In this article (18 sections)
In short: an AI customer service agent answers questions on WhatsApp, your website or email using your company's information, and hands to a person anything it can't resolve. It's implemented in five steps: choose which questions it handles, gather the information, connect it to your systems, test it with real conversations and roll it out gradually. On WhatsApp, you also pay Meta per message, a cost separate from the development.
What it does, and what it doesn't
A customer service agent isn't a menu with buttons. It understands what the person writes in their own words, looks for the answer in your company's information and, when needed, checks or updates data in your systems: the status of an order, the time of an appointment, the balance of an account.
What it shouldn't do is make things up. A well-built agent answers only with what it knows, and when it doesn't know or the customer is upset, it hands off to a person with a summary of the conversation. That handoff is part of the design, not a failure.
Which questions to automate
Not everything should be automated at first. The best candidates are questions that repeat and have clear answers:
Type of question | Example | Level of automation |
|---|---|---|
General information | Hours, location, payment methods | Full |
Status checks | "Where is my order?" | Full, if connected to your system |
Simple requests | Change an appointment, resend an invoice | High, with validations |
Product questions | Prices, compatibility, requirements | High |
Complaints | Duplicate charge, damaged product | Partial: the agent gathers details and hands off |
Sensitive cases or anger | Threat to cancel, serious complaint | A person handles it |
The practical rule: start with the three or four questions that come up most. If you don't have that data, spend a week counting them in the channel you use today.
The five steps of an implementation
1. Define the scope
Which questions it handles, on which channels and within which limits. Also what counts as "resolved" and when it should hand off. This definition prevents the most common mistake: an agent that tries to answer everything and gets part of it wrong.
2. Gather the information
The agent needs reliable material: frequently asked questions, change policies, price lists, manuals. If that information is outdated or contradicts itself, the agent repeats it. This step almost always reveals that the problem wasn't the AI, but that nobody had written the rules down.
3. Connect it to your systems
To look up an order or book an appointment, the agent connects to your system through an integration: an automatic link between two programs so they exchange data. Each action is defined separately and with limits; for example, it can look up an order but not change a price. For more on where time and money get lost when systems are connected, see how much it costs to integrate two systems.
4. Test with real conversations
Hundreds of past conversations are taken and you check what the agent would have answered. You measure how many it answers correctly, how many it hands off and where it gets things wrong. You fix, and repeat. Nothing goes live until serious errors are under control.
5. Roll it out gradually
First it handles a share of the questions, or a time window (for example, nights and weekends). A person reviews the conversations. When the results are good, the scope expands.
The channel: WhatsApp and its separate cost
WhatsApp is where many customers write in most markets. To connect an agent, you use Meta's official platform, which lets you receive and send messages from your own system, with delivery status updates and approved templates for reaching out first (WhatsApp Business Platform documentation).
This introduces a cost that doesn't depend on development. Meta charges per message delivered from the business to the customer, with rates that vary by category and by the recipient's country; messages the customer sends to the business are not charged. In addition, as of October 2026, its pricing page indicates that service messages within the 24-hour window are charged after the first 1,000 per number per month (WhatsApp Business Platform pricing). Terms change, so verify them when you budget.
How to measure whether it works
Before launch, define three numbers and review them every week:
- Share of questions resolved without human intervention. That's the direct benefit.
- Share handed off to a person. It shouldn't be zero: it shows the agent knows its limits.
- Customer satisfaction, with a short question when the conversation closes.
If the first number rises while the third falls, the agent is resolving quickly and badly.
Common mistakes
- Launching without organized information. The agent is only as good as what you give it.
- Not offering a way to reach a person. A customer stuck with a machine leaves.
- Not reviewing conversations. In the first weeks you learn more by reading than by measuring.
- Giving it too many permissions. Each action on your systems is authorized one at a time.
For a broader view of which company tasks an agent handles well, this guide to AI agents for business complements this article.
What it costs
The cost of a customer service agent at Atomix starts at USD 3,000 for the agent, plus monthly maintenance from USD 300; the final price depends on how many systems it checks and what actions it takes. Before quoting, we deliver a free working demo in 24 to 48 hours using your real questions, and in most cases it reaches production in under a month. Third-party usage costs, such as the messaging platform and the AI model, are budgeted separately. The service is described on our page about AI agents for business. If you want to see how it would answer your questions, you can book a call.
Frequently asked questions
Does the agent replace my support team?
It takes over the repetitive part and leaves your team the cases that require judgment or empathy. In practice, the team handles fewer simple questions and more complex cases.
What happens if it doesn't know the answer?
It's configured to say so and hand the conversation to a person, with a summary of what has already been discussed, so the customer doesn't have to repeat everything.
Can it support several languages?
Yes. An agent can reply in the language the customer writes in, as long as the base information is available or can be translated with review.
How long until it's running?
On most projects, under a month. You start from a demo built on your real questions, which lets you decide based on something concrete.
Is my customers' data safe?
It depends on how the agent is built. You limit which data it can access, log its actions and define where the information is stored, and keep regular backups of the data the system depends on.