When to use an AI agent and when to use a person

What an agent does well

• Answering questions that repeat every day: hours, requirements, payment options, shipping.
• Covering after-hours messages and logging everything for the next day.
• Collecting details for an appointment, a quote or an order.
• Organizing the inquiry before handing it off, so the person doesn't start from scratch.
• Replying quickly to many messages at once during peak times.

What a person should handle

• Complex or high-ticket sales, where trust is built through conversation.
• Complaints and upset customers.
• Sensitive topics: health, debt, personal or legal problems.
• Cases that aren't in the agent's information or that need a decision.
• Negotiating prices or special terms.

How to combine them

1. Define which topics the agent answers alone and which it always hands off.
2. Have the agent say it's an assistant and offer to connect with a person.
3. When it hands off, pass along the conversation with full context.
4. Review conversations every week to adjust answers and limits.

A good indicator is how many conversations the agent resolves on its own and how many it hands off. If it hands off almost everything, it's missing information. If it never does, it's probably answering things it shouldn't.

Before launch

Test it with real conversations from the last month and measure what it costs to run: AI usage and WhatsApp messages both cost money. That way you know whether it pays off before opening it to all your customers.

An example of splitting the work at a clinic

To make it concrete, think of a clinic that gets a lot of inquiries on WhatsApp. Here's how the work could be split between the agent and the team:

• The agent answers questions about hours, location, accepted insurance plans and how to prepare for a test.
• The agent collects the details for an appointment request and leaves them ready for the front desk to confirm.
• After hours, the agent answers common questions and logs the ones that need a person.
• The front desk handles rescheduling with special conditions, complaints and any question about symptoms.
• Any medical question is answered by a professional, never by the agent.

The result is that the team stops answering the same question ten times a day and spends its time on the conversations that truly need a person.

How to tell if the split is working

After the first few weeks, it's worth looking at a few simple numbers. You don't need a complex dashboard: reviewing them once a week is enough.

1. What share of conversations the agent resolves without handing off.
2. How long it takes a person to pick up a handed-off conversation.
3. Which questions the agent couldn't answer, so you can add them to its information.
4. Whether there were wrong or out-of-place answers, and why.
5. What customers think, if you ask them at the end of the conversation.

With that data you can fine-tune everything: which topics to add to the agent, which to remove and when it should hand off sooner.

The most common mistake

The most common mistake is launching the agent without defining what it must not do. An agent without limits tries to answer everything, and that's where made-up answers and promises the business can't keep show up. Starting with a few topics handled well and adding more over time is slower, but it builds trust with both your team and your customers.