A software services company was drowning in repetitive support emails — the same twenty questions, asked in two languages, around the clock. I built Sanad Desk: an AI assistant that answers from their actual documentation, escalates gracefully to humans, and gets measurably better every month. Six weeks from kickoff to production.

Two support staff were spending most of their day copy-pasting answers that already existed in the help docs. Response times stretched to 12+ hours overnight and on weekends, and customers in Arabic often got slower answers because only one agent spoke it fluently. Hiring more staff would cost far more than the problem justified.

I built a retrieval-augmented assistant: the company's docs, FAQs, and past resolved tickets are ingested into a vector database, and the AI answers only from that verified content — with citations, in the customer's language. When confidence is low, it hands off to a human with full conversation context. An admin dashboard shows every conversation, flags weak answers, and lets staff correct them, feeding improvements back into the knowledge base.

The critical early decision was scoping what the AI should refuse to answer — billing disputes and refunds always route to humans. We ran a two-week shadow phase where the AI drafted answers that staff approved before sending, which built the team's trust and produced a corpus of corrections that sharpened accuracy before the assistant ever spoke to a customer unsupervised.