Building a Privacy-First AI CRM Assistant for My Friend's Business

Matt operates a small business, but his daily intake process was becoming cumbersome. Clients communicated through email, website forms, and messaging apps. The incoming data varied greatly; some inquiries were promising sales leads, others were support questions, some were irrelevant junk, and some contained insufficient information.

He required a system to collect these inquiries, determine their nature, extract relevant structured data, and automatically generate the next course of action. Privacy was a major concern, as sending client information to off-site, closed-source language models was not an option. In response to a Hacktoberfest Weekend Challenge, Matt developed a secure, locally hosted AI CRM intake automation system written in Go.

The system encompasses several key functions: it ingests incoming messages, classifies them as sales leads, support tickets, junk, or incomplete, extracts key data such as names, contact information, and business requirements, drafts an appropriate response, and notifies Matt's team for review and approval before any updates are made to the CRM or any emails are sent out.

The prototype is capable of processing inquiries locally and providing structured data for further processing. The backend system was constructed primarily in Go, adhering to the principle that AI analyzes the data, while deterministic code dictates the actions. The AI's role is limited to classification, extraction, and drafting, and it does not directly interact with emails, CRM records, or perform arbitrary actions.

All consequential actions, such as sending emails or updating CRM records, require human approval. By deploying open-source AI locally using Ollama, Matt addressed the issue of data privacy, avoided costs associated with third-party API usage, and maintained control over the AI model, prompting mechanisms, and output formatting.