Looking Back on the First 100 Days of Our Website's AI Chatbot (spin.atomicobject.com)

Three months after deploying an LLM-powered chatbot trained on curated pages from their site, Atomic Object reports the first 100 days as a successful learning experiment rather than a lead-generation silver bullet. The agent was given explicit instructions to assist prospective custom-software clients, solicit contact details for a Marketing Specialist to follow up, redirect job-seekers to the careers page, and avoid inventing facts. After internal stress-testing and a disclosure that users were chatting with an automated tool, the bot handled ~1,900 messages from ~170 visitors. Topic breakdowns showed prospect behavior: process questions (27.5%), services (19.8%), pricing (15%), with many visitors describing concrete project ideas but few voluntarily sharing contact info in-chat. For the AI/ML community this is a pragmatic case study in applying a site-specific LLM: manual curation of training material mattered, safety and instruction-tuning limited scope and hallucinations, and internal piloting revealed real failure modes. Key implications—especially for high-cost, bespoke services—are that such agents surface intent and content gaps more reliably than qualified leads, so teams should prioritize strong website content, internal stress tests, and use chatbot transcripts to refine both FAQ/marketing assets and handoff processes rather than expecting direct conversions.