GenAI assistant with evaluation and guardrails

The challenge

An enterprise wanted to deploy an internal AI assistant to help knowledge workers find information across internal documentation, policies and procedures. Several proof-of-concept implementations had been built, but none had been approved for production use. The risk and compliance team had legitimate concerns about data leakage, hallucinated policy guidance and the inability to audit what the assistant was telling employees.

The technology team had focused on prompt engineering and model selection but had not addressed the operational requirements — logging, evaluation, access control and human review — that would satisfy governance requirements. The gap between "working demo" and "production-ready system" was wider than anyone had initially estimated.

What we did

Phase 1 — Requirements and red lines. We worked with risk, compliance and legal teams to define explicit requirements: what the assistant could and could not do, what data it could access, what logging was required, and what review processes were needed. This produced a clear specification that the engineering team could build against.

Phase 2 — Retrieval and grounding. We built a retrieval-augmented generation (RAG) pipeline that grounded the assistant's responses in authoritative internal documents. Every response included source citations, and the system was designed to decline questions where it could not find relevant source material rather than generating speculative answers.

Phase 3 — Evaluation harness. We built an automated evaluation framework that tested the assistant against a curated set of questions with known correct answers. The framework measured accuracy, source attribution quality, hallucination rate and boundary adherence. Evaluations ran automatically on every model update and configuration change.

Phase 4 — Operational guardrails. We implemented full logging of all interactions, content filtering for sensitive topics, rate limiting per user, and a human review queue for flagged responses. Dashboards showed usage patterns, quality metrics and compliance status in real time.

Measurable results

The assistant was approved for production use — the organisation's first AI feature to clear the governance process. The evaluation harness became the standard for all subsequent AI initiatives, reducing the time from prototype to production approval by an estimated 60%.