Knowledge Base + RAG — Go Global UAE

Problem

Go Global UAE managers were overwhelmed with repetitive client questions — about documents, admission process, and regulations. Outside business hours, clients waited days for a response and hot deals were lost.

At the same time, every new employee took hours of personal explanation from the department head. Onboarding was a bottleneck for both the team and company growth.

Solution

The bot became second-line support and onboarding mentor simultaneously.

A client writes in Telegram → gets a response in seconds, formatted as if a live manager replied. Not "here's a FAQ link" — a specific answer to a specific question with full conversation context.

A new employee starts work → gets access to the bot in "all about policy and process" mode. No need to bother the manager — the bot knows where everything is and how everything works.

A partner-ambassador asks a question → sees only the information they're supposed to see, without access to internal manager materials.

Outcome

The most valuable number — close rate grew by ~35%. Out of every 100 potential clients, 35 additional ones now convert. Before, those deals were lost in the window of "manager is away, client went cold."

Managers are freed up by half — not "half were fired," but "half their time is freed for complex negotiations." Simple questions the bot handles itself.

The most unexpected effect — the onboarding mentor role ceased to exist as a function. No firings, the work simply stopped being work. New employees complete onboarding via bot in a couple of days instead of two weeks. Managers returned to sales.

200+ clients per month get instant responses outside business hours. On weekends. During peaks. At night. This isn't optimization — it's a new category of service.

Lessons Learned

The key lesson — access control matters more than generation quality. Real value came when we correctly separated roles: what a manager vs client vs partner sees. The same question from different people requires different answers — that's business logic the model can't assemble on its own.

On a corpus of 1200+ pages, hybrid retrieval performed noticeably better than pure vector search. Technical terms and rare names got lost in semantic similarity but were caught by keyword search.

The sharpest moment at launch — the first version occasionally "filled in" details that weren't in the knowledge base. The fix wasn't swapping the model — it was strict constraints in the system prompt and a verification step before answering. The prompt is business logic.