The problem
A barbershop or studio lives in its messages. They arrive at eleven at night, on Sundays or mid-appointment. Whoever doesn’t get an answer in time goes somewhere else, and the slot nobody confirmed stays empty.
Many of these businesses also pay for ads. They pay for people to message them and then can’t keep up with the replies. For a small business, every empty slot is money that doesn’t come back.
What it does
The agent lives on the WhatsApp Business API, Instagram and Messenger. It answers questions about prices and hours, books and reschedules, saves spots in classes, sends reminders before each appointment and messages whoever didn’t show up.
The owner sees everything in a dashboard: bookings, customers and what the agent has handled. If they want to step into a conversation, they can. But day-to-day no longer depends on them holding the phone.
How we got here
We started with dentists. Then we opened up to dentists, studios and barbershops, and that’s when we added the AI agent, because almost nobody was doing it well. Later we decided to focus on dental, and ran into the fact that health has many sensitive situations an agent shouldn’t handle alone.
So we jumped to 1:1 appointment businesses like barbershops: more recurring and higher tickets, many already invest in ads, and almost all of them run everything on WhatsApp. Studios and class businesses stayed with us.
We didn’t pick the vertical we liked most. We picked the one the agent could handle well.
The hardest part
The agent. Not because it’s hard to build, but because there are very specific cases that only show up with real customers: the person who reschedules three times, the ambiguous message, the question nobody anticipated.
The only way to solve them is to monitor constantly and adjust until the agent is as smart and precise as possible. There’s no finished version; there’s a version that gets fewer things wrong every week.
What it cost me to learn
Our first customer was an early-stimulation school. It wasn’t our ideal customer and didn’t have a clear problem. We built features during the sale, they tried the agent for less than two weeks and never used the platform.
That’s where the rule I build by today came from: one validated pain, for one kind of customer, first.
Read the piece →How it’s built
The agent is orchestrated with LangGraph and runs on Vertex AI with Gemini Flash, connected to the WhatsApp Business API, Instagram and Messenger. The business dashboard is built in Next.js.
The conversation inbox is the same infrastructure Kipu uses. What we learn in one product we use in the other.
