AI apps need production operations

Google Cloud currently explains why many AI apps fail in production: a local prototype can now be built in hours, but an enterprise deployment hits networks, dependencies, approvals, operating rules and risk tolerance. Google cites a blunt figure in the post: only 5 percent of AI prototypes make it to production.

What is changing

The relevant news is not that prototyping is faster. That is already expected. What matters is how clearly vendors and practitioners now separate a demo from production operations.

Google argues from a developer and platform perspective: unconstrained agentic orchestration can accelerate technical debt when changes are not isolated, tested and controlled. Large platforms cannot let every agent idea immediately touch production paths.

AWS shows the other side with monday.com. Production agents need their own harness. monday.com describes the runtime as replaceable, while the enterprise-specific logic lives in the harness: evaluation, plugin composition, connections to Slack, monday and GitHub, and review against internal standards. The architecture also separates live state, sessions, memory, secrets and event processing across services such as Amazon Bedrock, Amazon EKS, Amazon SQS, Amazon SNS, Amazon ElastiCache and Amazon EFS.

Why it matters

For CIOs and IT leaders, this is an important shift. The question is no longer: “Can we build an AI demo quickly?” It is: “Can we turn it into a system that is versioned, observable, repeatable, reviewable and stoppable?”

The AWS post on Amazon Quick and NVIDIA NeMo Relay reinforces the point. Agentic workflows are not only connected; they are evaluated, profiled and optimized. Enterprise AI is therefore moving closer to established software and platform disciplines: SDLC, SRE, security, FinOps and change management.

DACH perspective

For organizations in the DACH region, production operations matter because privacy, auditability and operational accountability are hard to add cleanly after the fact. An AI use case should therefore have an operating contract before rollout: owner, permitted data classes, model path, tool permissions, logging, evaluation set, cost limit, review requirement and rollback rule.

The pragmatic next step is a production gate between pilot and scale. Go-live should not be decided by the nicest prompt, but by whether harness, state, monitoring and accountability are robust enough.