Making Enterprise Gen AI “Reliable” “The irony of GenAI progress? Models are getting smarter — but leaders are losing confidence in shipping them.”

Atindriyo Sanyal, combining experiences from building Siri at Apple to scaling Uber’s Michelangelo AI platform, offers a critical examination of the reliability challenges confronting generative AI (GenAI) in enterprise settings. His insights highlight the paradox facing AI leaders: despite advances in model intelligence, organizational confidence in deploying these technologies beyond prototypes remains weak due to subtle but impactful system failures and escalating operational complexities.

The keynote dissects key barriers to production readiness, including the high failure rate of retrieval-augmented generation (RAG) applications and the inherent risks introduced by multi-agent frameworks. Sanyal advances a framework centered on three pillars essential to reliable AI systems: rigorous evaluation protocols, continuous observability mechanisms, and robust guardrails that ensure operational safety and compliance. He further illustrates these principles through innovative solutions like lightweight Luna models, which facilitate instantaneous, scalable guardrails that can transform the enterprise AI deployment landscape.

• Nearly 50% of enterprise RAG applications fail to progress beyond the prototype stage due to trust erosion.
• Silent, nuanced errors in AI systems undermine leader confidence and stall adoption.
• Multi-agent system complexity compounds risk and complicates reliability assurance.
• The three foundational pillars—evaluation, observability, and guardrails—are paramount to building trustworthy AI.
• Innovations such as Galileo’s Luna models demonstrate scalable approaches to embedding guardrails and enhancing system responsiveness.