Build AI Agents for WhatsApp | Tutorial

Important Terminologies At a Glance

The seven core concepts that power every AI Agent you build on AiSensy.

AI Agent

A WhatsApp chatbot that understands intent in natural language, learns from your business data, and takes real actions through APIs — without keywords, scripts, or a single line of code.

Knowledge Base

The memory of your AI Agent. Train it on websites, articles, FAQs, or PDFs, and it answers customer queries using your real business content instead of generic AI replies.

AI Orchestrator

The brain of your AI Agent. Define your business context, tone, workflow rules, and restrictions so the agent knows exactly how to think, behave, and respond on every conversation.

Tool Calling

The hands of your AI Agent. Connect any REST API, CRM, or eCommerce platform to fetch live data, book appointments, look up orders, and reply with formatted WhatsApp messages.

Ai Routing

The traffic controller of your account. Reads every incoming message, classifies user intent, and routes the conversation to the most relevant flow — even when no keyword matches.

Confidence Score

A 0 to 1 rating the AI assigns to every routing decision. Above 0.95 is an exact match, below 0.75 it won't route — keeping your agent accurate instead of guessing.

AI Messages

The credits that fuel every intelligent response. Each KB answer, tool call, or Orchestrator reply consumes one message from your monthly pack — track and top up anytime.

How AI Agent Builder Works

Every AI Agent built on AiSensy runs on two layers working together to deliver one seamless WhatsApp experience.

How to set up the Knowledge Base

The Knowledge Base is where you teach your AI Agent everything it needs to know about your business, without writing a single line of code.

You can create multiple knowledge bases inside a single flow and toggle each one on or off as needed. Here's how to set one up.

1. Login to your AiSensy account

Once you log in, you'll land on AiSensy dashboard, your home screen for everything WhatsApp Business API. Here, you'll see your API status, quality rating, remaining quota, current plan, and quick links.

2. Click on Flows in the left sidebar

Click the Flows icon in the left navigation rail. This opens the Flow Builder index page, your home for all chatbot flows and the new AI Agent capabilities.

On this page you'll see the Active Flows card (tracks how many flows are live and how many are AI-Enabled), the AI Messages card (where you purchase the AI Messages addon that fuels intelligent responses), tabs for Your Flows | Templates | AI Routing | Test Number, and a + Create Flow button.

3. Click "+ Create Flow"

Click on + Create Flow to start a new flow.

4. Open the Knowledge Base tab

Once a flow is open, four tabs appear at the top:

• BUILDER: the familiar drag-and-drop canvas where you design conversation flows.

• KNOWLEDGE BASE: train this flow's AI on your business data.

• TOOL CALLING: configure API integrations the AI can trigger.

• AI ORCHESTRATOR: tell the AI how to behave with structured instructions.

Click on KNOWLEDGE BASE tab.

2. Click + Knowledge Base

Click the + Knowledge Base button on the top-right to create a new one. Give it a clear name like "Product FAQs" or "Pricing Policies" so you can identify it later.

3. Add your content

Expand your new knowledge base by clicking the chevron. You'll see a Content counter (showing the number of sources added) and an Add content section with three options:

• 🌐 Sync content from a website: Paste your URL and AiSensy will auto-discover and crawl every child page on your site. Each discovered page is processed and indexed as a separate document.

•📄 Create a new article: Write content directly inside AiSensy. Great for FAQs, policies, or any custom information you want the agent to know.

•📎Upload a document: Drop in PDFs, DOCX files, product catalogues, brochures, or any policy documents. Text is extracted and indexed automatically.

Click the + next to the option you want and add your source.

4. Wait for indexing

Once added, AiSensy processes the data in the background. Each source goes through three statuses:

• Processing: content is being parsed and indexed. Not yet available for retrieval.

• Completed: content is fully indexed and ready for the AI to use.

• Failed: processing hit an error. Check the error message and retry.

You'll see the source count update under Content once your knowledge base is ready to use.

Note: When a user asks a question, AiSensy searches your knowledge base using three complementary methods: semantic search (matches meaning), keyword search (matches exact terms), and topic search (matches intent). Results are combined and ranked to surface the most relevant content.

5. Toggle it live

Use the toggle next to your knowledge base to enable or disable it anytime. Use the edit pencil to rename it. Use the trash icon to delete it. Click the chevron to expand and add more sources.

You'll see a success toast at the top: "Knowledge base updated successfully."

That's it. Your AI Agent can now answer questions using your actual business content!

How to set up Tool Calling (Optional)

A great agent doesn't just answer questions, it gets things done. Tool Calling lets your AI Agent trigger API integrations to fetch live data, book appointments, check order status, update CRMs, and send the result back as a properly formatted WhatsApp message.

Follow the steps below to configure your first tool.

1. Open the Tool Calling tab

Inside the flow editor, click the TOOL CALLING tab. The page header reads Tool Calling with the subtitle "Configure API integrations triggered by specific keywords in user messages."

If you haven't added any tools yet, you'll see the empty state with a centered + Add New Tool button. You can also click the + Add New Tool button on the top-right at any time.

Tool Calling empty state, "No tools added yet" with hammer-and-wrench icon, body text "Add tools to allow your AI to perform actions like searching databases or sending emails via API", and a centered + Add New Tool button

2. Name your tool and add a description

Give your tool a clear, intent-based name like "Calendly", "Order Status", or "CRM Lookup".

In the Tool Description field, describe what the tool does in plain English. The AI uses this description to decide when to call it and to extract the right parameters from the user's message. Be specific.

For example: "Fetches available time slots from Calendly for the next 7 days."

2. Name your tool and add a description.png [ SCREENSHOT: Tool Calling expanded form, Tool name field with green lightning icon in header, ▶ Test / Save / Cancel buttons; Tool Description textarea; API Endpoint with GET dropdown and URL field; Headers table with Key/Value columns; Request Fields with Auto-detect button ]

3. Configure the API Endpoint

Choose the request type from the dropdown, GET, POST, PUT, PATCH, or DELETE, and paste your API endpoint URL.

You can use dynamic placeholders like {book_call} to pass user data into the call.

4. Add Headers

Under HEADERS, click + Add Headers and enter any required headers, like authentication tokens or Content-Type: application/json.

5. Add Request Fields

Under REQUEST FIELDS, click 🪄 Auto-detect to let AiSensy auto-fill the fields from your endpoint, or click + Add Field to add them manually.

Specify each field's name and type. These are the values your tool sends to your API. The AI will automatically extract them from the conversation when the tool is called.

6. Map the API response to attributes

Under RESPONSE ATTRIBUTE MAPPING, click + Add Mapping to capture specific values from the API response and save them as attributes.

For example, you can map response.policy_number to a policy_number attribute, then use that attribute later in the flow.

6. Map the API response to attributes.png

7. Configure how the response shows up on WhatsApp

Under RESPONSE CONFIGURATION, choose your WhatsApp Message Type:

• Text Message: A simple text reply.

• Text With Buttons: Text plus interactive quick reply buttons.

• Media With Buttons: Image or document plus quick reply buttons.

• List Message: A WhatsApp List for users to choose from multiple options.

SCREENSHOT: Tool Calling Response Configuration, WhatsApp Message Type picker with four options (Text Message, Text With Buttons, Media With Buttons, List Message), Header / Body / Footer fields with Insert Variable buttons, and a WhatsApp message preview on the right

8. Fill in the message content

• HEADER (OPTIONAL): A short top line. Click {} Insert Variable to add response values. Example: 🚗 Policy: {{policy.vehicle_number}}.

• BODY: The main message. Use {{field}} syntax or click Insert Variable to pull in mapped response values. Example: Policy No: {{policy.policy_number}}.

• FOOTER (OPTIONAL): A short bottom line. Example: {{policy.days_remaining}} days remaining.

A live WhatsApp message preview on the right shows exactly how your message will look.

7. Configure how the response shows up on WhatsApp.png

9. Test the tool

Click the ▶ Test button at the top to validate your tool configuration. This makes sure your endpoint works and your variables resolve correctly.

10. Save the tool

Hit Save to activate the tool. Your AI Agent will now call this API whenever the conversation calls for it, and reply with the WhatsApp message you've designed.

Note: Tool Calling supports any REST API. If you can call it from Postman, you can call it from your AI Agent.

How to set up the AI Orchestrator

The AI Orchestrator is where your agent really comes alive. It reads the conversation, considers the user's history, weighs the available knowledge and tools, and decides the best next step.

Follow the steps below to set up your AI Orchestrator.

1. Open the "AI Orchestrator" tab

Inside the flow editor, click the AI ORCHESTRATOR tab. You'll see the Agent instructions screen with the subtitle "Tell your bot how to behave. Fill in the sections below, or describe your goal and let us draft them from your knowledge base and tools."

SCREENSHOT: AI Orchestrator full view, Agent instructions header, Sections / Raw / Undo tabs, Generate / Templates / Save buttons, five sections each with their icon and 0 chars counter: About your business, Tone & personality, How the bot should work, What to avoid, Anything else

2. Choose "Sections" or "Raw" mode

Use Sections mode for a structured, guided setup with separate fields for each part of your agent's brain. Switch to Raw if you'd rather write the full prompt yourself.

Tip: Switch to Raw mode to insert knowledge bases or tools directly using the / slash menu.

3. Fill in "About your business"

Tell the bot what your business does, who it serves, and the main things it should know about.

For example:

"We are a motor insurance provider in India. Customers reach out for policy lookups, quotes, purchases, KYC, and claims. Most users buy two-wheeler and four-wheeler policies."

4. Fill in Tone & personality

How should the bot sound? Friendly, formal, concise, playful, be specific.

For example:

"Friendly and professional. Greet returning users by name when available. Use simple Hinglish-friendly English. Acknowledge frustration before offering solutions."

5. Fill in "How the bot should work"

Step-by-step rules for the most important things the bot does. Reference your knowledge bases and tools where the bot needs them.

For example:

"1. For renewals, look up the user's existing policy in {{knowledge_base_<id>}} first.

2. Always confirm vehicle details before quoting.

3. Use {{tool_call_<id>}} to fetch the latest premium.

4. Never submit an application without completing KYC."

6. Fill in "What to avoid"

Topics the bot must refuse or hand off to a human. Be concrete about scope.

For example:

"Only motor insurance. Do not answer questions about health, life, or travel insurance. For complex disputes, escalate to a human agent."

7. Fill in "Anything else"

Other rules. Pricing display, mandatory disclosures, escalation triggers, response length, etc.

For example:

"Always show prices in INR with the GST line item. If claim amount > ₹5 lakh, escalate immediately. Keep replies under 80 words unless the user asks for more detail."

8. Use Generate or Templates

Don't want to start from scratch? Click Generate to let AiSensy auto-draft your instructions from your knowledge base and tools. Or pick a ready-made starter from Templates.

Pro Tip: Keep your instructions focused on your business domain. AiSensy already handles general conversational rules, formatting, and content safety, so you don't need to repeat them.

9. Save and go live

Hit Save, and your AI Orchestrator is now configured. Use Undo anytime to roll back changes.

How to set up AI Routing (Optional)

AI Routing is what makes your AI Agent smart at the top level. It reads every incoming message and connects the user to the right flow, even when they don't use any keywords.

For example, if a user sends "Hi, I'm looking for a 2BHK in Gurgaon," the AI Router instantly classifies the intent as Property Search and triggers that flow, even though the user never typed the word "property."

This step is optional. If you only have one single flow active, you don’t need agent routing.

Follow the steps below to set up AI Routing.

1. Open the AI Routing tab

On the Flow Builder index page, click the AI Routing tab. You'll see the AI Router toggle at the top with the description: "Route incoming messages to the most relevant flow using AI, even without keyword matches."

SCREENSHOT: AI Routing tab full view, AI Router toggle (off), Orchestrator Prompt section with Auto-Generate from Flows button, warning callout "X flows have no name", and the five sections Role & Purpose, Available Flows, Routing Behaviour, Confidence Guidelines, Important Rules

2. Toggle the AI Router on

Switch the AI Router toggle on to enable intent-based routing across all your flows.

3. Click Auto-Generate from Flows

The fastest way to get started is to click Auto-Generate from Flows on the right of the Orchestrator Prompt section. AiSensy will draft a complete routing prompt based on your active flows, which you can then customize.

The prompt is broken into five editable sections:

• Role & Purpose: defines what the AI router does.

• Available Flows: auto-populated list of your flows with names, IDs, descriptions, sample user phrases, and the information each flow collects.

• Routing Behaviour: defines how the AI structures its routing decisions, including confidence scores and bridge messages.

• Confidence Guidelines: sets the routing thresholds. 0.95+ is an exact match, 0.85–0.94 is a strong match with minor ambiguity, 0.75–0.84 is plausible (route only if no active flow), and below 0.75 means do not route. Answer normally instead.

• Important Rules: guardrails like "never route during a form fill," "one flow per response," and "fall back to plain text if no flow matches."

4. Name your flows clearly

If you see the warning "X flows have no name. Naming your flows helps the AI route messages accurately," go back to Your Flows and rename any untitled ones.

The AI Router uses flow names to route messages, so vague names like "Untitled" or "Flow 1" hurt accuracy.

5. Save and test

Click Save Prompt to lock in your routing logic. Then head to the Test Number tab and send a test message to see your AI Router in action.

It's that simple!!

Testing the AI Agent

Before you go live for real customers, AiSensy lets you test the full agent on a sandbox of up to 3 designated WhatsApp Testing Numbers

Before you go live for real customers, AiSensy lets you test the full agent on a sandbox of up to 3 designated WhatsApp Testing Numbers, each with 500 free AI messages.

Your flow stays in Testing only mode until you purchase an AI Messages addon, which means it'll only respond to those testing numbers.

Here's a quick checklist I'd recommend going through before you turn the bot on.

1 — Ask a question that's directly in your Knowledge Base. The answer should be accurate and specific.

2 — Rephrase the same question. Say it a different way, the way a real customer would. The agent should still get it right — this tests whether semantic search is working.

3 — Ask something completely outside the scope of what you've set up. The agent should deflect gracefully — it should say something like "I'm sorry, I don't have information on that" rather than making something up. If it's hallucinating, tighten your Orchestration prompt in the restrictions section.

3 — If you've set up Tool Calling, trigger it with a real query. Confirm the API call goes through and the response displays correctly.

5 — Deliberately trigger escalation. Send "I want to talk to a human" and confirm the chat routes to your inbox.

6 — Test edge cases. Typos, very short messages, incomplete sentences. Real users don't write perfectly.

Once you're happy with the testing experience, purchase an AI Messages addon from the AI Messages card on the Flow Builder index page (or from the Upgrade Plan dialog). Once the addon is active, remove the Testing only label on your flow to make it live for all users.

Your AI Agent is now ready to handle real customer conversations on WhatsApp. 🎊

Congratulations, you just activated your first AI Agent!

Pricing & access: How AI Messages work

AI Agent Builder runs on two plan tiers. Learn about them.

AI Messages card on the Flow Builder index page, "Purchase AI Messages for Bot Builder

AI Agent Builder runs on two plan tiers, depending on what you need:

At ₹2,500 per month, you get 5 keyword-based chatbots. These are your standard flows that trigger on specific keywords — solid for structured use cases.

At ₹3,500 per month, you get 5 chatbots plus 7,000 AI messages per month. This is what unlocks the AI Agent Builder. So your customers can now have real conversations, not just keyword-triggered responses.

And based on your business volume, you can choose a higher AI message pack — 7,000, 15,000, 30,000, 50,000, or 1,00,000 messages per month. So whether you're a small business just getting started or a high-volume operation handling thousands of queries, there's a pack that fits your requirement.

Two things to keep in mind on usage. First — if you exceed your monthly AI message limit, the agent won't stop. The extra messages will simply be billed separately and reflected in your next invoice. Second — if you have AI messages left over at the end of the month, those don't carry forward. They lapse. So it's worth picking a pack that closely matches your actual usage rather than over-buying.

FAQ

Frequently asked questions answered

Q. What language does the AI agent respond in?

Right now the AI agent responds in English only. However, it can understand queries in any language as long as you've added that context to your Knowledge Base. So if a customer types in Hindi or Spanish, the agent will understand the intent and reply back in English. Full multi-language response support is coming very soon — check the description of this video for the latest updates on that.

Q. Can the AI agent send images or media?

By default, the AI responds in plain text. However, if you've set up Tool Calling and configured a media message type, the agent can reply with images or media for those specific tool responses.

Q. Can I use the AI agent without setting up Tool Calling?

Yes, absolutely. Knowledge Base and Orchestration are enough to get a fully functional AI agent running. Tool Calling is an add-on for when you want the agent to fetch live data from your systems.

Q. Can the AI agent and keyword chatbots run together?

Yes. They coexist. AI Routing decides which one handles each incoming message based on the context of the query.

Q. Do I get free messages to test before going live?

Yes — you get 500 free AI messages for testing. You can add up to 3 numbers and test the bot thoroughly on WhatsApp before turning it on for your customers.

Q. How do I know if my Knowledge Base is working correctly?

After adding your sources, wait for the status to show Completed. Then test the agent by asking questions directly from your KB content. If answers are vague or off, it usually means the content in your KB needs more detail and structure.

Q. Can I test the AI agent on a free plan?

Yes, you can. However, to test it on an actual WhatsApp number, you need the WhatsApp API connected. Without that, you can still test the agent directly inside the Agent Builder itself — you just won't see it on WhatsApp until your API number is set up.

Q. I updated my Knowledge Base — do I need to change anything else or will it just work?

It depends on what you changed.

If you've just added more information, go to the Orchestrator and make sure the updated Knowledge Base is synced there. The agent pulls responses based on what's referenced in the Orchestrator, so if it's not updated there, the agent won't know about the new content.

If you've changed something important like pricing, policies, or how you want the agent to handle certain queries — don't just update the KB. Go into your Orchestrator instructions and update the relevant sections too. Add custom instructions on how the agent should communicate that information and how it should respond.

Either way, always test the agent after making any changes before switching it back on for your customers. Don't assume it'll work correctly — spend 5 minutes testing the key scenarios and confirm the responses look right.

Q. Can the AI agent remember what a customer said in a previous conversation?

No. The AI agent works within a single session. Once the conversation ends, it starts fresh the next time the customer messages. It does not retain memory across sessions, so if a customer messaged you last week, the agent has no context of that when they come back.

Q. My bot gave a wrong answer. What do I do?

First, check if that information exists in your Knowledge Base — if it doesn't, add it. Go into your Orchestrator and add a specific instruction around that topic so the agent knows exactly how to respond. Then test it again before going live.

Q. Can I turn the AI agent off and go back to just keyword chatbots?

Yes, completely. You can toggle the AI agent off anytime and your keyword chatbots will take over. Nothing breaks, nothing gets deleted. You're always in control.

Q. Does the AI agent work 24/7?

Yes, that's the whole point. Once it's live, it responds to customer queries round the clock without any intervention from your team.

Q. My website gets updated regularly. Do I need to manually re-sync the Knowledge Base every time?

Yes, for now you'll need to manually re-sync when there are significant changes. The agent doesn't auto-detect website updates. Make it a habit to re-sync whenever there's a meaningful change — new product, updated pricing, changed policy — anything the agent might be asked about.

Q. What if a customer asks something completely random or off-topic?

The agent will deflect. If it's been set up correctly in the Orchestrator under Restrictions, it'll say something like "I'm sorry, I can only help with X" and not make up an answer. This is why filling in the Restrictions section properly is important — it stops the agent from going off-script.

Q. Can I have multiple AI agents running at the same time?

Yes. You can build multiple agents for different use cases — one for customer support, one for sales, one for FAQs — and AI Routing will direct incoming messages to the right one based on the query.

Q. Can the AI agent handle a high volume of messages simultaneously?

Yes. Unlike a human agent who can handle one conversation at a time, the AI agent handles unlimited conversations simultaneously. No queue, no wait time for the customer.

Q. What if my Tool Calling API goes down or returns an error?

If the API call fails, the agent won't be able to fetch the data and the response will either be incomplete or it'll fall back to a default message. This is why testing your tool thoroughly before going live is important — and it's also a good idea to add a fallback instruction in your Orchestrator for what the agent should say if it can't retrieve information.

Q. Do I need a developer to set this up?

No. The entire setup — Knowledge Base, Orchestration, Tool Calling — is done through a UI with no coding required. For Tool Calling, you will need an API endpoint from your developer or your platform, but the configuration inside AiSensy is no-code.

Q. Will the AI agent work if my customer sends a voice note?

Not as of now. We will update the description of this tutorial once voice note support is added to our WhatsApp AI Agents, so make sure to check the description of this video.

Q. What if a customer is being abusive or spamming the bot?

The agent will keep responding based on the message content. If this is a concern, you can add instructions in the Orchestrator to handle repeated or aggressive messages — like escalating to a human agent after a certain point.

Q. What happens when the AI agent can't resolve a query — does it just stop responding?

No, it doesn't leave the customer hanging. When the AI agent hits a dead end — either because it doesn't have the answer, the customer explicitly asks for a human, or it fails to resolve the query after multiple attempts — it escalates the conversation.

What that means practically is the chat gets transferred to your Live Chat inbox inside AiSensy. From there, one of your human agents will need to pick it up and take over the conversation manually.

So make sure someone on your team has notifications on for the Live Chat inbox — because the AI has done its job at that point, and the ball is in your court. The customer is waiting on the other end.

This is actually one of the best things about the system — the AI handles the volume, and your human agents only step in for the conversations that actually need them.