Implementing AI is not a “finished project.”

It’s a change that can’t yet be clearly scoped or precisely quantified.
The latest business roundtable focused on why it’s so hard to bring AI safely into production. The discussion was lively from the start. A strong community of experts is forming, real-world experience is growing, and so are new obstacles—ones that are worth tackling together.

A few key takeaways:

Data has its own lifecycle.
Models age. Without observability into both data and model behavior, you won’t notice changes. Once data structure shifts, models can start to fail.

“We don’t know how much this will cost—and that’s worse than knowing it will be expensive.”
Budgeting AI without understanding operations is a mistake. It’s not just about training and development, but also infrastructure and lifecycle management—and AI lifecycles are not linear.

“It’s unclear who is responsible for AI, and when. Infrastructure, development, or someone else?”
Typical MLOps approaches today often break the fundamentals of agile development when run in parallel. Models get thrown over the wall like code used to be, and the DevOps mindset disappears. That’s a mistake. We need teams and platforms that control the entire lifecycle end to end.

“AI is hard to version—not because it’s technically impossible, but because context gets lost.”
Model versions alone aren’t enough. You need snapshots of data, configurations, computations, and behavior. Otherwise, there’s nothing to audit.

How to move forward:

AI must behave like a component—versioned and deployable. That’s why we package it into infrastructure building blocks (Docker containers, microservices), exactly the same way any other cloud-ready software is built.