October 3, 2026

An AI Assistant for 500+ Employees: What Actually Mattered

Lessons from a RAG assistant that resolves about 65% of HR, IT and operations questions without a human, and why the hand-off matters as much as the answers.

  • rag
  • enterprise-ai
  • knowledge-management
Two colleagues working on laptops at a shared marble table

Most internal questions already have an answer somewhere: a policy PDF, a runbook, a wiki page nobody can find. The problem isn’t missing knowledge. It’s that getting to it takes a ticket and a few hours.

At SBER, I led the team that built a RAG-based AI assistant for 500+ employees across HR, IT and operations. It handles about 65% of queries without a human, cut response time from hours to seconds, and saved the equivalent of 3 FTE.

Here’s what mattered more than I expected.

Ground every answer in your own content

Retrieval-augmented generation is the right default for internal assistants. The knowledge changes, it’s specific to the organisation, and people need to trust that the answer reflects current policy. Answering from retrieved internal content, rather than the model’s general knowledge, is what makes that possible.

Retrieval quality is the ceiling

If the right passage isn’t retrieved, no prompt will fix the answer. Most of the “the model hallucinated” reports in a RAG system are really retrieval misses. That means retrieval deserves its own measurement and its own iteration loop, separate from generation.

Design the hand-off, not just the answers

The headline number is that about 65% of queries are resolved without a human. The other 35% matter just as much. An assistant that confidently answers questions it shouldn’t is worse than no assistant. Routing the rest to the right people is what makes the system shorten the queue instead of becoming another dead end.

Measure in the units the business uses

“Answer quality improved” doesn’t get a project funded. Response time (hours to seconds), deflection (about 65%) and effort saved (3 FTE) do. Choose those measures up front and instrument for them from the first release.

The takeaway

An enterprise RAG assistant succeeds on ordinary things: retrieval that finds the right passage, an honest hand-off when it can’t, and metrics the business already cares about.

— Kamish Upadhyay, Senior GenAI / AI Architect. Get in touch if you’re building something similar.

Keep reading.

When Fine-Tuning Earns Its Cost

A practitioner's guide to choosing between prompting, RAG and fine-tuning, and to the parameter-efficient methods (PEFT, LoRA, QLoRA) that make fine-tuning practical.

Text-to-SQL That Business Users Can Trust

Generating SQL is the easy part. Validation, orchestration and returning a chart instead of a table are what got 300+ users to stop filing report requests.

Putting a Voice Agent in Front of Bank Customers

Accuracy, latency and cost all compound in a speech pipeline. Here is how we got a STT → LLM → TTS banking agent to 92% intent accuracy at 1.8 s p95 and ₹0.48 a turn.