AI Agents for Customer Support: A Safe WhatsApp Setup

Can an AI agent really answer customers on WhatsApp and other channels?

Yes. An AI agent can read a customer's message, look for the answer in material you give it and reply in the chat, whether that chat is on WhatsApp, your website, email or a social channel. With the right connections it can also do small jobs, such as booking an appointment or opening a ticket.

It differs from an older keyword bot, which follows a fixed menu and breaks when the customer words things differently. An agent copes with loose wording and follow-up questions because it generates each reply. That same quality creates the main risk. NIST, the US standards body, lists confabulation among the risks of generative AI: confidently stated but false content, which it calls a natural result of how these models are designed. So nobody can honestly quote you a rate of correct answers for your business before testing on your own conversations.

Your business stays responsible for what the agent says. In 2024 a tribunal in British Columbia, Canada held Air Canada liable after its website chatbot gave a customer wrong advice about bereavement fares, and it rejected the airline's argument that the chatbot was a separate entity responsible for its own actions. That ruling does not bind courts in Saudi Arabia, Jordan or Iraq, so ask a local adviser about your own obligations. Until you hear otherwise, plan as if a wrong answer from the agent is a wrong answer from you.

What should an AI agent handle, and what should stay with people?

Give the agent questions with one correct answer that you have already written down, plus small actions that are easy to undo. Keep anything that needs judgment or authority with staff. A useful test: if a new employee in their first week would have to ask a supervisor, the agent should hand it over too.

Good first jobs include opening hours, branch locations, the documents a customer should bring, delivery areas, booking or rescheduling, and collecting details before a person takes over. Prices, stock and order status belong on that list only if the agent reads them from a source someone keeps current.

An agent will not notice that a long relationship is going wrong, and it cannot know a policy you never wrote down. Leave these to people:

Complaints, angry customers and anything that could end in a dispute or a legal claim.
Refunds, discounts, credit terms and any exception to a written policy.
Medical, legal and financial advice. A clinic's agent can book the visit, and a clinician answers the question about symptoms.
Contract changes, account closures and other steps that are hard to reverse.
Any case where the customer asks for a person or the agent has already failed to help.
Anything that needs a full card number, financial account number or ID card number. WhatsApp's policy tells businesses not to ask for these in the chat at all.

What do Meta's rules say about AI on WhatsApp?

Meta's rules treat AI companies offering an assistant differently from businesses using automation to serve their own customers. Its WhatsApp Business Platform terms, last updated September 23, 2026, prohibit AI providers from using the platform when AI is the primary, rather than incidental or ancillary, functionality on offer. The definition names large language models, generative AI platforms and general-purpose assistants, and Meta decides what counts as primary.

An agent that answers questions about your own products, orders and bookings reads as incidental to your business, while an open ask-me-anything assistant is what the clause targets. That is our reading of it. The terms do not mention customer service by name, so keep the agent on your business topics and ask your provider or adviser if your case looks close. The same section allows a business to keep an AI provider as its solution provider.

The rules below shape how you set an agent up. Meta changes them, so read the current text before launch:

Automation is allowed when responding inside the 24-hour window, and the policy requires prompt, clear and direct escalation paths. It lists in-chat transfer to a human agent, a phone number, email, web support, in-store visits and a support form.
The window runs for 24 hours from the customer's last message or call. Once it closes, only pre-approved templates can go out, so the agent cannot keep sending free-form follow-ups.
You need opt-in before contacting people on WhatsApp, and you must respect requests to stop.
WhatsApp Business Platform data may not be used to build or improve AI systems, apart from fine-tuning a model for your exclusive use. Ask any vendor in writing whether your chats train shared models.
The policy limits health information where local regulations require heightened safeguards, and it restricts some regulated goods and services. Clinics and regulated businesses should confirm their position with an adviser.

How do you give an agent accurate knowledge?

Give it a small set of checked documents, each with a named owner, and remove anything out of date. An agent working from a folder of old price lists and draft policies will quote them with full confidence.

Keep facts apart from behavior. Facts such as prices, policies and delivery terms belong in knowledge articles that can change without touching the agent's instructions. The instructions cover tone, what to ask for, what never to promise and when to hand over. Our guide on how to build a knowledge base covers writing and maintaining those articles. Before you connect anything:

Start from questions customers really ask. Pull the last few months of chats, group them by topic and write answers for the most common ones.
Write one clear answer per question and spell out the conditions, such as which branches, which customers and until what date.
Remove conflicts. Two files with different return periods will produce answers that change from one customer to the next.
Give each article an owner and a review date, and add the agent's knowledge to the checklist for any price or policy change.
Tell the agent what to do when the answer is missing: say it does not have that information and offer a person.
Sync only web pages and files you control. OWASP warns about indirect prompt injection, where instructions hidden in websites or files change a model's behavior.

How should an agent hand off to a person?

A handoff should be easy for the customer to trigger and should bring the whole case with it. On WhatsApp you need an escalation path anyway, so design it before you write anything else.

The person who receives the case should see the full chat, a short summary, the reason for the handoff and what the agent has already told the customer. Each handed-over chat needs one owner, which is what a shared inbox gives you. Make sure the agent stops answering once a staff member takes over, so the customer does not get two voices at once.

Plan for the hours when nobody is there. A handoff at 11 at night that lands in an empty queue leaves the customer waiting with no idea for how long. Have the agent state your working hours, collect the details and tell the customer when someone will reply.

Hand over when any of these happens:

The customer asks for a person, in any wording or dialect.
The topic is on your staff-only list.
The customer is upset or keeps repeating the same question.
The knowledge has no answer, or the agent has tried twice without helping.
An action fails, or the request falls outside what the agent is allowed to do.

Can an AI agent reply well in Arabic?

It can reply in Arabic, but you have to judge the quality on your own customers' messages, because a general claim of Arabic support says little about how Saudi, Jordanian or Iraqi customers actually type.

Decide the voice first. Plain Modern Standard Arabic with a polite tone is a safe default for business replies, and the agent should answer in the language the customer used. Write your most used knowledge articles in Arabic too, with your own terms for products and services, so the agent is not translating English policies as it goes.

Check these with real examples:

The same question in Saudi, Jordanian and Iraqi wording.
Arabic typed in Latin letters, and messages that mix Arabic and English.
Spelling variants and typing mistakes, which are common on phones.
Gender. Arabic addresses men and women differently, so decide whether to use neutral wording or ask.
Numbers, dates and currency, including digits written both ways (٣ and 3) and amounts in SAR, JOD or IQD.
Product and brand names, which the agent should not translate or rename.
Voice notes, if your customers send them, and what the agent does when it cannot understand one.

How do you test an AI agent before it goes live?

Test it privately against real past conversations, then release it in stages so your team reviews replies before customers depend on them.

Build a test set of a few dozen real messages with personal details removed, and write down the answer you would accept for each. Have a support person score the results, since they know what a wrong answer costs. Run the same set after every change to instructions or knowledge, because a fix in one place can break an answer elsewhere.

Include hard cases on purpose: a question the knowledge cannot answer, a demand for a discount, an angry customer, someone asking for a person, and a message telling the agent to ignore its instructions. OWASP calls that last one prompt injection and recommends limiting a model's access to what it needs, with human approval for privileged operations. Apply this to actions too: give the agent only the ones it needs, and make sure test runs do not touch real customer records.

Then roll out in steps:

Suggested replies only. Staff see each draft, decide whether to send it and mark the bad ones.
Automatic replies for one or two simple topics on one channel, during working hours when someone is watching.
A daily review of a sample of chats in the first weeks. Fix the knowledge before you add topics.
Wider scope and after-hours replies, once reviews have stayed clean for a period you set in advance.
A named person who can pause the agent, and who has tried doing it.
A clear statement to customers that they are talking to an automated assistant and how to reach a person. Ask an adviser about local rules on this and on privacy, including where messages are processed.

How do you measure whether an AI agent is working?

Measure the quality of answers first and volume second. A high share of chats handled by the agent may only mean that customers gave up.

Record your current numbers before launch so you have your own baseline, because outside benchmarks will not tell you how the agent does with your customers. Review weekly at the start and at least monthly after that. If wrong answers rise, cut the agent back to the topics that hold up. Track these:

Wrong answers, from a sample of replies that a person has scored.
Answer rate, meaning chats the agent finishes without a person, read together with repeat contacts.
Repeat contacts, meaning customers who come back about the same issue within a few days.
Handoff rate and the reasons behind it. Neither a high nor a low figure is good by itself, so look at why people were handed over.
Time to reach a person after a handoff.
Unanswered questions, which give you the next articles to write.
Customer ratings and complaints, compared with chats that staff handled.
Cost per conversation, including AI usage and the staff time spent reviewing.

How does Repzo handle this?

Repzo Workstation has two kinds of AI. Fox AI is the built-in assistant for your team: it drafts replies in the inbox, summarizes conversations and tickets and answers questions about your data, and sensitive actions need approval first. For customers, you build AI agents that work in the shared inbox on chat, email, WhatsApp and social channels, and on tickets.

Each agent has a playbook covering its identity, tone, working hours, what to ask for, guardrails and when to hand off, with ready-made sections in English or Arabic. Its knowledge comes from Content Hub articles, uploaded files or a synced website, and language is a setting on the agent. You choose the functions it may run, such as booking an appointment, creating a ticket or sending a WhatsApp template.

Each deployment has a mode, including Suggest a reply and Auto-reply, so you can follow the staged rollout above. Handoff settings cover the message the customer gets, who receives the conversation and a case summary. Handed-over chats appear under AI handoffs with a brief explaining why, and staff can use Take over, Resume AI or a private Instruct AI note.

You can test a draft in the editor, or use Test in inbox, where you write as the customer and choose Simulate actions or Run real actions. Watch one detail: until your first publish, every save goes live, and knowledge and deployment changes always apply immediately. Test before you connect a live channel.

For measuring, Feedback groups the replies your team rated good or bad and lists questions the agent could not resolve. Analyze shows run counts, answer and handoff rates and average cost, and it compares test sets across versions. Activity shows recent runs with their reasoning, tool calls and handoffs.

These tools support your review and do not guarantee correct answers, so the checks in this guide still apply. AI work is paid from a prepaid AI balance, and customer-facing agents come with the Inbox & Calls add-on. Every plan starts with a 7-day free trial. To find out what fits your channels, ask our sales team.

Frequently asked questions

Does WhatsApp allow AI chatbots for customer service?

WhatsApp's policy allows automated replies inside the 24-hour window, with a clear path to reach help. Meta's terms bar AI providers when AI is the primary functionality on offer. They do not name customer service, and Meta decides what counts as primary, so keep the agent on your own business topics.

How accurate will an AI agent be?

Nobody can tell you in advance. Accuracy depends on your knowledge, your customers' wording and how narrow the agent's job is. Score a test set from your own conversations, then keep scoring a sample of live replies.

Will an AI agent replace my support team?

Plan on keeping people. Someone has to take handoffs, deal with complaints and exceptions, review replies and keep the knowledge current. Whether it frees up staff time is something to measure in your own business once it is running.

Can an agent message a customer after the 24-hour window closes?

Only with a pre-approved template, and only if the customer has opted in. Free-form replies stop until the customer writes or calls again.

Should customers know they are talking to AI?

We recommend telling them at the start of the chat and explaining how to reach a person. That sets fair expectations and avoids misleading anyone. Ask an adviser whether local rules require specific wording.

Can one agent work on WhatsApp, Instagram and email?

Yes, if your software connects those channels, and the same knowledge can serve all of them. Rules and habits differ by channel, so check each platform's terms and test separately. A reply that suits email may be far too long for WhatsApp.