- Only 6% of enterprise AI Proof-of-Concepts (POCs) make it to production.
- AI product development often starts with technological research, not customer problems.
- Private AI deployments are shifting the industry from SaaS back to license models.
- Measuring AI product success requires intimate customer understanding beyond usage metrics.
The landscape of AI product development for enterprises is undergoing a radical transformation. As companies rush to integrate artificial intelligence, many are discovering that traditional product management playbooks and sales cycles no longer apply. This shift demands a new approach, particularly as the industry grapples with the complexities of moving from impressive demos to scalable, production-ready solutions.
Eran Erez, host of the 'Productive' podcast, recently sat down with Ord Egan, Product and Strategy Officer at AI 21, to dissect these evolving challenges. AI 21, a leading AI lab developing language models from scratch, specializes in complex AI systems for enterprise clients. Egan highlights that building AI products in 2025 is unique due to two main factors: the nature of horizontal AI technology itself, and the drastically different enterprise engagement model compared to previous years.
One of the most striking revelations is the high failure rate of AI POCs. AWS data indicates that a mere 6% of enterprise AI initiatives move from POC to production. This stark reality stems from what Egan describes as the 'prompt and pray' approach, where initial demos, often achieving 50-60% functionality, create a 'wow factor' that masks underlying complexities. The journey from this initial impressive demo to a robust, 80-90% reliable solution is a significant hurdle, requiring traditional software development practices, extensive evaluation, and substantial R&D investment. Enterprises are now seeking 'Enterprise Grade' solutions, demanding proven data sheets, robust evaluation tools, and strong internal skill sets.
The shift towards 'Private AI' introduces another layer of complexity. Driven by regulatory concerns, data privacy requirements, and a desire for greater ownership, many organizations want to host AI models on-premise or within their Virtual Private Clouds (VPCs). This move effectively transitions the industry from a SaaS model back to a license-based one, presenting product managers with new challenges in onboarding, pricing, updates, and crucially, measuring adoption and success. Without direct access to usage data, maintaining an intimate and proactive relationship with customers, focusing on expansion into new use cases, and understanding their specific value perception becomes paramount.
“The challenge here is that it's like, sometimes we talk about it, to boil the ocean, because it's like everything, it's all types of things that all people can ask for, and here you need to manage, here you need to understand the technology and manage risks smartly.”
- Or Dagan, Chief Product & Strategy Officer at AI21
