Summary
What happens when your AI pilot works too well to stay a pilot? In this episode of The WorkOps Podcast, host Jeet Mukerji talks with Ivan Nosov, Head of HR Tech and Global Total Rewards Director at Campari Group, about taking AI in HR from proof of concept to production. Ivan shares why he started with Copilot Studio knowing it wouldn't last, where RAG breaks down at enterprise scale, and how his team built a digital twin router app on Claude that loads only the context each question needs. He also lays out the adoption playbook that made it stick: winning over regional directors and COE champions first, running hackathons for awareness, and addressing the work people hate instead of the work they love. A practical conversation for HR, people ops, and workplace technology leaders navigating AI transformation.
Chapters
00:00 Introduction
01:30 From IT engineer to HR leader
03:45 Should AI transformation sit in HR or IT
05:55 How non technical HR can get started
07:20 Copilot Studio as a proof of concept tool
09:30 Harness, routing, and the limits of RAG
15:00 Building the digital twin router app on Claude
17:40 Context engineering and distilling tacit knowledge
19:30 Winning adoption through champions
Takeaways
- Start with the accessible tool to prove the concept, then move on. Copilot Studio validated Campari Group's AI ideas, but production required control over the harness that abstracted platforms can't offer.
- Routing beats RAG at scale. Loading only the relevant context for each question makes AI more targeted, more efficient, and far less likely to hallucinate.
- Campari Group's digital twin runs on Claude with deliberate model selection: Sonnet for efficiency, Haiku for helper queries, and Opus for final artifacts because of its design taste.
- Adoption spreads through champions. Win over regional directors and COE heads first, stay with them until the outputs click, and they will share it further than any rollout plan.
- Context engineering is the real work. Distill tacit knowledge into a compact, unambiguous knowledge base: 250,000 tokens in total, with no session loading more than 50,000.
Connect with the Guest
LinkedIn: https://www.linkedin.com/in/ivan-nosov/
Sponsor
This episode is brought to you by Kinfolk, the AI service desk built for HR.
See more at kinfolkhq.com
This episode is brought to you by Kinfolk, the AI service desk built for HR.
See more at kinfolkhq.com