On a lean team, the answer is not to pick one of the two. The AI agent covers the first reply and the hours when nobody is there to answer, and the person keeps what requires a decision.
The rest of this post shows how that split works at businesses I serve and how to find where it falls in yours.
Who answered before
At the businesses I serve, before the agent, the person who answered WhatsApp was the owner or one employee. That is my general answer, not a measurement, and I did not confirm it case by case.
In my opinion, for a team that size the question is not who to replace with whom. The question is what stays with each one.
What stays with the agent
Not every message arrives during business hours. At a car audio store I serve, 38.9% of customer messages arrived outside the store’s hours: 8,816 of 22,650, between January 19 and August 18, 2026. The hours used in the count are the rule in the store’s registration, and I was the one who applied that rule to the data.
Between October 11 and November 10, 2026, that store’s agent sent 5,516 replies. Of those, 2,108 went out outside the store’s hours.
- 5,516replies from the agent
- 2,108replies outside the store's hours
From April to July 2026, that agent’s median response time stayed between 3 and 6 seconds, inside and outside business hours. I have no measurement from before the agent to compare that time with.
That is where the agent works: on the first reply and on the messages that arrive while the store is closed. It joins consecutive messages from the same customer, answers based on the store’s catalog and rules, and simulates installment plans in 3, 6 and 10 payments.
At three businesses I serve, 2,703 of 7,661 agent replies went out outside business hours, 35.3%, between October 11 and November 10, 2026. That comes to about 45 hours, on an estimate of 1 minute per reply. That premise has no measured basis and only gives an order of magnitude.
What stays with the person
The same store agent hands the conversation to a person when the subject goes beyond what it knows. In other words, it does not solve everything on its own.
At a dental clinic I serve, the split is built into the agent’s design. It works like a receptionist that handles the conversation, qualifies and hands off. It does not book appointments. The clinic’s receptionist does the booking.
- 01Handlesthe agent receives the patient's message
- 02Qualifiesthe agent qualifies the patient's request
- 03Hands offthe agent passes the conversation on
- 04Booksthe receptionist schedules the appointment
In my opinion, the criterion for the split is this. If the answer already exists in the catalog, the rules or a table, it can stay with the agent. If it depends on someone deciding, such as booking a time slot or handling a case outside the rules, it stays with the person.
What these numbers do not show
The numbers above measure service: a message that arrived and a reply that went out. There is no period before the agent in the data, because it was already running from the first month of collection. There is also no record of sales or of separate human replies in the messages.
So it is not possible to measure how much time the person stopped spending, or to compare human service with the agent’s. What can be stated is coverage of hours.
Make the split with your own conversations
Take the conversations from an ordinary week on the company WhatsApp. Sort each one into two piles. The first is “ready answer”: the answer already existed in the catalog, the price list or the business rules. The second is “needs a person”: someone had to decide something. Also mark which conversations arrived outside your hours.
The ready-answer pile shows what an agent can cover. The needs-a-person pile shows what stays with you or your employee. If a good part of the first pile arrived outside working hours, that is where the agent starts working.
If you want to see how the WhatsApp agent I build works, the details are in services.