Enterprise AI deployments are accelerating, but the velocity of change in models, tools and platforms is quietly creating a new and costly trap: costly failure to build for AI architectural replaceability.

As Google Cloud Next 2026 confirms that agentic AI has become the organizing logic for the entire Google LLC product portfolio, the harder question for enterprise leaders is not what to build — it is how to build it with replaceability front of mind. That distinction is where many organizations are quietly failing, according to Paul Lewis (pictured), chief technology officer at Pythian Services Inc., a managed AI operations and advisory firm.

“Whatever you implement, your only non-functional requirement that matters is replaceability,” Lewis said. “Ensure that anything you do, the tool can be replaced, the model can be replaced, the team can be replaced, the expertise can be replaced, because I guarantee you within weeks — not months, not quarters, not years — it will be different.”

Lewis spoke with theCUBE’s John Furrier and Alison Kosik at Google Cloud Next, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed AI architectural replaceability, enterprise AI readiness and the gap between production demos and real-world deployment. (* Disclosure below.)

AI architectural replaceability and the production gap

The shift from build to operate is the defining challenge at this year’s event, Lewis noted. After conducting about 50 customer workshops last year, Pythian found no consistent pattern of AI maturity across enterprises — responses ranged from “I’ve never heard of it” all the way up to the organizations reporting billion-dollar investments.

“Last year was all about ‘build,'” he said. “Unfortunately, the vast majority of those [pilots] never actually went to production. You ended up doing a lot of internal education.”

The most common failure mode is not technical — it is the gap between a polished five-minute demo and the months of design and implementation work behind it. Enterprise friction, including departmental approvals and change management, does not disappear because of AI; it has to be navigated through it, Lewis explained. Pythian has responded by building five practice areas — from a field CTO advisory function to managed AI operations — to help customers bridge that divide.

“I can put an agent in production that’s 70% accurate,” he said. “You kind of want it in the 90s. That takes prompt change and model change and data source change. They have lifespans just like an application does and that requires a team.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of Google Cloud Next:

(* Disclosure: Pythian sponsored this segment of theCUBE. Neither Pythian nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE