Richer AI Agents Hub

The problem

The business bought cheap, low-intent leads in volume. Reps burned their day on people who would never buy, and the genuinely interested ones waited hours for a first reply — by which point they had cooled. The instinct was to hire more reps. The actual problem was that no qualification happened before human time was spent.

What I built

• Built a multi-agent architecture: one shared platform, several independent agents, each with its own WhatsApp number, prompt and configuration.
• Each agent opens the conversation in Hebrew within seconds, asks the qualifying questions a rep would ask, handles the common hesitations, and books a meeting when the lead is ready.
• Anything outside the agent's scope hands off to a human with the conversation history attached, so the rep starts informed instead of starting over.
• Instrumented every conversation with Langfuse, so prompt changes could be judged on booking rate rather than on how the output read.

The hard part

Knowing when to stop talking

The failure mode of a qualification agent isn't a wrong answer, it's refusing to let go — pushing for a booking when the person has said something that a human would recognise as a stop signal. Getting the handoff boundary right mattered more than any prompt tuning, because a bad handoff costs a lead permanently while a clumsy sentence costs nothing.

Outcome

Of 7,414 leads processed in the first eleven weeks, 1,664 held a real conversation, 293 booked a meeting and 86 closed. Reps stopped spending their day on unqualified leads.