photo of a damaged product, and ask for a person (Preview article). Preview shows the actions, such as "Handed over to [team]", so you can confirm the handover reaches the right team with the details collected.

Step 6: Supervise for a month

Go to Trengo AI > AI Agent, select the agent and set its mode to Supervised. The agent drafts each reply and proposes escalations, closes and labels; a team member approves, edits or assigns, and anything sent goes out under that person's name (Supervised Mode). This is the human check, and while it runs every message to the customer is written or approved by a person. The 30-day clock is per account, starts the first time anyone switches the mode on, and does not pause when you switch it off. Keep a tally of drafts approved unchanged against drafts edited.

Step 7: Switch to Autopilot and read the dashboard

When the tally says the drafts are good enough to send unchanged, change the mode to Autopilot. From here the agent talks to customers directly, which is why the notice from step 3 matters now. The dashboard tracks only Autopilot agents, with up to 30 days of data: resolution rate, assistance rate, and the average number of messages and time before each outcome.

What Trengo's articles warn about

• The agent knows only what you connect. It cannot reach Shopify, Notion or any other system unless the integration is set up, and it cannot read video messages (abilities and limits).
• Old knowledge gives wrong answers. The knowledge base article asks for regular reviews, because outdated information leads to incorrect replies.
• Several short replies cost more than one. Meta now bills service messages one by one, so Trengo suggests combining replies and turning off "Send customer replies on Enter" (WhatsApp pricing change).

The balance that pays Meta's charges and the AI surcharges tops itself up automatically once it is empty, and the top-up adds an estimate for the rest of the billing cycle, up to €500 of usable balance (plan and pricing article). That default keeps the agent answering through a busy week and moves the spending decision from you to the meter. Check Settings > Transactions weekly in the first month on Autopilot.

The free version of this test

Before any of this, the tally from step 6 can be run by hand: take 100 recent WhatsApp questions from the inbox, sort them into the topics your snippets cover, and count how many a snippet answers word for word. If that share is small, the knowledge base needs work before any AI does. The Trengo glossary entry covers plans and add-ons.

FAQ

Does Supervised Mode use up AI conversations?

No. Trengo applies no AI surcharge while the agent is supervised, and the customer only ever sees the human's name on replies (Supervised Mode).

Can I extend the 30 days?

Yes, by asking your Trengo customer success manager. When the period ends, the agent stops drafting and its configuration is kept (Supervised Mode).

Is Supervised Mode available for every agent?

No. It is not available for Voice agents or for AI Agents embedded in an AI Journey, so test those in Preview before they go live (Supervised Mode).