Set up WhatsApp AI agents

• One agent answers each message. LarkFlow works out what the customer wants and picks the best enabled agent for it.
• Agents only use your data — your knowledge base, catalogue, orders and availability. They are instructed never to invent prices, stock, order status or appointment times.
• Some messages always go to a person. Before the AI replies, LarkFlow checks for requests to speak to a human, refunds, disputes or fraud claims, legal threats and abusive language — these are handed to your team instead.
• Your team can take over at any time. Once a conversation is assigned to a teammate, the agents stop replying in it.
• AI replies use your plan’s monthly AI actions.

Open AI agents → Knowledge. The more accurate your knowledge, the better the answers. Add any mix of:

1. 1

   Step 1: Your website

   Enter your site address (e.g. https://yourbusiness.com) and start the crawl. You’ll see Crawling started; pages are added as sources. Use Re-crawl after you update your site.

2. 2

   Step 2: Business details

   Add a title and write the essentials your website may not cover — opening hours, delivery areas, payment terms, return policy, common questions — then Save.

3. 3

   Step 3: Documents

   Under Documents (PDF, DOCX), upload price lists, brochures, catalogues or policies. You’ll see Document uploaded, processing… while it’s read.

Go to the Agents tab. Switch on the agents you want:

Open an agent to fine-tune it, then click Save configuration:

• Role & responsibilities — what this agent is for, e.g. “Answer questions about our cleaning services and pricing”.
• Tone — e.g. “Friendly and professional”.
• Max response length (chars) — keep WhatsApp replies short.
• Response style — e.g. “Short, use bullet points for lists”.
• Escalation rules — extra situations to hand to your team, e.g. “Escalate pricing negotiations and complaints”.
• Custom instructions — anything else the agent must always or never do.

To add a custom agent, enter an Agent name, click Create, and describe its role. Its behaviour is defined entirely by what you write.

• Sales searches your product catalogue for names, prices and stock. Add products in LarkFlow or sync them from Shopify, WooCommerce, Odoo, Zoho, Business Central or a custom REST endpoint (RFQ → Products / ERP).
• Commerce reads the customer’s orders in Ecommerce, matched by their WhatsApp number.
• Service reads your booking availability — next step.

1. 1

   Step 1: Open booking settings

   Go to Settings → Booking and click Add availability.

2. 2

   Step 2: Add a weekly window

   Choose a Service (or leave All services), the Day of week, Start time and End time, the Slot length (minutes) and Capacity per slot (how many bookings one slot can take). Repeat for each day you’re open.

3. 3

   Step 3: Check your timezone

   Times use your company’s timezone from Company settings — set it before adding windows.

The Service agent offers customers a few of these real open slots. When the customer clearly confirms one, LarkFlow re-checks it is still free and creates the booking, and the customer gets a confirmation. If the slot has just been taken, the customer is asked to choose another. Bookings appear under Upcoming bookings, where your team can cancel them.

Open the Test Agent tab. Two ways to test your draft (unpublished) setup:

1. 1

   Step 1: Simulate a conversation

   Type customer messages under Simulate a conversation. Nothing is sent to a real customer, but it does use a real AI call. Use Reset conversation memory to start over.

2. 2

   Step 2: Add test numbers

   Add your own WhatsApp numbers under Test numbers. Messages from these numbers use your latest draft even before you publish — so you can try it on a real phone.

On the Overview tab, click Publish. Publishing can take a few minutes. Customers already mid-conversation keep the current version until the new one is active. Every time you change knowledge or agent settings, test and publish again.

Track results in Analytics: customers, AI messages, human handoffs, handoff rate, failed responses and estimated cost.

On the RFQ Agent tab:

• Enable RFQ detection — recognises purchase requests in WhatsApp messages, extracts the items and drafts a quotation.
• Auto-send quotation via WhatsApp — off by default. When on, the quotation is sent once the customer has confirmed the exact product. Manual sending from the quote page always works.
• Follow up after no response (hours) — sends one automatic follow-up if a quotation gets no reply. Set to 0 to turn off.

When a WhatsApp message arrives, LarkFlow tries these in order and stops at the first one that handles it:

1. 1

   Step 1: A flow waiting for this reply

   If an automation flow asked the customer a question, their answer goes back to that flow.

2. 2

   Step 2: Automation flows

   An active flow whose trigger matches (e.g. a keyword) runs. See automation flows.

3. 3

   Step 3: AI agents

   If no flow ran and no teammate is assigned to the conversation, the agents reply.

4. 4

   Step 4: AI Bot (basic auto-reply)

   If the agent team didn’t reply and AI Bot → Auto-reply is on, the simpler auto-reply answers.

RFQ detection runs alongside all of these when it’s enabled.

Agents don’t reply at all

Check you’ve clicked Publish, at least one agent is switched on, the conversation isn’t assigned to a teammate, and no automation flow is catching the message first.

Answers are vague or wrong

Add the missing facts under Knowledge (business details are quickest), re-crawl your website, test, then publish again.

Sales agent doesn’t know prices or stock

It only uses your catalogue — add or sync products first.

Booking agent never offers times

Add availability in Settings → Booking and check your company timezone.

“Could not publish”

Read the error shown, fix the agent it mentions, and publish again. Wait for an earlier publish to finish before starting another.