Enterprise AI Has A Production Problem

Abhijit Kakhandiki is senior vice president and general manager of Digital Business Automation at BMC Software, Inc.

Enterprises aren’t struggling to build AI. They’re struggling to make it matter.

I recently spoke with the CIO of a large global enterprise whose data science teams had built more than 200 AI models, yet fewer than 10 were in production.

That gap is the defining pattern of enterprise AI today. Organizations are rich in experimentation but poor in production. McKinsey’s 2025 global AI survey found that while AI use is widespread, most organizations remain in experimentation phases, with only about one-third reporting they have begun scaling AI programs across the enterprise. It’s tempting to blame talent, tooling or model performance, but the issue is that enterprises treat AI as a technology problem when it’s an operational one. The hard part is running AI reliably, governing it and integrating it into the workflows the business already depends on.

Big data and machine learning followed the same trajectory, where early innovation outpaced enterprise readiness, and later operational discipline caught up. AI is now at that inflection point. What distinguishes the organizations moving forward is not better models, but an understanding of what it takes to make AI part of how the business runs.

Three persistent breakdowns explain why AI stalls before production.

1. Starting With Technology Instead Of Business Outcomes

Thousands of vendors populate the AI and data landscape with an excess of tools, and enterprises get caught up in assembling technically sophisticated solutions without establishing what success looks like.

A team I met with built a complex generative AI solution with RAG pipelines, vector databases that needed regular refresh and fine-tuned models. When asked which business KPI it was improving, there wasn’t a clear answer. Months later, despite significant investment, the system remained stuck in pilot.

Leaders begin asking operational questions: