How to launch a WhatsApp AI agent: templates first, prompt second

By the end of this guide you'll have an AI agent answering customers on your own WhatsApp Business number. It will reply instantly inside the 24-hour window, fall back to an approved template outside it, honour opt-outs before the AI reads them, and hand photos and edge cases to your team with the thread intact.

The order of work matters more than the prompt. On a call the agent can say anything at any time. On WhatsApp, Meta decides what may be sent and when: free-form replies only within 24 hours of the customer's last message, approved templates for everything else. Teams that write the prompt first and the templates last ship an agent that promises "I'll send that tomorrow" and then can't. Build the template map first.

Before you start

• A Meta business portfolio with a WhatsApp Business Account, and a number that can receive a verification code and isn't in use on the WhatsApp Business app.
• Business verification done or under way. It raises your messaging limit and template allowance.
• A display name that matches your brand (Meta reviews it).
• A working agent with its knowledge base and tools. If you don't have one, start with how to set up an AI voice agent; the same agent will answer on WhatsApp.
• A record of opt-in for anyone you plan to message first, and a named owner for handoffs.

1. Decide who starts each conversation

List every journey the agent will run and mark who sends the first message. That one column decides how much template work you have.

If customers write first (support, order questions, replies to a click-to-WhatsApp ad), the agent mostly answers inside the window and needs a few templates for follow-ups. If the business starts conversations (reminders, renewals, a lead who went quiet), every first message is a template, needs opt-in and counts against Meta's messaging limit.

One rule catches almost everyone: a phone call does not open the WhatsApp window. Only the customer messaging that number does. Callaro measures the window from the person's last message to the connected number, so a brochure your voice agent sends mid-call to someone who has never written to you goes out as a template. Plan one for it.

2. Connect your number

Callaro connects through Meta's Cloud API directly, using Meta's embedded signup, or through a provider you already use: 360dialog, Gupshup, Interakt, WATI, AiSensy or Twilio. With embedded signup you complete Meta's own popup; Callaro then subscribes to the account's webhooks and can register the number for you.

A direct Cloud API connection (or 360dialog) carries Meta's full set of rich messages, including catalogs, WhatsApp Flows and native checkout; other providers carry a reduced set. If you want a storefront in the chat, confirm this before you move a number.

A number can be live in only one Callaro workspace at a time, because Meta delivers each event once and two workspaces would both answer every message. Keep a separate test number.

3. Build the template map before the prompt

List every message the agent might need to send while the window is closed. A typical set:

• a follow-up that reopens the conversation
• an appointment or visit confirmation
• a link the customer asked for (document, form, payment page)
• a callback notice
• a verification code, if you identify visitors on other channels

Callaro has starter templates for the first four. You edit the wording in any language, the editor checks structure against Meta's template rules before you submit, and you submit for review from the same screen. Meta's automatic review can take up to 24 hours. Use named variables rather than numbered ones.

Then map each approved template to an intent, the action the agent can take, and bind its variables to values the platform already knows. The intent also carries the purpose (service, transactional or marketing), so the model never decides that a promotional message counts as service.

Keep utility templates free of offers. Meta moves a utility template to marketing when it mixes an update with a promotion, and from then on it falls under marketing opt-outs and limits. Callaro's scheduled sync picks up category, status and quality changes. If Meta pauses a template, or its variables stop matching the binding, the intent is flagged in the dashboard and removed from the agent's options. The agent stops offering what it can't send, rather than failing at send time.

4. Give the agent a WhatsApp style, not a second agent

In Callaro one agent keeps its instructions, knowledge, tools and guardrails, and adapts its style per channel. Each customer has one memory across channels, so a WhatsApp reply the morning after a call starts from what was said on the call.

What to add for WhatsApp:

• Offer choices as buttons, not open questions. Meta allows up to three reply buttons and ten rows in a list; when the agent offers more than three options, Callaro sends a list.
• Show product cards only for items in your data. Links, media and product IDs must come from what the business gave the agent (instructions, variables, tool results), never from the customer's own messages, and an invented one is refused. A customer can't talk the agent into sending a made-up link.
• Keep each reply to one short message. WhatsApp shows bold, italics and lists, but not headings.

5. Design for bursts, because every reply now costs

People write in bursts: "hi", then "I booked last week", then "need to change the date". An agent that answers each line sends three replies to one question.

That matters more since 1 October 2026, when Meta started charging per message for service messages inside the customer service window, as it already did for templates. Every reply is now a line on your Meta bill. Callaro bills AI usage per conversation on tokens rather than per message, but the Meta side grows with how chatty the agent is.

Callaro waits briefly after the last inbound message, so a burst gets one considered reply, and sends one message per turn. Before launch, also set a daily spend cap with an alert and check the limit on AI replies per conversation, after which a long thread goes to a person.

6. Make consent and opt-outs work before the AI does

Meta requires opt-in before you message someone first. It must name your business and make clear what the person is agreeing to receive. Store when and how it was given.

In Callaro, opt-out and help keywords are handled before the AI sees the message. STOP records an opt-out from everything and lets exactly one confirmation reply through. HELP gets your fixed help text, never an improvised answer. Opt-outs can also be recorded for one purpose, so someone who stops offers still receives appointment confirmations.

Business-initiated sends wait out quiet hours in the recipient's own time zone, using the stricter of your market's rule and your agent's sending window. Replies to a customer who just wrote are never delayed.

Keep the agent about your business. Since 15 January 2026, Meta's business terms have barred general-purpose AI assistants as the main product on WhatsApp (a rule now contested in the EU), while AI that supports a company's own customers is allowed.

7. Decide what goes to a person

Write the handoff rules before launch, not after the first complaint.

• Photos, videos and files with no words go to a person by default, with a short reply to the customer and a note on the thread. A photo of a damaged parcel needs human judgement. A caption counts as text, and the agent answers it.
• Voice notes are transcribed and answered like text. The agent can also reply with a voice note in its own call voice where available, with text as the fallback.
• Requests for a person, disputes and anything legal go to your team with a summary. A teammate can take over from the inbox and hand back with "Let AI answer"; while they type, the AI holds its reply. Handoff alerts go out by email and webhook.

8. Connect WhatsApp to the rest of the journey

WhatsApp earns its keep when it isn't on its own:

• From a call to WhatsApp. Turn on the in-call send tool for your voice agent and map it to the same intents. When a caller asks for the brochure, the agent sends the approved template and tells the caller whether it went, or why not.
• From WhatsApp to a call. When a customer writes "can someone just call me", the agent can place a call that starts with the chat's context.
• Follow-ups. A follow-up promised in a chat can go out at the time the customer chose: free text if the window is open, your follow-up template if not. The follow-ups guide covers cadence design.

For your CRM, subscribe your webhook to the WhatsApp conversation events you need, such as started, handoff and completed; subscription is opt-in per event. The CRM connection guide covers deduplication and record IDs.

9. Test from a real phone, then ramp up

Builder test chats are kept apart from production: they don't fire your webhooks or follow-up rules. So run a final test from a real handset on the real number and check the outcome: captured data, webhook delivery, any follow-up.

Test the closed-window path on purpose by sending a follow-up to a test number that hasn't written for more than a day. The provider's view of the window can differ from Callaro's, for example when the person wrote from another device. If the provider refuses a free-form message, Callaro sends the intent's approved template instead, or your re-engagement template once per conversation, and leaves a note for the team.

Then ramp up. Meta's messaging limit for messages outside the window starts low for a new portfolio and is shared by every number in it. It rises with business verification and a record of high-quality sends, so watch your quality rating and each template's rating.

Gotchas from real deployments

• The missing template. The most common launch problem we see is a template that was never mapped, not the AI. If the agent can promise it, map it.
• Counting the window from your own reply. The 24 hours run from the customer's latest message. Your agent's replies don't extend it.
• The phone number as the only key. WhatsApp is rolling out usernames during 2026, and webhooks can arrive with a business-scoped user ID and no phone number. Callaro keeps one identity per person when that ID appears, but your CRM join shouldn't rely on the phone number alone.

Common questions

Can the agent message someone first?

Yes, with an approved template, the person's opt-in, and outside quiet hours.

What does it cost to run?

Two parts: Meta's per-message charges, paid to Meta or through your provider, and AI usage billed per conversation. Fewer, fuller replies keep both down. See how costs are structured.

Launch checklist

• [ ] Number connected, not on the WhatsApp Business app, live in one workspace only
• [ ] Business verification submitted and display name approved
• [ ] Journeys listed, with who sends the first message
• [ ] Templates approved and mapped for every message the agent can promise
• [ ] Utility templates free of offers; named variables bound to known values
• [ ] WhatsApp style written: buttons over open questions, one short message per reply
• [ ] Daily spend cap and alert set
• [ ] Opt-in stored; STOP and HELP tested; quiet hours checked per market
• [ ] Handoff owner named; photo handoff tested
• [ ] In-call send, call-me-now and follow-ups mapped to the same intents
• [ ] Webhook subscribed to the WhatsApp events your CRM needs
• [ ] Real-handset test done, including the closed-window path
• [ ] Quality rating and flagged intents reviewed weekly

Once the template map is solid, the rest is the same work as any channel: clear instructions, grounded tools, clean handoffs. For the bigger picture, read the complete guide, or go to the WhatsApp agent page to connect a number.