Why we built a RAG-based AI HR assistant

At CTI Digital, our Innovation Team has been building a range of AI tools in collaboration with the wider workforce, aiming to increase productivity and reduce time-intensive, low value tasks.

In this blog post, we’ll share one of these AI projects – a RAG-based HR assistant that answers the most routine policy and process questions.

Colleagues can access the HR Assistant in two ways – via a bot in Slack or as a knowledge base in our internal LLM, with information drawn only from CTI's own internal documentation. We can ask questions of the assistant in natural language, resulting in faster answers, and fewer repetitive questions landing in HR’s inbox.

The challenge

As much as 30% of HR time is spent on administrative ‘quick questions’, an issue that is acknowledged by Tom Barnes, Head of People at CTI Digital.

“Before we created the AI HR assistant, a significant amount of time was spent answering the same everyday questions - from logging time and reporting sickness to understanding our ways of working,” says Barnes.

“Those questions matter,“ he adds, “but by making answers quicker and more accessible, we’ve been able to focus more of our energy on the work that has the greatest impact: developing our people, strengthening the employee experience and giving everyone the tools, support and opportunities they need to thrive at CTI.”

The aim was to create a win-win. Not just helping HR to be more efficient and engaged, but also reducing employee frustration at having to search policy documents themselves, or simply in waiting for a reply.

The solution

“We indexed the HR knowledge base – including policies, FAQs, and leave entitlements – in a vector store,” said Saiful Thoombath, Digital Innovation Specialist at CTI Digital.

“When a query comes in, semantic search retrieves only the most relevant policy sections before passing them to the LLM, rather than dumping the full document set into context. This keeps responses grounded in accurate, company-specific information rather than the model's general training data, and keeps token usage efficient as the knowledge base grows.”

“We orchestrated the workflow (Slack listening, query routing, retrieval, and response delivery) end to end. We run inference on a private, locally-hosted LLM rather than a public API like OpenAI, Claude, or Gemini. That means no HR data is sent to third-party AI providers for processing. The query is received via Slack, retrieved against the vector store, and processed entirely on internal infrastructure,” noted Thoombath.

The Results

In its first five months, the assistant has saved roughly 30 hours per month handling over 300 questions, across diverse subjects including the office, IT equipment, security, timesheets, and absence.

The assistant has proven particularly useful for new starters, who can lean on the bot where necessary in their first few weeks, to ask for a refresher on key induction FAQs.

As was expected, this work is gradually freeing up HR time to focus on more complex questions.

Next steps and learnings

“RAG is relatively straightforward to implement when the knowledge base is structured and bounded,” says Thoombath. “HR policies, onboarding guides, IT FAQs, and compliance documentation are ideal candidates, since they're authoritative, largely static, and have clear scope.”

The CTI Innovation Team is exploring other internal systems that may benefit from similar RAG integration. But the logical next step for the HR assistant itself is to experiment with moving from reactive Q&A to proactive onboarding support, guiding new starters through mandatory training, company platform walkthroughs, and induction videos at the right point in their first weeks, rather than relying on HR to guide them from start to end. The goal will be to save administrative time when a new starter joins, without precluding the all-important face-to-face time between our colleagues.

If you’re interested in AI assistants or privately-hosted LLMs, or you’d like to talk to our Innovation Team, drop us a message.