Field Notes: Why Legacy Databases, Not LLMs, Define AI Integration Work

An AI integration consultant working with a B2B software company describes a typical week spent mapping data systems rather than writing code, highlighting that the real challenge lies in understanding where reliable data lives across tools like Salesforce, NetSuite, Zendesk, and a decade-old MySQL database. The client had requested an AI assistant to answer customer questions from a knowledge base and CRM, but the consultant found that a legacy MySQL app — not the modern SaaS platforms — was the true source of truth for customer entitlements. Without reading that database, the AI assistant risked generating incorrect responses about product access, which could trigger support escalations. The consultant outlines a decision framework for choosing integration patterns — synchronous calls, event-driven queues, scheduled workers, or agent loops — arguing that most so-called 'agentic' tasks are better handled by simpler, deterministic workflows. The article concludes that data governance and system mapping must precede any LLM implementation, and that skipping this step is the most common reason AI integrations fail months after launch.