How AI-drafted replies work in the WhatsApp inbox — and why they never send themselves

A message lands in the shared team inbox: a customer asking where their order is, three days past the promised date. Before anyone types a word, there's a blank reply field and a decision to make about tone, specifics, and what to promise next. That's the exact moment the AI-drafted-reply feature is built for — not to answer the customer, but to give the human working the conversation something better than a blank field to start from.

What actually generates the draft

The feature sits inside the shared team inbox, attached to a conversation that already has an inbound customer message in it. An LLM suggests a reply for that conversation — a starting draft for a human on the team to review, edit, or send. The suggestion appears inline in the inbox, next to the conversation it belongs to, visible to whoever is working it.

Generating a draft and sending a message are two separate actions, and only one of them reaches the customer. A draft is just text sitting in the inbox until someone acts on it.

What the human does with it

Once the draft appears, there are exactly three things a team member can do with it:

• Send it as written, if it's accurate and reads right
• Edit it — fix a detail, adjust the tone, add something the LLM didn't know — then send the edited version
• Ignore it entirely and write a reply from scratch

All three are normal outcomes. The only message that leaves the inbox and reaches the customer over WhatsApp is the one a person explicitly hits send on — whether that's the draft untouched, the draft after edits, or something typed independently. It never sends itself: there's no timeout, no auto-send-if-unedited-after-a-few-minutes, no background dispatch. A draft can sit there unsent indefinitely; it's just text in the inbox until a human acts on it.

Why it stops there, deliberately

The reason it stops at a draft rather than sending outright comes down to what an LLM drafting a reply from a conversation can't independently verify: the exact status of a specific order, a refund amount, a delivery date, a personal circumstance the customer mentioned earlier in the thread, a promise about timing that a human on the team may or may not be able to keep. Get one of those wrong in a message a customer receives as coming directly from the business, over a channel where they keep the full chat history and can screenshot it, and the cost isn't a UI bug report — it's a real customer relationship taking the hit for something no person actually decided to say.

So the design leaves the narrower, more reliable job to the AI — producing a plausible starting point, structure, and phrasing — while keeping a person accountable for every word that actually goes out. A draft is a starting point, not a decision. The decision, and the responsibility for it, stays with whoever hits send. That division holds even where messages do go out through automated paths elsewhere in the product: the MCP server's WhatsApp tools, for instance, can send a text or a template message directly, but that's a human-authored integration calling a specific, pre-approved action — not an LLM freely composing and dispatching a reply on its own judgment. The draft feature in the shared inbox is the one place an LLM is involved in composing a live reply, and it's also the one place that composition is explicitly walled off from sending.

How it's billed, and when it's actually worth turning on

AI-drafted replies are billed separately from every plan tier, as a usage-based add-on called WhatsApp AI Credits — not a flat monthly charge, and not bundled free into any plan, including the higher tiers that unlock other WhatsApp features like automation rules. It stacks alongside whatever base plan an account is already on, from the Pay-As-You-Go API-only tier up.

That billing shape is worth being direct about: it's usage-based, so it's not something every account benefits from turning on by default. For a team fielding a handful of WhatsApp conversations a day, writing each reply from scratch costs a few minutes, and the AI credits add a line item without much to show for it. The case for it gets stronger as conversation volume goes up — a team handling enough inbound messages that a meaningful share of them are variations on the same kind of question (order status, appointment confirmation, a policy question) is where a drafted starting point has the most room to help, versus composing every response from a blank field.