AI agents
An AI agent answers customers’ WhatsApp messages for you, using only the knowledge you give it: FAQs you paste in and web pages you point it at. When nothing in that knowledge is close enough to the question, it doesn’t guess: by default it hands the chat to a person.
AI agents live under Automate → AI Agents in the sidebar.
When an agent answers
Section titled “When an agent answers”Flows always come first. An agent answers a message only when:
- nobody is assigned to the chat. Agents never talk over a person;
- no flow is running for the customer; and
- no flow matched the message. Flows take keyword, first-message and voice-note messages first.
See who answers a message for the full order. Once an agent has started talking with a customer, it keeps answering that chat until it hands it to a person.
Create an agent
Section titled “Create an agent”-
Open AI Agents, press New agent, type a name and press Create. A new agent starts Disabled.
-
On the Instructions tab, describe the assistant under Who is this assistant?: the business, what it helps with, anything it should know about how to talk to customers. Set the Tone, for example “friendly and professional”.
-
On the Knowledge tab, add what it should answer from (see below).
-
On the Settings tab, choose When should this agent respond?
-
Try it on the Test tab.
-
Press Disabled in the top bar to switch it to Enabled.
Every change saves itself as you type (“Autosaves” / “Saved” in the top bar). You can also enable, disable or delete an agent from the AI Agents list. Deleting an agent deletes its knowledge base too.
Knowledge
Section titled “Knowledge”The agent answers only from its knowledge. With no sources it has nothing to answer from, and every question goes to a person (or, with handoff switched off, gets no reply).
On the Knowledge tab, choose:
- Paste FAQ / text: give it a title and paste your questions and answers, price lists, policies, anything written down.
- Website URL: give it a title and a web page address.
Then press Add to knowledge base.
How indexing works
Section titled “How indexing works”A new source shows pending while whatbot reads it and indexes it for search. The list updates on its own when it is done:
| Status | Meaning |
|---|---|
| pending | Being read and indexed |
| embedded | Ready. The agent can answer from it |
| failed | Couldn’t be read. The reason is shown next to it |
Click a pasted-text source to see what it contains. A website source’s fetched text isn’t kept for preview.
Some limits worth knowing:
- A Website URL source reads that one page, not the whole site. Add each page you want covered as its own source. Menus, headers and footers are skipped.
- A website is read once, when you add it. If the page changes, delete the source and add it again.
- Up to about 200,000 characters are read from each source.
Project knowledge from the CRM
Section titled “Project knowledge from the CRM”If your workspace is connected to the Zetta CRM (Settings → CRM Connection), the Knowledge tab can also pull the CRM’s published project knowledge. Press Choose projects, pick up to 2 projects and press Attach selected. They appear as sources marked Project data. The refresh button next to Choose projects pulls the attached projects’ latest content again.
When the agent responds
Section titled “When the agent responds”Pick one on the Settings tab:
| Option | The agent answers |
|---|---|
| All unhandled messages (fallback brain) | Any message that no flow and no other agent handles. |
| When a keyword matches | Messages containing one of its keywords. |
| Customer’s first message | The first message it gets in a chat no AI agent has talked in before, when no flow took it. |
| Manual only | Nothing on its own. |
Keywords follow the same rules as flow keywords: case doesn’t matter, a keyword matches at the start of a word, the longest match wins, and each keyword belongs to only one flow or agent. Keywords are saved when you click away from the box.
One catch-all agent per workspace
Section titled “One catch-all agent per workspace”Only one agent in the workspace can be the fallback (All unhandled messages), and only one flow or agent can own Customer’s first message. A disabled agent still holds its slot. If another agent or flow already owns the slot, the Settings tab warns you and the change is refused. Change the other one first, or use the Message routing card at the top of the Flows page, which shows who owns both slots and lets you swap them.
Conversation memory
Section titled “Conversation memory”Conversation memory is how many earlier messages of the chat the agent re-reads before each reply, from 2 to 40. More keeps it on track in long chats, but each reply takes a little longer. Fewer is quicker, but it forgets sooner.
Handing off to a person
Section titled “Handing off to a person”On the Handoff tab:
- Handoff phrases: comma-separated words or phrases, such as
talk to a human, call me, agent. If the customer’s message contains any of them, the agent hands off at once. - Hand off to a human when the answer isn’t in the knowledge base: when
nothing in the knowledge is a close enough match for the question:
- On (the default): the customer is told a teammate is coming, and the chat is handed off.
- Off: the agent stays silent, and the customer gets no reply.
On a handoff, the customer gets the message “Let me connect you with one of our team members who can help you better. 🙂”. The chat is assigned to the least-busy available teammate, or stays unassigned if nobody is available, and an internal note says the AI handed it over. The agent then stays silent in that chat. It doesn’t pick the chat up again on its own.
Test an agent
Section titled “Test an agent”The Test tab is a chat window that runs the real agent against your
knowledge. Nothing is sent on WhatsApp, and nothing is charged. Under each
reply you see what the agent did (reply, handoff, or no_answer when
handoff is switched off) and how closely the best piece of knowledge matched,
for example match 72%. A handoff in the test means the message contained a
handoff phrase or the knowledge base had no good enough answer: add or
improve a source and ask again.
The Conversations tab lists the most recent 50 chats the agent has handled, each with the full back-and-forth between the customer and the agent.
What the agent’s replies look like
Section titled “What the agent’s replies look like”- It is told to write in the customer’s language when that is obvious.
- It uses WhatsApp formatting: bold for names, prices and key numbers, and • bullets for lists.
- Replies are kept short, under about 900 characters.
- It is told not to mention documents, sources or tools, and just answer.
Agents and flows
Section titled “Agents and flows”An agent and a flow are separate. A flow can’t call an agent as one of its steps. They share the work by order: flows handle what they are triggered for, and an agent picks up the rest. To end a flow and let the agent take over the following messages, end the flow without assigning the chat to a person.
Billing
Section titled “Billing”Each message an AI agent sends costs ₹0.20, taken from your prepaid wallet. It appears on the Billing page as AI agent message. The handoff message is a message the agent sends, so it is charged too.
Not charged:
- anything on the Test tab;
- a question the agent stays silent on;
- a reply the agent wrote but didn’t send because the customer’s 24-hour window had closed;
- AI reply suggestions and summaries in the inbox.
See Billing & wallet for how the wallet works.