In this article (14 sections)
In short: an AI agent for WhatsApp can handle the repetitive questions whose answers already live in your data, such as prices, opening hours, order status, availability and appointments. It hands anything that needs judgment, an exception or a negotiation to a person. It doesn't replace your team; it takes the first level of questions, at any hour. The WhatsApp rule that shapes the design most is the 24-hour window: outside it, you can only write with approved template messages.
This article explains what an agent can resolve on WhatsApp, what it shouldn't resolve, how it works underneath in business terms, and what it costs.
What it can handle well
An agent works best where the question is common and the answer sits in one of your systems. The most frequent cases:
- Product or service questions: what's included, how much it costs, opening hours, where you are.
- Status checks: an order, a delivery, a claim, a booking, an invoice.
- Scheduling: book, move or cancel an appointment against a real calendar, such as Google Calendar.
- Lead qualification: ask the basics (what they need, when, approximate budget) and leave the contact ready for your sales team.
- Document intake: a photo of a receipt, a prescription or a form, which the agent reads and records.
- After-hours questions: what used to wait until Monday.
The difference from a menu of numbered options is that the customer writes the way they talk, and the agent understands, asks what's missing and answers with real data instead of a fixed text.
What it shouldn't handle
This is what is usually left out when these systems are sold, and what a serious implementation takes the most care over:
- Complaints with strong emotion or legal weight. An angry customer needs a person.
- Price exceptions, returns outside policy, discounts. These are business decisions with consequences.
- Medical, legal or financial advice beyond general information.
- Anything the agent can't find in your data. If it doesn't know, it should say so and hand off, not make something up.
- Irreversible actions (charging, canceling, deleting) without a confirmation or a person in the loop.
A well-designed agent has an explicit list of what it can do and a clear handoff to a human for everything else. When the conversation leaves that perimeter, it goes to your team with the full history, so the customer doesn't have to repeat anything.
How it works underneath, without jargon
Piece | What it does |
|---|---|
WhatsApp Business number | Where messages come in and go out |
Official connection provider | Enables sending and receiving in a legitimate way |
The agent (AI model) | Understands what the customer writes and decides what to do |
Your data | Catalog, prices, calendar, orders: the source of the answers |
Business rules | What it can do, what it can't, when to hand off |
Inbox for your team | Where a person takes over the conversation when needed |
The agent doesn't answer from memory. It checks your systems at the moment of the conversation. That's why quality depends more on your data being in order than on which model you use.
The WhatsApp rules that shape the design
According to Meta's official documentation, two rules matter from the start of any project:
- The 24-hour customer service window. It opens when the customer writes to you and renews with every message they send. While it's open, you can reply freely. When it closes, you can only send templates approved by Meta.
- Per-message billing. Since July 2025, you pay per delivered message, based on its category and the recipient's country. Messages the customer sends are not charged, and each number gets a monthly allowance of free service messages, with charges after that as Meta specifies. Check the current page before projecting costs.
The practical consequence: if the agent has to write first (reminders, follow-ups), that happens with templates, and each one has a cost.
What it costs
There are two different costs, and it helps not to mix them:
- Building the agent, which you pay once. Atomix's price for an agent like this is USD 5,000 to 10,000, depending on how many systems it consults and what actions it can take, with a free working demo in 24 to 48 hours before we quote.
- The monthly running cost, paid to third parties: WhatsApp messages at Meta's rates, use of the AI model and hosting. Each is paid directly to its provider and scales with the number of conversations.
To see what to automate first, read AI agents for business: what they actually automate.
Examples by type of business
- Clinic or medical office: the agent offers open slots, books the appointment and sends the reminder. Clinical questions always go to a person.
- Distributor or retailer: reports the status of an order, the account balance and product availability, and takes repeat orders from known customers.
- Real estate agency: answers about available properties, qualifies the prospect and books the visit with the agent.
- Business services: takes the quote request, asks the basic questions and passes an organized case to the sales team.
In all of them the agent's job is the same: resolve the simple things and hand over, well prepared, the ones that aren't simple.
How to measure whether it works
Define three numbers before you start and track them every month:
- Share of conversations resolved without human intervention. A healthy starting point is lower than what any vendor promises; it rises as the answers are tuned.
- Time to first response. It should drop to seconds.
- Share of correct handoffs. That is, how often the agent passed a conversation to a person when it should have.
Also review a sample of conversations every week during the first month. It's the only reliable way to find weak answers before a customer does.
How to implement it without unnecessary risk
An order that works: pick three or four types of question that repeat, write the right answers with your team, test the agent with real conversations from the last month, and only then open it to customers, with the handoff to a person active from day one. Scope grows when the numbers justify it.
If you want to see what this would look like for your business, you can explore Atomix's AI agent services.
Frequently asked questions
Can the agent make mistakes?
Yes, like any new employee. You reduce it with organized data, a limited scope, clear handoff rules and a weekly review of conversations. What is not acceptable is invention: if it can't find the information, it should say so and pass the conversation to a person.
Do I need the official WhatsApp API?
For an agent that serves customers on an ongoing basis, yes. It is the legitimate and stable way to connect a system to WhatsApp, and it has clear rules for templates and billing. Unofficial shortcuts don't come with those guarantees.
Does the customer know they're talking to an agent?
It's best to say so in the first message and offer to reach a person at any time. It builds trust and avoids confusion.
Can I keep my current number?
Usually, yes, but that number stops working as a regular app account and runs through the platform instead. Decide this with your team beforehand, because it affects how you serve customers today.
How long does it take to go live?
The demo is ready in 24 to 48 hours. A first production version, with a few types of question, usually takes less than a month.
If you want to evaluate your case with a concrete conversation, you can book a call.