A playbook for managers evaluating whether to automate customer service.
In Costa Rica, almost all of these conversations happen over WhatsApp, so much of the decision comes down to how an agent behaves on that channel — with your real customers and your real volume.
You don’t need to understand how the AI works under the hood. You need to confirm the agent will understand your business.
Here are the five criteria worth resolving before you sign, along with the concrete questions you can ask any vendor.
01 · Validate the scope
Before talking about models or features, define what problem the agent will solve and where its limits end. That decision is built jointly between the business and the vendor.
- Which queries can it resolve from day one?
- Which queries will still need a person?
- How does that scope change over time?
An agent that tries to solve everything usually ends up solving things poorly.
02 · Test how it behaves once it stops being a demo
A demo works with one conversation. Your business needs it to work with many, especially during your most demanding peaks.
- Is the agent built for your daily volume of queries?
- How does it behave during your demand peaks?
- What happens if you run a campaign and traffic doubles?
Scalability is also part of your customer’s experience.
03 · Understand the economic model
Two proposals with the same starting price can end up very different as usage grows: every conversation carries a real processing cost behind it.
- Is the cost per conversation fixed or variable?
- Who absorbs the consumption when volume rises?
- What can drive the bill up without anyone noticing?
- What strategies exist to keep the cost under control?
You’re not just buying technology, you’re also buying a financial model.
04 · The most important asset is the information
The value isn’t only in the agent, but in the knowledge it accumulates with every conversation: the questions, the patterns, your customers’ preferences.
- Where are the conversations stored?
- Can you access that data whenever you want?
- What happens to that information if you switch vendors?
That data is your business, not a technical detail in the contract.
05 · Define how it represents your business
An agent doesn’t just answer questions: it speaks on behalf of your brand and decides when to hand a case off to a person. You know how your customers should be handled.
- What can the agent say, and what can’t it say?
- How does it maintain your company’s tone?
- Does it have all the information it needs for its purpose?
- How is the quality of its responses measured?
Trust depends as much on what the agent says as on what happens when it can’t help.
In summary
AI can help your business. But only you know your operation.
- Validate the scope — what it can resolve, and what it can’t.
- Test the operation — what happens when it scales or fails.
- Understand the cost — the economic model and its growth.
- Protect your assets — data, conversations, and configuration.
- Define how it represents your business — brand, tone, and escalation.
Frequently asked questions
What is a conversational AI agent?
It’s a system that handles conversations with your customers, usually over WhatsApp or your site’s chat, interpreting what they ask in natural language instead of making them navigate a menu of options. Unlike a flow-based chatbot, it can answer queries nobody programmed one by one, and decide when to hand the case off to a person.
How much does a conversational AI agent cost?
You’ll find everything from subscriptions of 150 or 200 dollars a month to projects in the several thousands. The difference is rarely the technology: it’s how much comes already solved, and how much gets built around your operation.
A low subscription is usually a ready-made platform, with a standard configuration and a cap on included conversations. It’s a good way to start, especially if your case looks like a lot of other cases.
When the agent needs to know your product, follow your rules or connect to your systems, a setup cost appears and the monthly figure goes up: there’s dedicated infrastructure, and every conversation consumes processing from the AI model. If the channel is WhatsApp, Meta’s fee is added on top.
With either model, three questions organize the comparison: what the monthly fee includes, what happens when you go past that cap, and what the month when things go really well would cost.
How long does it take to implement?
It depends far more on the information than on the development. An agent with a well-defined scope and organized documentation can be up and running in a few weeks. What stretches the timeline is almost always the same thing: scattered information, processes nobody has written down, or a scope that keeps growing during the project.
Does an agent replace my customer service team?
No, and it’s worth not buying one with that expectation. What it does well is absorb repetitive queries and filter out the ones that truly need a person, so your team arrives with context on the cases that are worth their time.
Transparency: At Entrópica, we use AI as part of our research and editing process. The content reflects our knowledge, experience, and judgment.