CRM for EDISAN with an LLM agent
EDISAN
EDISAN wanted a bot that felt like a person. We told them it's almost impossible to make it undetectable, but they decided to move forward anyway. The agent detects and qualifies leads, makes the initial property offer, and hands off to a real salesperson when needed. We also imported all the data from their previous CRM into the new system, which they now use to manage what they sell and what they don't. Something they didn't have before. The hardest part was making sure the agent didn't make things up. EDISAN has a very specific way of speaking to clients and they wanted to replicate it exactly. It took a long process of evaluating models until we found the right fit: we ended up with a heavily tuned GPT-4o for communication, plus other models inside the graph for filter detection, routing, and information extraction.
Challenges
- Making sure the agent didn't invent things. EDISAN had a very specific way of talking to clients and they wanted to replicate it exactly.
- Finding the right model. Long evaluation and tuning process until we found the one that best matched their communication style.
- Migrating all the data from the previous CRM without losing anything.
Solutions
- GPT-4o for client communication, with heavy prompt and context work.
- Different models inside the graph for filter detection, routing, and information extraction.
- Full migration of existing data into the new CRM.
- Automatic handoff to a real salesperson when the lead is qualified.