24/7 customer service on WhatsApp with AI: what to automate and what not to

Updated on October 29, 2026

24/7 customer service on WhatsApp with AI works well for low-risk requests with an objective answer: order status, a copy of an invoice or payment slip, frequently asked questions and scheduling. Complaints, cancellations, disputed charges and anything health-related should be logged by the AI and wait for a human, with the expected response time stated to the customer. WhatsApp’s policy allows automation inside the 24-hour window, as long as there is a fast, clear path to reach a person.

This guide gives you a decision matrix, the after-hours flow, how to handle holidays and SLAs, and what it costs to keep service open overnight on the official WhatsApp Business API.

What WhatsApp’s policy requires from automated service

The WhatsApp Business Messaging Policy says businesses may use automation to respond within the 24-hour window, but must also have “prompt, clear, and direct” escalation paths. The examples it lists include in-chat transfer to a human agent, phone, email, web support and a support form.

In practice, this means two things for 24/7 service:

• The bot can’t be a dead end. Every conversation needs a “talk to an agent” option, even at 3 a.m. After hours, that option logs the request and tells the customer when someone will reply.
• The assistant must be about your business. Since January 15, 2026, Meta restricts general-purpose chatbots on the platform. An assistant that answers about your products, orders and services is still allowed. The details are in the article on Meta’s 2026 AI policy.

Matrix: what AI handles after hours and what waits for a human

The rule behind the matrix is simple: AI resolves the request when the answer comes from data in your system or from approved text, and a mistake is cheap. When money is in dispute, emotions run high or health is at stake, the AI acknowledges, logs and hands off.

Topic	After hours	Why
Order status and tracking	AI resolves	Answer comes from the ERP or the carrier
Copy of a boleto (Brazilian payment slip) or PIX code	AI resolves	Data from the billing system, no negotiation
FAQs (hours, address, returns, deadlines)	AI resolves	Approved text in the knowledge base
Booking, rescheduling and confirmation	AI resolves	The calendar has fixed slot rules
Simple quote from a price list	AI resolves	Price comes from the list, no discount
Complaint	AI logs, human replies	The customer wants to be heard by someone
Contract or subscription cancellation	AI logs, human replies	Involves contract rules and retention
Disputed charge, refund, negotiation	AI logs, human replies	Money in dispute
Symptoms, medication, test results	AI points to emergency services, human replies	Health risk
Sensitive data (documents, medical reports)	Human, through an appropriate channel	Requires extra care under data protection law
Upset customer or one who asked for an “agent”	Human, without pushing the bot	Required by policy and common sense

Two caveats about health. First, the AI must not suggest diagnoses or doses: the after-hours reply should point to emergency services (in Brazil, SAMU 192, the public ambulance service) when the case seems serious. Second, clinics and labs handle sensitive personal data under the LGPD, Brazil’s data protection law (Law 13,709/2018, art. 11). This article is not a substitute for advice from your lawyer or data protection officer. The article WhatsApp for clinics covers allowed use in healthcare.

Want to try it while you read? Create a Vectalk account and send your first message through the official API. Create free account →

How the after-hours flow works

The API has no built-in “away mode.” Your system decides, based on the incoming message webhook. The article on welcome, away and business-hours messages shows the code for the business-hours rule. Here the focus is on what happens after a message arrives outside business hours:

1. Classify the intent. The AI identifies the topic (order, payment slip, booking, complaint, etc.).
2. Check the matrix. If the topic is in the “AI resolves” column, it answers with data from your system.
3. Escalate when needed. If the topic is in the “human” column, the AI confirms what it understood, creates a prioritized ticket and says when the customer will hear back.
4. Offer a way out at any time. A “Talk to an agent” button is available in every reply.
5. Hand the context to the human. In the morning, the agent opens the ticket with a summary, without asking the customer to repeat everything.

Handoff and the summary for the agent are covered in from bot to human on WhatsApp.

An example configuration your backend can read to decide:

{
  "timezone": "America/Sao_Paulo",
  "business_hours": {
    "mon-fri": ["08:00", "18:00"],
    "sat": ["09:00", "13:00"]
  },
  "holidays": ["2026-11-02", "2026-11-15", "2026-11-20", "2026-12-25"],
  "ai_resolves": ["order_status", "invoice_copy", "faq", "booking"],
  "always_human": ["complaint", "cancellation", "charge_dispute", "health"],
  "after_hours_sla": {
    "complaint": "next_business_day_10am",
    "cancellation": "next_business_day_12pm",
    "health": "next_business_day_9am"
  }
}

And the routing logic in Node.js:

function decide(intent, now, cfg) {
  const open = withinBusinessHours(now, cfg); // accounts for timezone and holidays
  if (cfg.always_human.includes(intent)) {
    return open
      ? { action: 'transfer' }
      : { action: 'log', due: cfg.after_hours_sla[intent] };
  }
  if (cfg.ai_resolves.includes(intent)) return { action: 'ai_replies' };
  return open ? { action: 'transfer' } : { action: 'log', due: 'next_business_day' };
}

Note the last line: anything the AI doesn’t recognize goes to a human. It is better to err on the side of an agent than to make up an answer.

An away message that doesn’t frustrate

The after-hours message is many people’s first contact with your business. It works better when it:

• says what the AI can do right now (“I can check your order, send a copy of your invoice or book a slot”);
• says when an agent will reply, with a day and time, not “soon”;
• offers buttons for the most common topics instead of asking the customer to describe everything;
• doesn’t promise what the bot can’t deliver.

Sample text to send within the window, as an interactive message with buttons:

“Hi, Ana! Our team is back tomorrow at 8 a.m. Meanwhile, I can help with your order, an invoice copy or a booking. If you’d rather talk to a person, tap Agent and I’ll put your request in the queue.”

Holidays and peak dates

Holidays are where 24/7 service fails most, because the business-hours rule gets hardcoded. Three precautions:

• Calendar as data, not code. Keep national, state and city holidays in a table your support team can edit.
• Reduced on-call team. On peak dates such as Black Friday and the year-end holidays, have an agent on call just for the “always human” queue. The Black Friday on WhatsApp guide has the sending calendar for those dates.
• A holiday-specific away message. “Today is a holiday, we’re back Monday at 8 a.m.” frustrates less than the generic message.

SLA: promise what you can deliver

The after-hours SLA is the promise the AI makes on behalf of your team. Set a deadline per topic (as in the JSON above) and measure whether it’s met. The metrics that matter most here are time to first human response and the share of tickets the AI resolves without a transfer. The article WhatsApp customer service metrics explains how to calculate first response time and handling time from webhooks.

Watch the 24-hour window: if the customer wrote at 11 p.m. on Friday and your team only replies on Monday, the window has already closed. To resume contact, you’ll need an approved template, not free-form text. That’s why, for sensitive topics, you should try to reply within 24 hours of the customer’s last message (the Saturday shift in the example helps with that) and keep an approved re-engagement template ready for when the deadline goes past that limit. The article on the 24-hour customer service window explains the full rule.

How much 24/7 service costs

Replies within a window opened by the customer are service messages. Since October 1, 2026, each phone number gets 1,000 free service messages per month; from the 1,001st on, each service message delivered to Brazil costs R$ 0.0350 (Brazilian reais). The free tier doesn’t cover utility templates, which since the same date are also charged when sent inside the window (R$ 0.0350 for recipients in Brazil).

Situation	Category	Meta cost (Brazil, since Oct 1, 2026)
AI answers a question inside the window	Service	Free up to 1,000/month per number; then R$ 0.0350
AI sends a booking confirmation as a template	Utility	R$ 0.0350
Agent resumes the conversation after 24 h with a template	Utility or marketing, depending on content	R$ 0.0350 or R$ 0.3217

A late-night bot that replies in many short messages burns through the free tier faster. Combine information into a single reply and use buttons, which cut down on back-and-forth. The article on utility messages in the customer service window details what changed in October 2026.

On top of Meta’s cost, there’s the cost of the AI model, billed by the provider you choose, and of your backend infrastructure.

How to do this with Vectalk

Vectalk is your access to the official API: it delivers Meta’s webhooks to your system and sends the replies your system produces. Vectalk has no customer service inbox or built-in AI in its dashboard; business-hours logic, the AI and the agent queue live in your backend or in tools connected to the API, such as n8n, Typebot or a CRM.

What you set up in Vectalk:

• Instance webhook pointing to the service that runs the AI (webhooks documentation).
• Ice breakers and commands for the number, such as /order and /agent, through the conversational automation endpoint.
• Utility templates to resume conversations after 24 hours, on the templates screen with the Template Library.

Billing is prepaid: each delivered message is deducted from your balance, and you don’t register a card with Meta. Create your account at apps.vectalk.com.br/register.

Frequently asked questions

Can the AI answer overnight with no human available?

Yes, as long as there is a clear escalation path, such as logging the request for a human with a stated deadline, or providing a phone number and email. The policy requires the path to exist, not that a human be online at all times.

Can the bot cancel a contract on its own?

Technically yes, but we don’t recommend it. Cancellation involves contract terms, fees and retention. The AI should log the request with date and time, confirm receipt and leave the decision to the team.

What if the customer wrote after hours and the 24-hour window closed?

Send an approved template to resume contact. If the content is the answer to the customer’s request (for example, “your ticket has been reviewed”), it tends to qualify as utility. See how categories work in WhatsApp template categories.

Do AI-sent messages count toward the 1,000 free service messages?

Yes. Meta doesn’t care whether a human or a bot wrote it: every free-form reply inside the window is a service message and counts toward the number’s monthly free tier.

Do I need to disclose that it’s an AI?

The WhatsApp policy cited above has no specific rule on this, but saying it’s a virtual assistant prevents frustration and makes it clear when the customer should ask for an agent. Check with legal counsel about rules in your industry.

---

Official sources consulted on October 29, 2026: WhatsApp Business Messaging Policy, WhatsApp Business Platform pricing, Law 13,709/2018 (LGPD), Vectalk documentation: webhooks and conversational automation.