WhatsApp is often the easiest channel for reaching users outside the US - no carrier spam filtering to fight, no 10DLC registration, and reach into markets where SMS delivery is unreliable. It also has its own rulebook, and that rulebook does not look anything like SMS.
Get it wrong and the failure mode isn't a filtered message - it's a message that silently fails to send, or a number that gets restricted entirely. Here is what actually governs WhatsApp messaging for an AI agent.
The 24-hour window is the whole game
This is the single most important rule and the one most teams miss first. Once a user messages your business, you have a 24-hour window in which you can send them anything - free-form text, no restrictions, no pre-approval. Once that window closes, you can only reach them with a pre-approved message template, and only for approved categories.
For an agent that's actively conversing with a user, this is invisible - the conversation stays inside the window naturally. It becomes visible the moment your agent needs to follow up later: confirming an appointment three days out, sending a reminder after the conversation went quiet, re-engaging someone who hasn't replied in a week. All of that falls outside the 24-hour window and requires a template.
What this means practically: if your agent has any workflow that reaches out first, or follows up after a gap, you need templates ready for those cases before you need them. Building the flow first and discovering the window limit in production is a common and avoidable mistake.
Templates: what they are and how approval works
A template is a pre-written message format, submitted for approval before you can use it - things like "Your appointment with {{1}} is confirmed for {{2}}." Variables get filled in at send time, but the surrounding structure and category can't change without going through approval again.
Templates fall into a few categories - utility (appointment reminders, order updates), marketing (promotions), and authentication (OTP codes) - and the category affects both approval scrutiny and cost. Utility and authentication templates are reviewed for being genuinely transactional, not promotional in disguise; marketing templates face the most scrutiny and, in most markets, require explicit opt-in that's more rigorous than for the other categories.
Approval typically takes anywhere from minutes to a couple of days. Templates get rejected for looking too generic, containing content that reads as spam, or having variables that could be filled with anything (a genuinely open-ended {{1}} invites rejection - the more specific the template's fixed language, the smoother approval goes).
Practical implication for an agent: don't design a conversation flow that assumes free-form text can go out at any time. Map out which messages happen inside an active conversation (fine, no template needed) versus which ones are agent-initiated after a gap (need a template, submitted and approved ahead of time).
Opt-in is stricter than SMS, and it's enforced
WhatsApp requires an explicit opt-in before you send a user anything outside of them messaging you first - and "explicit" is interpreted narrowly. A user giving you their number on a web form is not opt-in. A user checking a box that specifically names WhatsApp as the channel they're agreeing to receive messages on is closer to what's expected.
This matters more on WhatsApp than SMS because Meta enforces it partly through the quality rating system described below, not just through legal liability after the fact - a pattern of sending to users who never opted in shows up as a quality problem before it shows up as anything else.
The quality rating that can restrict your number
Every WhatsApp Business number has a quality rating - visible to you, invisible to the recipient - that Meta tracks based on block rates, negative feedback, and how recipients interact with your messages. It moves through tiers, and a number that drops to the lowest tier can have its messaging limits reduced or, in sustained cases, get restricted from sending altogether.
The rating responds to the same things that determine any legitimate messaging pattern: sending only to people who opted in, respecting the 24-hour window rather than trying to force free-form messages through templates in bad faith, keeping message content relevant to what the recipient actually expects from you. A quality drop is a signal that something about the sending pattern needs to change, not just a number to route around.
For an agent, the practical read: the same discipline that keeps a phone number's reputation healthy generally - real opt-in, relevant content, respecting the channel's structure - is what keeps WhatsApp's quality rating healthy specifically. It is not a separate problem from general sender reputation; it's the same problem with WhatsApp's own scoring layer on top.
The short version
Inside an active 24-hour conversation, WhatsApp behaves like free-form messaging - send what you want. Outside it, every message needs a pre-approved template, categorized correctly, submitted ahead of the moment you need it. Every recipient needs genuine opt-in to the channel specifically. And the quality rating is a live signal, not a one-time gate - it responds to sending patterns the same way carrier reputation does for voice and SMS, just with WhatsApp's own enforcement behind it.
Design the agent's follow-up and re-engagement flows around the 24-hour window from the start, and WhatsApp becomes one of the more reliable channels available, not a source of silent failures.