Product
WhatsApp AI agents: what's allowed in 2026 and how to use AI safely
AI can answer, sort and draft WhatsApp messages, but Meta now restricts some AI uses and still requires a route to a human. Here is what the rules say, where AI helps most, and how Buddy AI keeps a person in charge of every send.
The Buddy Team · · 11 min read
Product
AI on WhatsApp
A WhatsApp AI agent is software that reads customer messages on WhatsApp and uses AI to answer them, sort them or draft replies for a person to send. Businesses may use AI to serve their own customers. Since 15 January 2026, Meta's terms bar providers from offering general-purpose AI assistants as the main thing people use through the WhatsApp Business Platform.
This guide explains the rule precisely, the other policies every AI setup on WhatsApp must follow, the difference between an autopilot bot and an AI co-pilot, and exactly what Buddy AI does today (it is early) versus what is planned.
10M+
conversations a week running through business AIs by April 2026, up from 1 million at the start of the year
Meta, Q1 2026 earnings call
What is a WhatsApp AI agent?
"AI agent" covers a wide range of software. It helps to separate three levels, because the rules, risks and costs are different for each.
| Level | What it does | Who sends the message | Typical risk |
|---|---|---|---|
| Rules-based automation | Replies to keywords or events with fixed messages | The system, from text you wrote | Low; it only says what you told it to |
| AI-assisted (co-pilot) | Sorts messages by intent and drafts replies | A person, after reviewing the draft | Low to medium; a human checks before anything goes out |
| Autonomous AI agent | Reads and answers customers on its own, sometimes taking actions such as bookings | The AI | Higher; wrong answers, invented facts and missed escalations reach customers directly |
Most businesses benefit from the first two long before they need the third. Rules handle the predictable (welcome messages, keyword answers), AI handles the messy middle (understanding what a reply means, getting a first draft down), and people handle anything that needs judgement. Our guide to WhatsApp auto replies covers the rules-based layer in detail.
Are AI chatbots allowed on WhatsApp in 2026?
Yes, for businesses using AI to serve their customers. What changed is aimed at AI companies distributing their assistants through WhatsApp. Meta's terms for the WhatsApp Business Platform now say:
Providers and developers of artificial intelligence or machine learning technologies, including but not limited to large language models, generative artificial intelligence platforms, general-purpose artificial intelligence assistants, or similar technologies as determined by Meta in its sole discretion ("AI Providers"), are strictly prohibited from accessing or using the WhatsApp Business Platform, whether directly or indirectly, for the purposes of providing, delivering, offering, selling, or otherwise making available such technologies when such technologies are the primary (rather than incidental or ancillary) functionality being made available for use.
TechCrunch reported the change in October 2025, with the restriction taking effect on 15 January 2026. It named general-purpose assistants such as ChatGPT and Perplexity as affected, and noted that a business using AI for customer service, such as a travel company's support bot, remains permitted. The deciding test is whether AI is the primary product or incidental to a business serving its own customers.
| Scenario | Allowed? | Why |
|---|---|---|
| A fashion store uses AI to answer sizing and delivery questions | Yes | AI is ancillary to the store serving its customers |
| A clinic's bot handles appointment questions and hands over to staff | Yes | Structured business use with a human route |
| A team uses AI to draft replies that staff review and send | Yes | AI assists people; the business is the sender |
| A company offers a general "ask me anything" AI assistant over WhatsApp | No | The AI assistant is the primary functionality |
This is our reading of the published terms, not legal advice. If AI is central to what you sell, read the terms in full and take advice.
What rules apply to any AI on WhatsApp?
The AI Providers restriction is only one part. Any automated or AI-driven setup also has to meet the everyday rules of the platform.
- A clear route to a human. The WhatsApp Business Messaging Policy says: "You may use automation when responding during the 24-hour window, but must also have available prompt, clear, and direct escalation paths." It lists options such as transfer to a human agent in the chat, phone, email, web support, in-store visits or support forms.
- The 24-hour window. Free-form replies, including AI-written ones, are only possible within 24 hours of the customer's last message. After that, the business must use an approved template (Meta). An AI agent cannot start conversations with free text.
- Opt-in. AI does not change consent. You can only message people who opted in to hear from you on WhatsApp.
- Every reply has a price. From 1 October 2026, service replies inside the 24-hour window are charged at the utility rate (Meta). An agent that replies to every "ok, thanks" now costs money. Our October 2026 pricing guide has the detail.
- Quality still counts. If customers block or report your number because a bot frustrated them, your quality rating drops, and with it your ability to grow your messaging limit.
Should AI answer customers on its own, or help your team answer?
Fully autonomous agents make sense for some high-volume, well-defined support flows, especially when they are connected to accurate order and product data. For most businesses, and for anything that touches prices, refunds, health, money or promises, a co-pilot is safer and usually good enough.
| Autopilot agent | Co-pilot (AI assists, human sends) | |
|---|---|---|
| Speed | Instant, around the clock | Fast, but depends on someone being available |
| Accuracy risk | Invented details reach customers | A person catches mistakes before sending |
| Escalation | Must be designed and tested carefully | Built in: a person is already in the loop |
| Cost from 1 Oct 2026 | Every reply it sends is billed | You choose which replies are worth sending |
| Tone and brand | Varies with the prompt and model | Your team edits to your voice |
| Best for | Narrow, repetitive questions with reliable data | Most small and mid-sized teams, sales conversations, sensitive topics |
Buddy is built around the second column. Our principle is co-pilot, not autopilot: AI can understand, sort and draft, but nothing high-impact goes to your customers without a person approving it.
What does Buddy AI do today?
Buddy AI is early. Two capabilities are built and available on workspaces where Buddy AI is enabled. When it is not enabled, Buddy works exactly the same, with keyword rules in place of AI.
1. Suggest a reply with AI, in the Inbox
When you open a conversation in the Buddy Inbox, a "Suggest a reply with AI" button drafts a response to the customer's latest message. It reads the recent conversation (up to the last 12 messages), writes one warm, concise reply of one to three sentences, and matches the customer's language, so a question in Portuguese or French gets a draft in the same language.
It is instructed not to invent prices, policies or facts it does not know, and to offer to check instead. The draft goes into your reply box. You read it, edit it and decide whether to send it. Nothing is sent automatically. The usual WhatsApp rules still apply: if more than 24 hours have passed since the customer last wrote, Buddy explains that a free-text reply is not possible and points you to an approved template.
2. AI intent classification for automations
Buddy Automations have a trigger called "When a reply is categorized". It fires when an incoming reply is recognised as one of five intents: Interested, Question, Purchased, Follow up or Not now. With Buddy AI enabled, the category is decided by AI, which copes with typos, slang and any language. Without it, or if the AI call fails, Buddy falls back to keyword rules.
That means an automation can react to what someone meant, not just the words they used. A reply of "yh send me the link pls" or "quero comprar" can be treated as Interested even though neither contains the word "interested".
Automation actions are the same as for any trigger: send a message (text or an approved template), add or remove a tag, wait, or notify the team in the app. On the visual canvas you can also branch on category, message text or tags.
How to set up AI-assisted triage in Buddy
Here is a practical setup for a small team, using only what is built today. Imagine a skincare brand in Manchester that sends a monthly WhatsNew to about 3,000 subscribers and gets a few hundred replies.
- 1Interested replies: automation on "When a reply is categorized" as Interested. Add the tag
hot-leadand notify the team, so someone replies while the customer is still in the chat. - 2Questions: on Question, notify the team. Your team opens the Inbox, uses "Suggest a reply with AI" for a first draft, fixes any facts and sends.
- 3Purchased: on Purchased, add the tag
customerand send a short thank-you. The customer has just messaged you, so a free-text reply inside the 24-hour window works, though it is billed at the utility rate from 1 October 2026. - 4Not now: on Not now, add the tag
not-nowand do nothing else. Use the tag to leave them out of the next WhatsNew or two. - 5Opt-outs are separate: replies of STOP, unsubscribe or opt out unsubscribe the contact automatically, whatever the AI thinks.
This meets the escalation requirement naturally, because every conversation that needs judgement lands with a person. It also keeps costs down: automations only send where a reply adds value. For more on running replies as a team, see our WhatsApp team inbox guide.
What is planned for Buddy AI?
These are on our roadmap for after Buddy 1.0 launches on 12 October 2026. They are planned, not built, and each will keep a person in charge of the final action.
Drafting WhatsNews
Turn launch notes or a rough idea into a campaign draft that you review, edit and approve before it is scheduled.
Segment suggestions
Suggest audiences from tags, sources and engagement history, for example people who clicked last month's launch but did not reply.
Engagement summaries
Summarise what happened after a send: common questions, themes in replies and who needs follow-up.
Further out, Buddy MCP is planned as a controlled way for AI clients to work with a Buddy workspace, with scopes, confirmations and audit logs. Meta has also announced a WhatsApp Business MCP (15 September 2026) aimed at developer tasks such as setup, templates and webhook testing. Our WhatsApp MCP guide explains the difference.
Which WhatsApp AI approach is right for you?
| Option | Good fit when | Things to check |
|---|---|---|
| Meta Business Agent | You want Meta's own agent answering on WhatsApp and Messenger | What it knows about your business, how handover works, and future subscription costs Meta has said are coming |
| Build your own agent on the Cloud API | You have developers and accurate internal data | Escalation paths, the 24-hour window, per-reply costs, testing |
| An engagement platform with AI assistance (such as Buddy) | You want audience, campaigns, inbox and automations in one place, with AI helping your team | Which AI features are live versus planned, and who approves sends |
Whichever you choose, the fundamentals are the same: consented subscribers, approved templates for outbound messages, fast human escalation, and messages people are glad to get. Our pillar guide to picking a WhatsApp marketing platform covers the rest of the stack.
A checklist for using AI on WhatsApp responsibly
- Tell people how to reach a human, and make sure someone actually responds.
- Keep AI away from prices, refunds and promises unless it is reading accurate, current data.
- Review AI drafts before sending, especially in the first weeks.
- Do not let AI start conversations. Outbound messages outside the window need approved templates anyway.
- Treat opt-outs as absolute, independent of any AI classification.
- Watch reply costs from 1 October 2026 and cut replies that add nothing.
- Check your quality rating regularly, and pause any automation that drives blocks or reports.
Frequently asked questions
Did WhatsApp ban AI chatbots in 2026?
Not for businesses serving their own customers. From 15 January 2026, Meta's terms prohibit AI providers from using the WhatsApp Business Platform to offer AI technologies, such as general-purpose assistants, when the AI is the primary functionality. A business using AI for customer service is still allowed.
Do I need a human agent if I use an AI chatbot on WhatsApp?
You need a prompt, clear and direct escalation path. Meta's Business Messaging Policy allows automation in the 24-hour window but requires escalation options such as handover to a human in the chat, phone, email or a support form.
Can AI send WhatsApp messages automatically in Buddy?
Buddy AI drafts and classifies; it does not send on its own. Suggested replies go into your reply box for a person to edit and send. Automations you create can send messages, but those are messages you wrote or templates you chose.
Is Buddy AI available on every plan?
Buddy AI is early and available on workspaces where it has been enabled. Without it, Buddy uses keyword rules for reply categories and the Inbox works as normal. See the pricing page or contact us for details.
Does AI-classified intent work in other languages?
Yes. When Buddy AI is enabled, reply classification handles typos, slang and any language, and suggested replies match the customer's language.