AI inbound sales agent: 4-second responses across 40 enterprise demos

The situation

• An AI sales agent startup had landed 40 enterprise demos in two weeks on the strength of their Lovable UI.
• The actual agent, supposed to qualify inbound leads in chat and update HubSpot, wandered off-topic, lost memory mid-conversation, and silently failed to log activity to the CRM.
• The founders were running every demo manually behind the scenes to make the agent look smart.

What we did

• LangGraph durable agent state machine with tool retries and traces
• 38 callable tools across qualification, scheduling, CRM, and pricing
• Pinecone retrieval over product docs and 200 prior call transcripts
• Bidirectional HubSpot sync, leads, activities, deal stage
• Refusal + handoff-to-human rules for adversarial prompts
• 60-case eval harness running on every change

The result

• Of the 40 demos, 14 closed within 30 days, a 35% close rate at an average ACV the founders haven't shared publicly.
• The agent now runs unattended, and the team has shifted from running demos to coaching the agent. They've since added a Salesforce variant for a separate enterprise prospect.