WhatsApp AI agent for business: what it can do, what it cannot, and the rules

A WhatsApp AI agent for business answers customers' typed messages from your own prices, policies and stock, and passes the chat to a person when it should. This guide covers what it can and cannot do, how it copes with Kuwaiti Arabic, what its replies cost after Meta's 1 October 2026 pricing change, and which Meta rules apply.

For the broader idea of agents that take actions, not only answer, see what AI agents are.

Rule-based chatbot vs AI agent

A rule-based WhatsApp bot follows a tree you design in advance. The customer taps a button or types a keyword, and the bot sends the reply attached to that branch. It is predictable and cheap to run. It breaks the moment someone asks something the tree did not anticipate.

A knowledge-grounded AI agent reads the customer's message as free text, looks up the relevant facts in material the business has supplied (price list, delivery areas, opening hours, return policy, stock), and writes a reply from those facts. It can also take actions you allow, such as recording a lead or sending a payment link.

You can combine the two: buttons for flows with a fixed shape, and the agent for everything typed.

How the agent handles one WhatsApp message

1. A customer messages your business number. This opens a 24-hour customer service window.
2. Your messaging platform passes it to the agent with the recent conversation.
3. The agent checks the handoff rules first: did the customer ask for a person, or is this a complaint or refund?
4. It looks up the answer in the knowledge base or a connected system, such as stock or orders.
5. It sends one complete reply inside the window. Each delivered reply is a service message.
6. If it has no answer, it tags the chat, assigns it to a staff member and stops replying.

What a knowledge-grounded agent can do

• Answer factual questions about prices, delivery areas, opening hours, sizes, availability and policies, using the current version of your data.
• Collect details a person would otherwise ask for: name, area, preferred time, items, budget.
• Qualify and route leads and support requests to the right team, with tags your staff can filter.
• Record an order and send a payment link, if it is connected to your catalogue and a payment gateway. In Kuwait that usually means a KNET payment link sent inside the chat.
• Reply at 2 a.m. inside the 24-hour customer service window, when nobody is on shift.

What it cannot do, or should not

• Know what you did not give it. If the price list in the knowledge base is out of date, the agent quotes the old price. Keeping the knowledge base current is the main ongoing job, and it needs a named owner.
• Make judgement calls with money or liability. Refunds outside policy, discounts, medical or legal answers, and complaints belong with a person.
• Start a conversation in free text. The WhatsApp Business Messaging Policy says that outside the 24-hour customer service window you may only send approved message templates.
• Act as a general assistant. Meta's terms bar that use (see below). Have the agent decline off-topic requests.
• Be right every time. AI output can be wrong. Treat every reply as if a new employee had sent it: useful, and reviewed.

Meta's 2026 rules for AI on WhatsApp

Three rules matter, from two Meta documents. This is a technical summary, not legal advice. Read the originals before you launch.

1. The AI Providers clause (terms section 4.7)

The Meta Terms for WhatsApp Business Platform define AI Providers as providers and developers of large language models, generative AI platforms, general-purpose AI assistants, or similar technologies as determined by Meta. They are strictly prohibited from using the platform to provide those technologies when they are the primary (rather than incidental or ancillary) functionality on offer. Meta's developer documentation dates this to 15 January 2026 and notes an exception in markets where Meta is legally required to allow AI Providers.

In plain terms: a "ChatGPT on WhatsApp" product is out. A clinic's booking assistant, where AI is the means of serving the clinic's own patients, is the kind of use the clause does not target. Meta decides, in its sole discretion, which technologies count, so keep the agent's scope narrow and tied to your business.

2. No training on WhatsApp data

The same section says you may not allow WhatsApp Business Platform Data to be used to create, develop, train or improve machine learning models, with an exception for fine-tuning a model for your exclusive use. Ask any provider, in writing, whether your customers' conversations train their models.

3. Escalation paths for any automation

The WhatsApp Business Messaging Policy (last updated 23 September 2026) does not mention AI by name. It says you may use automation when responding during the 24-hour window, but must also have prompt, clear and direct escalation paths. The options it lists are in-chat human agent transfer, a phone number, email, web support, in-store visits, or a support form. The same policy requires opt-in before you contact people and requires you to respect opt-out requests.

What the agent's replies cost from 1 October 2026

Until 30 September 2026, Meta does not charge for non-template messages inside the 24-hour customer service window, and utility templates sent inside an open window are free too. From 1 October 2026, Meta charges service messages per message, at the same rate as utility messages in each market. A service message is any non-template message not powered by Meta Business Agent, so replies written by a third-party AI count. Utility templates sent in reply inside an open window, such as an order confirmation, are charged from the same date. Meta notes that the AI provider and the messaging provider typically bill separately.

Two practical effects:

• One complete answer costs less than three short ones, because each delivered message is billed on its own.
• Chats that start from a Click to WhatsApp ad or a Facebook call-to-action button open a 72-hour free entry point window if you reply within 24 hours. Meta says delivery inside that window stays free.

Kuwait's rates and worked monthly bills are in WhatsApp Business API pricing in Kuwait.

(Meta's separate 2026 charge on AI Providers applied only in markets where it must legally admit them: Brazil from 11 March 2026, and Italy and other European countries until 12 May 2026. Kuwait is not on that list.)

Handling Kuwaiti and Gulf Arabic

Many customers in Kuwait write to businesses in Kuwaiti dialect, not formal Arabic. They drop in English words, sometimes use Arabizi (Arabic in Latin letters with digits, such as "3" for ع), and switch language mid-chat. Area names arrive in several spellings: Salmiya, Salmiyah, السالمية.

The AraDiCE benchmark (COLING 2025) tested Levantine and Egyptian Arabic, plus a cultural test covering the Gulf, Egypt and the Levant. It found that Arabic-centric models outperform multilingual ones on dialect tasks, that significant challenges persist in dialect identification, generation and translation, and that models understand dialect better than they generate it. Gulf dialect was not in its dialect tests, which is one more reason to test on your own Kuwaiti chats. In practice:

• Separate understanding from replying. The agent must understand dialect. Its replies can be in clear, polite Arabic close to the customer's tone, rather than forced dialect that can sound off.
• Write synonyms into the knowledge base. Product nicknames, English loanwords and alternative spellings of areas and items help retrieval find the right fact.
• Normalise numbers. Customers mix Arabic-Indic digits (٠١٢) and Western digits (012) in phone numbers and order IDs. Convert both before lookups.
• Test on your own chats. Run a sample of anonymised past conversations through the agent before customers see it. Count wrong answers, not just answered ones.
• Match the customer's language. If a customer switches to English, the reply should follow.

Human handoff: the guardrails that matter

Triggers that should always hand off:

1. The customer asks for a person, in any phrasing ("أبي أكلم موظف", "can I talk to someone").
2. The knowledge base has no answer. Guessing is worse than handing off.
3. Money or liability: refunds, disputes, complaints, anything health or legal.
4. Repeated questions or clear frustration.
5. High-value leads your sales team wants to own.

What a good handoff looks like:

• The staff member sees the full conversation, not a summary alone.
• The agent tells the customer a person will reply, and roughly when if it is outside working hours.
• The agent stays silent once a person has taken over, so the customer does not get two voices.
• Every reply the agent sent can be traced back to the data it used.

Meta Business Agent or a custom agent?

On 3 June 2026 Meta announced Meta Business Agent, an AI agent that answers customer questions about a business, its products and pricing, recommends products and gathers customer information. Meta says it learns from Facebook Page posts, past WhatsApp chats and information you add, such as your website, product details and price lists, and you can tell it to transfer complex or sensitive topics to you. It is available to eligible businesses in select markets. The same announcement introduced the Meta Business Agent Platform for larger businesses, which Meta says works alongside the WhatsApp Business Platform.

On price, Meta said at launch that getting started is free, with paid subscription options to follow. Its developer documentation now says Meta Business Agent messages have been charged since 1 August 2026 at $2.00 per million tokens, which Meta estimates at roughly 4 to 5 US cents per message, and that they are charged even inside the 72-hour free entry point window.

For a single owner answering from one phone, check whether it is offered on your number. A custom agent makes more sense when the agent must read live stock or a CRM, when payments run through a Kuwaiti gateway, or when you need your own handoff rules, logs and data handling. The trade-offs between the app and the platform are in WhatsApp Business app vs API.

Waslo as a working example

Waslo is the messaging platform Rukn built and runs. It shows how the pieces above fit together:

• One team inbox across a web chat widget, WhatsApp, Instagram, Telegram, LINE, SMS and email, plus Twilio and API connections.
• A per-business AI agent that replies from that business's own knowledge base (its real prices, policies and stock) in the customer's language, Arabic or English.
• The agent answers common questions, qualifies leads, routes conversations and collects details. Complex or high-stakes cases go to a person with full conversation context, and a human agent can take over any conversation.
• It can take an order and send a KNET payment link from the business's own licensed payment gateway account inside the chat. Funds settle to the business and never pass through Rukn.
• The business can switch the AI agent off at any time.

On data: when a business enables the agent, message content and that business's knowledge base are sent to third-party AI providers to generate replies. Those providers are Rukn's sub-processors, bound by contract, and do not use the content to train their models. AI replies can be wrong, and the business remains responsible for them, as it would for replies from its own staff. Meta bills message charges directly to the business's own WhatsApp Business Account. Waslo itself is priced as a monthly subscription plus a one-time setup fee. On Waslo, how AI usage is priced is set out in your subscription quote. For hosting questions, see our note on data residency and cloud rules in Kuwait.

A launch checklist

1. For a custom agent, move the number to the WhatsApp Business Platform. The steps are in our WhatsApp Business API integration guide.
2. Export last month's chats and list the 30 most common questions.
3. Write the knowledge base from those questions, with a named owner for prices, stock and policies.
4. Write the no-go list and the handoff triggers before you write any prompts.
5. Test on anonymised real messages, including dialect, Arabizi and English.
6. Start narrow, for example after hours only or one product line, and widen when wrong answers are rare.
7. Review weekly: unanswered questions, handoffs and corrections. Each one is a knowledge base update.
8. Track four numbers each week: share of chats resolved without a person, handoff rate, wrong answers found in review, and monthly Meta message spend.

To have Rukn write the knowledge base, set the handoff rules and run the agent on Waslo for your team, see our WhatsApp and messaging services.