AI in the Official WhatsApp API: What Meta Allows, What It Prohibits, and How to Configure Without Putting Your Number at Risk

What still applies with or without AI

AI does not change the basic rules of the official API. Everything that applies to human service applies to automated service:

• 24-hour window. AI can respond freely while the customer’s open window is active. Outside it, any message initiated by the company must be an approved template — AI cannot "pull conversation" with free text.
• Consent. The company only initiates contact with those who have authorized receiving messages. Automating sending does not waive opt-in.
• Number quality. Blocks and reports caused by bad AI reduce quality just like those caused by a bad campaign.
• Commerce and content policies. Prohibited products and topics remain prohibited, and AI must be configured not to address them.
• LGPD. Personal data sent to the AI model require clear purpose, legal basis, and care with the provider processing them.

Where AI delivers results in customer service

• Immediate answers to frequent questions — hours, reference price, deadlines, payment methods, policies — based on information registered by the company.
• Understanding the customer’s written intent, including typos and long messages, to route to the correct department or flow.
• Lead qualification in natural conversation, collecting interest, timing, and size before passing to a salesperson.
• Scheduling and rescheduling integrated with the calendar.
• After-hours service, resolving simple issues and organizing complex ones for the team the next morning.
• Conversation summary for the agent, who takes over without needing to read fifty messages.
• Audio transcription, so voice messages enter automation and history.

The Four Limits Method

The difference between AI that sells and AI that burns the number lies in the limits set before it talks to the first customer.

Limit 1 — Topic

AI talks only about the company’s business and nothing else. Out-of-scope questions receive a polite response that brings the conversation back: "Here I can help you with orders, deliveries, and exchanges from the store. Can I help you with any of these?" This protects the company and keeps use within the platform’s purpose.

Limit 2 — Source

AI responds based on material provided by the company — catalog, table, policies, FAQs, funnel scripts. When information is not there, it does not invent: it says it will confirm and calls a person. Price, deadline, and conditions should never be guessed.

Limit 3 — Decision

Some decisions always remain with humans: discounts outside policy, deadline exceptions, serious complaints, cancellations with fees, sensitive health, legal, and financial issues. AI recognizes these cases and transfers.

Limit 4 — Transparency

The customer must be able to talk to a person whenever they want and must not be misled about talking to a human. A simple notice at the start and a clear exit — "just type agent" — prevent frustration and reports.

How to write AI instructions

Configuring AI is, in practice, a customer service document. A working model has six parts:

• Who it is: company name, assistant’s role, and tone of voice — friendly, direct, without emoji overuse.
• What it solves: the closed list of topics.
• Where information comes from: official material and explicit instruction not to answer anything not there.
• What it never does: discounts, promises, diagnoses, opinions about competitors, topics outside the business.
• When to transfer and what phrase to notify the customer.
• Examples of good responses for the most common questions, written by the team itself.

This document should be reviewed whenever price, policy, or product changes. AI with outdated material makes mistakes with the same confidence as when it is correct.

AI in campaigns: what changes

When the company sends an approved template — reminder, offer, reactivation — and the customer responds, the 24-hour window opens and AI can conduct the conversation. This is one of the highest-return uses: no one who responded to a campaign waits. Two precautions: AI must know which campaign started the conversation to respond with context, and must recognize responses like "stop" or "I don’t want to receive more" and immediately register the opt-out without trying to convince.

Escalation to human is part of the product

Good AI knows when to exit. Transfer triggers every configuration should have:

• The customer asks to speak to a person.
• AI did not find the answer in the company’s material.
• Signs of irritation — repeated messages, complaints, strong language.
• Topics on the human decision list.
• High-value opportunity deserving consultative service.
• Two consecutive attempts without AI understanding intent.

And the transfer must be good on the other side: the agent receives the summary, what was asked, and what AI already answered, and the customer is informed of the expected wait time.

Example configuration for an online store

• Initial notice: "Hi! I’m the virtual assistant of Aurora Store and can help you with orders, exchanges, sizes, and deadlines. If you prefer to talk to the team, just type agent."
• Material: catalog with stock, size chart, exchange policy, deadlines by region, payment methods.
• Can do: check order, inform deadline, guide exchange, suggest size by chart, send product link, register exchange request.
• Cannot do: grant discounts, promise deadlines outside the chart, approve exchanges outside policy, discuss topics unrelated to the store.
• Transfers when: product arrived defective, order is delayed beyond deadline, customer asks for agent, or AI did not find the information.
• After hours: resolves what is in the material and organizes transfer cases, informing that the team responds starting at 9h.

The numbers: how much AI frees up the team

With illustrative numbers from an operation receiving 3,000 conversations per month:

• Average human service time per conversation: 7 minutes — 350 hours per month.
• If AI resolves 45% of conversations alone (simple questions, order status, scheduling), that’s 1,350 conversations and about 157 hours freed.
• In transferred conversations, the summary reduces average time from 7 to 5 minutes: another 55 hours.
• Total: about 212 hours per month, equivalent to more than one full-time person.

The most important gain, however, is usually in immediate response outside business hours and during peaks — that’s when the company used to lose sales. And risk must be accounted for: if poorly configured AI causes blocks that reduce the number’s quality, the time savings are not worth it.

How to measure if AI is helping or hindering

• Resolution rate without human, checked by sample — truly resolved, not "customer gave up."
• Transfer rate and the most common reasons, showing what is missing in the material.
• Blocks and reports after conversations with AI.
• Weekly conversation review, focusing on wrong or invented answers.
• Conversion of conversations handled by AI compared to those handled only by humans.
• Satisfaction at the end of service, with a simple question.

Data and AI provider precautions

• Send only what is necessary to the model to respond. Documents, cards, and health data should not circulate without clear necessity.
• Know your provider: where data is processed, if used to train models, and how long it is stored.
• Update your privacy policy to inform about automation and AI use in customer service.
• Keep conversation records to audit responses and comply with data subject requests.

Mistakes that put the number at risk

• AI without topic limits, responding about anything.
• AI inventing price, deadline, or policy.
• Hiding that it is automation or making access to a person difficult.
• Using AI to initiate conversations with free text outside the 24-hour window.
• Automating sending to those who did not consent.
• Not reviewing conversations and discovering problems only through quality drops.
• Launching everything at once, without testing with part of the service.

Where to start

List the twenty questions the team answers most and gather official material that answers each. Define in writing the Four Limits — topic, source, decision, and transparency — and transfer triggers. Turn on AI first after hours or in part of the conversations, review daily in the first week and weekly thereafter. When resolution rate is good and blocks stable, expand. The right AI does not replace the team: it ensures no customer is left unanswered and that team members spend their time where it matters.